JoshuaW1997/tdprm-results

Experiment logs, synced 2026-10-19 03:52:31 every few minutes. tdprm_results.tar (7242 MB, tar -xf tdprm_results.tar -> tdprm_results/) holds every metric, trace, evaluation, state, queue and service log of the run; no model weights, environments or dataset text.

TD-PRM live status

Updated 2026-10-19 03:52:03 · refreshes automatically

Training queue (finished): training runs (97) done 97; other jobs (data, states, base evals) done 44

Step-credit evaluation queue (finished): 10 jobs, done 10 (see Step-credit evaluation)

GPUs

GPU	memory	util	job
0	67 / 80 GB	97%	train_llama3b-countdown-b_hc_v3-s1
1	55 / 80 GB	98%	train_qwen1.5b-kk-j_hc_v3-s1
2	55 / 80 GB	100%	train_llama3b-countdown-b_hc_v3-s1
3	58 / 80 GB	100%	train_llama3b-kk-b_hc_v3-s1
4	45 / 80 GB	0%	train_llama3b-kk-b_hc_v3-s1
5	0 / 80 GB	0%	idle
6	48 / 80 GB	97%	train_llama3b-countdown-grpo-s1
7	48 / 80 GB	100%	train_llama3b-countdown-grpo-s1

Final test

Greedy accuracy, v1 and v2 methods in separate tables. Δ: v1 methods vs grpo, v2 methods vs grpo_v2 (same cell), GRPO rows vs the base model. pass@8 on the in-domain set; mean@8 / pass@8 for every set are in evals/<run>/step_200.json. – = not tested on that set (runs that finished before ProverQA replaced HotpotQA; no checkpoints to re-test).

Llama-3B · trained on MATH

v1 methods (Δ vs grpo)

method	MATH500	GSM8K	PQ easy	PQ medium	PQ hard	pass@8 MATH500
base model	0.488	0.814	0.662	0.444	0.374	0.664
grpo	0.528 (+0.040)	0.834 (+0.020)	0.684 (+0.022)	0.490 (+0.046)	0.420 (+0.046)	0.728
s_hc	0.496 (-0.032)	0.827 (-0.007)	0.680 (-0.004)	0.492 (+0.002)	0.380 (-0.040)	0.710
s_gae	0.504 (-0.024)	0.835 (+0.001)	0.378 (-0.306)	0.330 (-0.160)	0.248 (-0.172)	0.712
s_gaed	0.492 (-0.036)	0.838 (+0.004)	0.016 (-0.668)	0.014 (-0.476)	0.012 (-0.408)	0.736
j_hc	0.496 (-0.032)	0.818 (-0.016)	0.074 (-0.610)	0.042 (-0.448)	0.018 (-0.402)	0.706
j_gae	0.526 (-0.002)	0.842 (+0.008)	0.646 (-0.038)	0.518 (+0.028)	0.392 (-0.028)	0.726
j_gaed	0.542 (+0.014)	0.839 (+0.005)	0.648 (-0.036)	0.506 (+0.016)	0.400 (-0.020)	0.738
prm800k	0.248 (-0.280)	0.450 (-0.384)	0.466 (-0.218)	0.418 (-0.072)	0.384 (-0.036)	0.336
qwenprm	0.444 (-0.084)	0.826 (-0.008)	0.496 (-0.188)	0.416 (-0.074)	0.412 (-0.008)	0.584
versaprm	0.396 (-0.132)	0.798 (-0.036)	0.542 (-0.142)	0.446 (-0.044)	0.396 (-0.024)	0.544

v2 methods (Δ vs grpo_v2; last column: the v1 version's MATH500 greedy and the v2 − v1 change)

method	MATH500	GSM8K	PQ easy	PQ medium	PQ hard	pass@8 MATH500	v1 version (MATH500)
base model	0.488	0.814	0.662	0.444	0.374	0.664
grpo_v2	0.518 (+0.030)	0.825 (+0.011)	0.674 (+0.012)	0.476 (+0.032)	0.404 (+0.030)	0.720	grpo 0.528 (-0.010)
s_gae_v2	0.526 (+0.008)	0.835 (+0.010)	0.610 (-0.064)	0.468 (-0.008)	0.408 (+0.004)	0.726	s_gae 0.504 (+0.022)
vp_gae_v2	0.528 (+0.010)	0.840 (+0.015)	0.702 (+0.028)	0.514 (+0.038)	0.386 (-0.018)	0.724	new in v2
qwenprm_v2	0.542 (+0.024)	0.845 (+0.020)	0.680 (+0.006)	0.486 (+0.010)	0.394 (-0.010)	0.704	qwenprm 0.444 (+0.098)

v3 methods (Δ vs v1 grpo, same seed; last column: the v1 counterpart's MATH500 greedy and the v3 − v1 change)

method	MATH500	GSM8K	PQ easy	PQ medium	PQ hard	pass@8 MATH500	v1 counterpart (MATH500)
base model	0.488	0.814	0.662	0.444	0.374	0.664
grpo (v1 reference)	0.528 (+0.040)	0.834 (+0.020)	0.684 (+0.022)	0.490 (+0.046)	0.420 (+0.046)	0.728
b_td_v3	0.502 (-0.026)	0.811 (-0.023)	0.678 (-0.006)	0.512 (+0.022)	0.418 (-0.002)	0.704	new in v3
s_hc_v3	0.448 (-0.080)	0.805 (-0.029)	0.646 (-0.038)	0.444 (-0.046)	0.396 (-0.024)	0.678	s_hc 0.496 (-0.048)

Llama-3B · trained on ProverQA

v1 methods (Δ vs grpo)

method	MATH500	GSM8K	PQ easy	PQ medium	PQ hard	pass@8 PQ hard
base model	0.488	0.814	0.662	0.444	0.374	0.750
grpo	0.516 (+0.028)	0.825 (+0.011)	0.942 (+0.280)	0.882 (+0.438)	0.724 (+0.350)	0.846
s_hc	0.142 (-0.374)	0.165 (-0.660)	0.902 (-0.040)	0.864 (-0.018)	0.748 (+0.024)	0.790
s_gae	0.500 (-0.016)	0.822 (-0.003)	0.906 (-0.036)	0.820 (-0.062)	0.658 (-0.066)	0.826
s_gaed	0.506 (-0.010)	0.812 (-0.013)	0.896 (-0.046)	0.822 (-0.060)	0.642 (-0.082)	0.780
j_hc	0.496 (-0.020)	0.819 (-0.006)	0.902 (-0.040)	0.836 (-0.046)	0.672 (-0.052)	0.758
j_gae	0.488 (-0.028)	0.813 (-0.012)	0.926 (-0.016)	0.878 (-0.004)	0.700 (-0.024)	0.828
j_gaed	0.502 (-0.014)	0.821 (-0.004)	0.942 (+0.000)	0.894 (+0.012)	0.644 (-0.080)	0.820
prm800k	0.118 (-0.398)	0.095 (-0.730)	0.886 (-0.056)	0.820 (-0.062)	0.654 (-0.070)	0.660
qwenprm	0.116 (-0.400)	0.089 (-0.736)	0.830 (-0.112)	0.800 (-0.082)	0.610 (-0.114)	0.624
versaprm	0.120 (-0.396)	0.090 (-0.735)	0.846 (-0.096)	0.826 (-0.056)	0.680 (-0.044)	0.686

v2 methods (Δ vs grpo_v2; last column: the v1 version's PQ hard greedy and the v2 − v1 change)

no finished runs yet

v3 methods (Δ vs v1 grpo, same seed; last column: the v1 counterpart's PQ hard greedy and the v3 − v1 change)

no finished runs yet

Qwen-1.5B · trained on MATH

v1 methods (Δ vs grpo)

method	MATH500	GSM8K	PQ easy	PQ medium	PQ hard	pass@8 MATH500
base model	0.388	0.337	0.622	0.374	0.356	0.640
grpo	0.570 (+0.182)	0.757 (+0.421)	–	–	–	0.800
grpo s2	0.576 (+0.188)	0.760 (+0.423)	0.642 (+0.020)	0.468 (+0.094)	0.334 (-0.022)	0.782
grpo s3	0.590 (+0.202)	0.747 (+0.410)	0.680 (+0.058)	0.488 (+0.114)	0.338 (-0.018)	0.794
s_hc	0.442 (-0.128)	0.684 (-0.074)	–	–	–	0.680
s_gae	0.598 (+0.028)	0.754 (-0.004)	–	–	–	0.780
s_gaed	0.548 (-0.022)	0.768 (+0.011)	–	–	–	0.788
j_hc	0.568 (-0.002)	0.750 (-0.008)	0.422	0.272	0.206	0.764
j_gae	0.566 (-0.004)	0.749 (-0.008)	0.670	0.498	0.342	0.778
j_gaed	0.552 (-0.018)	0.763 (+0.005)	0.666	0.462	0.320	0.780
prm800k	0.576 (+0.006)	0.749 (-0.008)	–	–	–	0.776
qwenprm	0.360 (-0.210)	0.669 (-0.088)	–	–	–	0.480
versaprm	0.428 (-0.142)	0.591 (-0.167)	0.446	0.482	0.362	0.604

v2 methods (Δ vs grpo_v2; last column: the v1 version's MATH500 greedy and the v2 − v1 change)

method	MATH500	GSM8K	PQ easy	PQ medium	PQ hard	pass@8 MATH500	v1 version (MATH500)
base model	0.388	0.337	0.622	0.374	0.356	0.640
grpo_v2	0.602 (+0.214)	0.776 (+0.440)	0.674 (+0.052)	0.462 (+0.088)	0.322 (-0.034)	0.774	grpo 0.570 (+0.032)
s_hc_v2	0.594 (-0.008)	0.748 (-0.028)	0.666 (-0.008)	0.466 (+0.004)	0.374 (+0.052)	0.786	s_hc 0.442 (+0.152)
s_gae_v2	0.588 (-0.014)	0.766 (-0.011)	0.666 (-0.008)	0.504 (+0.042)	0.324 (+0.002)	0.776	s_gae 0.598 (-0.010)
s_gae_v2_nogate	0.572 (-0.030)	0.763 (-0.014)	0.608 (-0.066)	0.482 (+0.020)	0.366 (+0.044)	0.778	s_gae 0.598 (-0.026)
s_gaed_v2	0.582 (-0.020)	0.754 (-0.023)	0.674 (+0.000)	0.482 (+0.020)	0.312 (-0.010)	0.780	s_gaed 0.548 (+0.034)
j_hc_v2	0.566 (-0.036)	0.762 (-0.014)	0.686 (+0.012)	0.494 (+0.032)	0.338 (+0.016)	0.792	j_hc 0.568 (-0.002)
j_gae_v2	0.598 (-0.004)	0.757 (-0.020)	0.658 (-0.016)	0.466 (+0.004)	0.336 (+0.014)	0.796	j_gae 0.566 (+0.032)
j_gaed_v2	0.588 (-0.014)	0.764 (-0.012)	0.692 (+0.018)	0.478 (+0.016)	0.380 (+0.058)	0.794	j_gaed 0.552 (+0.036)
vp_hc_v2	0.574 (-0.028)	0.756 (-0.020)	0.666 (-0.008)	0.470 (+0.008)	0.352 (+0.030)	0.780	new in v2
vp_gae_v2	0.578 (-0.024)	0.759 (-0.017)	0.624 (-0.050)	0.488 (+0.026)	0.334 (+0.012)	0.792	new in v2
vp_gae_v2_nogate	0.576 (-0.026)	0.748 (-0.028)	0.624 (-0.050)	0.448 (-0.014)	0.316 (-0.006)	0.778	new in v2
vp_gaed_v2	0.590 (-0.012)	0.756 (-0.020)	0.642 (-0.032)	0.476 (+0.014)	0.344 (+0.022)	0.776	new in v2
prm800k_v2	0.574 (-0.028)	0.760 (-0.016)	0.624 (-0.050)	0.450 (-0.012)	0.312 (-0.010)	0.780	prm800k 0.576 (-0.002)
qwenprm_v2	0.568 (-0.034)	0.748 (-0.028)	0.648 (-0.026)	0.498 (+0.036)	0.402 (+0.080)	0.784	qwenprm 0.360 (+0.208)
versaprm_v2	0.566 (-0.036)	0.748 (-0.028)	0.686 (+0.012)	0.476 (+0.014)	0.356 (+0.034)	0.784	versaprm 0.428 (+0.138)

v3 methods (Δ vs v1 grpo, same seed; last column: the v1 counterpart's MATH500 greedy and the v3 − v1 change)

method	MATH500	GSM8K	PQ easy	PQ medium	PQ hard	pass@8 MATH500	v1 counterpart (MATH500)
base model	0.388	0.337	0.622	0.374	0.356	0.640
grpo (v1 reference)	0.570 (+0.182)	0.757 (+0.421)	–	–	–	0.800
grpo (v1 reference) s2	0.576 (+0.188)	0.760 (+0.423)	0.642 (+0.020)	0.468 (+0.094)	0.334 (-0.022)	0.782
grpo (v1 reference) s3	0.590 (+0.202)	0.747 (+0.410)	0.680 (+0.058)	0.488 (+0.114)	0.338 (-0.018)	0.794
b_td_v3	0.584 (+0.014)	0.756 (-0.002)	0.668	0.472	0.364	0.766	new in v3
b_td_v3 s2	0.566 (-0.010)	0.760 (+0.001)	0.648 (+0.006)	0.480 (+0.012)	0.332 (-0.002)	0.774	new in v3
b_td_v3 s3	0.588 (-0.002)	0.759 (+0.012)	0.676 (-0.004)	0.482 (-0.006)	0.332 (-0.006)	0.766	new in v3
b_td_v3_nozv	0.598 (+0.028)	0.767 (+0.010)	0.658	0.478	0.318	0.778	new in v3
b_hc_v3	0.562 (-0.008)	0.760 (+0.003)	0.660	0.424	0.256	0.762	new in v3
s_hc_v3	0.530 (-0.040)	0.722 (-0.036)	0.628	0.420	0.302	0.742	s_hc 0.442 (+0.088)
s_hc_v3 s2	0.550 (-0.026)	0.731 (-0.029)	0.632 (-0.010)	0.472 (+0.004)	0.368 (+0.034)	0.770	–
s_hc_v3 s3	0.530 (-0.060)	0.715 (-0.032)	0.634 (-0.046)	0.402 (-0.086)	0.296 (-0.042)	0.742	–
s_gaed_v3	0.580 (+0.010)	0.761 (+0.004)	0.462	0.434	0.332	0.780	s_gaed 0.548 (+0.032)
j_hc_v3	0.530 (-0.040)	0.756 (-0.002)	0.338	0.306	0.210	0.732	j_hc 0.568 (-0.038)

Qwen-1.5B · trained on ProverQA

v1 methods (Δ vs grpo)

method	MATH500	GSM8K	PQ easy	PQ medium	PQ hard	pass@8 PQ hard
base model	0.388	0.337	0.622	0.374	0.356	0.886
grpo	0.182 (-0.206)	0.100 (-0.237)	0.832 (+0.210)	0.782 (+0.408)	0.666 (+0.310)	0.690
grpo s2	0.512 (+0.124)	0.690 (+0.353)	0.860 (+0.238)	0.762 (+0.388)	0.576 (+0.220)	0.754
grpo s3	0.472 (+0.084)	0.644 (+0.307)	0.876 (+0.254)	0.750 (+0.376)	0.588 (+0.232)	0.704
s_hc	0.154 (-0.028)	0.086 (-0.014)	0.850 (+0.018)	0.790 (+0.008)	0.656 (-0.010)	0.674
s_gae	0.190 (+0.008)	0.143 (+0.042)	0.844 (+0.012)	0.784 (+0.002)	0.648 (-0.018)	0.676
s_gaed	0.500 (+0.318)	0.736 (+0.636)	0.844 (+0.012)	0.794 (+0.012)	0.610 (-0.056)	0.790
j_hc	0.444 (+0.262)	0.569 (+0.469)	0.876 (+0.044)	0.832 (+0.050)	0.682 (+0.016)	0.752
j_gae	0.194 (+0.012)	0.123 (+0.023)	0.852 (+0.020)	0.806 (+0.024)	0.626 (-0.040)	0.638
j_gaed	0.484 (+0.302)	0.644 (+0.544)	0.894 (+0.062)	0.774 (-0.008)	0.596 (-0.070)	0.814
prm800k	0.134 (-0.048)	0.092 (-0.008)	0.848 (+0.016)	0.768 (-0.014)	0.612 (-0.054)	0.616
qwenprm	0.126 (-0.056)	0.069 (-0.031)	0.830 (-0.002)	0.806 (+0.024)	0.650 (-0.016)	0.680
versaprm	0.126 (-0.056)	0.091 (-0.009)	0.832 (+0.000)	0.788 (+0.006)	0.642 (-0.024)	0.666

v2 methods (Δ vs grpo_v2; last column: the v1 version's PQ hard greedy and the v2 − v1 change)

method	MATH500	GSM8K	PQ easy	PQ medium	PQ hard	pass@8 PQ hard	v1 version (PQ hard)
base model	0.388	0.337	0.622	0.374	0.356	0.886
grpo_v2	0.288 (-0.100)	0.318 (-0.018)	0.554 (-0.068)	0.568 (+0.194)	0.522 (+0.166)	0.720	grpo 0.666 (-0.144)
s_hc_v2	0.378 (+0.090)	0.572 (+0.254)	0.850 (+0.296)	0.740 (+0.172)	0.576 (+0.054)	0.888	s_hc 0.656 (-0.080)
s_gae_v2	0.464 (+0.176)	0.666 (+0.348)	0.886 (+0.332)	0.810 (+0.242)	0.648 (+0.126)	0.884	s_gae 0.648 (+0.000)
s_gaed_v2	0.404 (+0.116)	0.466 (+0.147)	0.632 (+0.078)	0.536 (-0.032)	0.502 (-0.020)	0.672	s_gaed 0.610 (-0.108)
vp_hc_v2	0.382 (+0.094)	0.495 (+0.177)	0.590 (+0.036)	0.620 (+0.052)	0.582 (+0.060)	0.678	new in v2
vp_gae_v2	0.200 (-0.088)	0.119 (-0.199)	0.710 (+0.156)	0.630 (+0.062)	0.474 (-0.048)	0.560	new in v2
vp_gaed_v2	0.452 (+0.164)	0.695 (+0.377)	0.572 (+0.018)	0.594 (+0.026)	0.548 (+0.026)	0.680	new in v2
prm800k_v2	0.308 (+0.020)	0.209 (-0.109)	0.686 (+0.132)	0.600 (+0.032)	0.502 (-0.020)	0.786	prm800k 0.612 (-0.110)
qwenprm_v2	0.508 (+0.220)	0.684 (+0.365)	0.862 (+0.308)	0.798 (+0.230)	0.652 (+0.130)	0.922	qwenprm 0.650 (+0.002)
versaprm_v2	0.454 (+0.166)	0.630 (+0.312)	0.850 (+0.296)	0.796 (+0.228)	0.642 (+0.120)	0.844	versaprm 0.642 (+0.000)

v3 methods (Δ vs v1 grpo, same seed; last column: the v1 counterpart's PQ hard greedy and the v3 − v1 change)

method	MATH500	GSM8K	PQ easy	PQ medium	PQ hard	pass@8 PQ hard	v1 counterpart (PQ hard)
base model	0.388	0.337	0.622	0.374	0.356	0.886
grpo (v1 reference)	0.182 (-0.206)	0.100 (-0.237)	0.832 (+0.210)	0.782 (+0.408)	0.666 (+0.310)	0.690
grpo (v1 reference) s2	0.512 (+0.124)	0.690 (+0.353)	0.860 (+0.238)	0.762 (+0.388)	0.576 (+0.220)	0.754
grpo (v1 reference) s3	0.472 (+0.084)	0.644 (+0.307)	0.876 (+0.254)	0.750 (+0.376)	0.588 (+0.232)	0.704
b_td_v3	0.306 (+0.124)	0.159 (+0.059)	0.844 (+0.012)	0.798 (+0.016)	0.608 (-0.058)	0.662	new in v3
b_td_v3 s2	0.448 (-0.064)	0.575 (-0.114)	0.870 (+0.010)	0.836 (+0.074)	0.710 (+0.134)	0.732	new in v3
b_td_v3 s3	0.412 (-0.060)	0.578 (-0.065)	0.816 (-0.060)	0.764 (+0.014)	0.648 (+0.060)	0.672	new in v3
b_td_v3_nozv	0.216 (+0.034)	0.146 (+0.046)	0.856 (+0.024)	0.802 (+0.020)	0.656 (-0.010)	0.690	new in v3
b_td_v3_nomask	0.194 (+0.012)	0.107 (+0.007)	0.860 (+0.028)	0.828 (+0.046)	0.668 (+0.002)	0.684	new in v3
b_td_v3_zs	0.380 (+0.198)	0.327 (+0.227)	0.854 (+0.022)	0.720 (-0.062)	0.592 (-0.074)	0.670	new in v3
b_hc_v3	0.406 (+0.224)	0.578 (+0.478)	0.844 (+0.012)	0.802 (+0.020)	0.676 (+0.010)	0.706	new in v3
s_hc_v3	0.338 (+0.156)	0.411 (+0.311)	0.856 (+0.024)	0.810 (+0.028)	0.630 (-0.036)	0.642	s_hc 0.656 (-0.026)
s_gaed_v3	0.190 (+0.008)	0.264 (+0.164)	0.810 (-0.022)	0.760 (-0.022)	0.600 (-0.066)	0.628	s_gaed 0.610 (-0.010)
s_gaed_v3 s2	0.104 (-0.408)	0.177 (-0.513)	0.780 (-0.080)	0.754 (-0.008)	0.586 (+0.010)	0.646	–
s_gaed_v3 s3	0.494 (+0.022)	0.719 (+0.076)	0.836 (-0.040)	0.744 (-0.006)	0.572 (-0.016)	0.788	–
j_hc_v3	0.488 (+0.306)	0.573 (+0.473)	0.878 (+0.046)	0.810 (+0.028)	0.642 (-0.024)	0.740	j_hc 0.682 (-0.040)

Step-credit evaluation (ProcessBench / PRMBench)

Each arm's step signal scored as an error detector (tdprm/prmbench, definitions in tdprm/prmbench/score.py). Queue: $TDPRM_HOME/queue/prmbench/plan.json; results: $TDPRM_HOME/prmbench/results/ and /home/colligo/tdprm_prmbench_report.md.

job	status	GPU	time	progress
prmb_answers	done	[0]	0.06 h	done
prmb_spectral_qwen1.5b	done	[1]	0.30 h	done
prmb_spectral_llama3b	done	[2]	0.31 h	done
prmb_jev	done	[4]	1.05 h	done
prmb_belief_processbench	done	[5]	0.24 h	done
prmb_belief_prmbench	done	[7]	0.32 h	done
prmb_prm_qwenprm	done	[0]	0.29 h	done
prmb_prm_prm800k	done	[3]	0.29 h	done
prmb_prm_versaprm	done	[6]	0.34 h	done
prmb_score	done	[]	0.04 h	done

Generated by python -m tdprm.prmbench.score (definitions: tdprm/prmbench/score.py docstring).

• OF (outcome-free): the arm's step value without the outcome, the same information a PRM gets. value = P(step correct) for PRMs, the Bellman value V(z_t) of the state after step t for latent arms (the run's final transition table); td = its change at step t.
• F (training-faithful): the step term the arm's training channel computes given the verified outcome R and the gold answer. v2 terms use an open gate (w = 1); the gate weight the run ended with is in the note. F outcome R (every step scored R) is what outcome-only GRPO knows.
• tuned: threshold fit on a fixed 20% calibration split, reported on the other 80%. @0.5 (full): the published protocol, for probability-valued PRM scores only, on the full set. step AUROC: erroneous steps vs other steps (ProcessBench: first erroneous step vs the steps before it and all steps of correct samples), 80% split.
• Outcomes: ProcessBench R known for 3400 items (benchmark labels); PRMBench R known for 6211 items (source: reader). Gold answers: ProcessBench 3385, PRMBench 6211.

ProcessBench

Published (threshold 0.5): qwenprm 82.4 / 77.6 / 67.5 / 66.3, avg 73.5; prm800k 68.2 / 62.6 / 50.7 / 44.3, avg 56.5.

method	signal	GSM8K	MATH	Olympiad	Omni	avg F1 (tuned)	step AUROC	avg F1 @0.5 (full)	items	note
prm800k	OF value	73.5	67.4	54.9	52.1	62.0	0.916	56.7	3400	4 runs share it
prm800k	OF td	74.6	68.0	53.7	52.0	62.1	0.862	–	3400	4 runs share it
prm800k	OF value (official input)	71.3	66.5	56.9	55.5	62.5	0.913	56.3	3400	4 runs share it
prm800k (v1 channel)	F credit	73.5	67.4	54.9	52.1	62.0	0.916	–	3400	2 runs share it
qwenprm	OF value	82.9	76.9	67.1	64.6	72.9	0.953	73.4	3400	4 runs share it
qwenprm	OF td	82.5	76.5	66.4	64.6	72.5	0.900	–	3400	4 runs share it
qwenprm	OF value (official input)	80.9	77.3	66.6	66.3	72.8	0.953	73.0	3400	4 runs share it
qwenprm (v1 channel)	F credit	82.9	76.9	67.1	64.6	72.9	0.953	–	3400	2 runs share it
versaprm	OF value	66.7	52.0	36.2	39.5	48.6	0.864	37.1	3400	3 runs share it
versaprm	OF td	60.4	49.8	43.0	41.7	48.7	0.824	–	3400	3 runs share it
versaprm	OF value (official input)	68.4	52.3	36.1	40.2	49.3	0.868	35.9	3400	3 runs share it
versaprm (v1 channel)	F credit	66.7	52.0	36.2	39.5	48.6	0.864	–	3400	2 runs share it
llama3b j_gae s1	OF value	10.8	21.1	15.2	20.4	16.9	0.535	–	3400	tables @ step 200
llama3b j_gae s1	OF td	12.9	22.6	14.9	17.4	16.9	0.494	–	3400	tables @ step 200
llama3b j_gae s1	F credit	43.0	36.2	20.4	18.8	29.6	0.774	–	3400	tables @ step 200
llama3b j_gaed s1	OF value	7.3	10.8	19.7	20.5	14.6	0.562	–	3400	tables @ step 200
llama3b j_gaed s1	OF td	11.5	21.4	15.7	17.3	16.5	0.495	–	3400	tables @ step 200
llama3b j_gaed s1	F credit	43.8	35.9	21.3	17.5	29.6	0.769	–	3400	tables @ step 200
llama3b s_gae s1	OF value	7.3	17.9	20.2	23.8	17.3	0.590	–	3400	tables @ step 200
llama3b s_gae s1	OF td	11.9	22.0	22.5	27.2	20.9	0.667	–	3400	tables @ step 200
llama3b s_gae s1	F credit	37.3	29.2	27.1	20.8	28.6	0.777	–	3400	tables @ step 200
llama3b s_gaed s1	OF value	11.9	14.0	8.9	7.9	10.7	0.420	–	3400	tables @ step 200
llama3b s_gaed s1	OF td	1.3	7.1	13.2	15.2	9.2	0.407	–	3400	tables @ step 200
llama3b s_gaed s1	F credit	39.8	32.9	24.5	18.6	29.0	0.755	–	3400	tables @ step 200
qwen1.5b j_gae s1	OF value	7.3	11.9	19.7	21.6	15.1	0.543	–	3400	tables @ step 200
qwen1.5b j_gae s1	OF td	14.1	16.5	16.9	18.5	16.5	0.509	–	3400	tables @ step 200
qwen1.5b j_gae s1	F credit	38.1	30.8	25.4	16.9	27.8	0.774	–	3400	tables @ step 200
qwen1.5b j_gaed s1	OF value	7.3	11.9	19.7	21.6	15.1	0.543	–	3400	tables @ step 200
qwen1.5b j_gaed s1	OF td	13.6	14.4	19.3	18.9	16.5	0.503	–	3400	tables @ step 200
qwen1.5b j_gaed s1	F credit	46.8	33.2	21.3	16.3	29.4	0.759	–	3400	tables @ step 200
qwen1.5b s_gae s1	OF value	17.9	24.1	22.5	25.9	22.6	0.615	–	3400	tables @ step 200
qwen1.5b s_gae s1	OF td	28.4	27.0	23.5	28.9	27.0	0.660	–	3400	tables @ step 200
qwen1.5b s_gae s1	F credit	39.0	30.5	19.5	18.0	26.7	0.740	–	3400	tables @ step 200
qwen1.5b s_gaed s1	OF value	17.9	24.1	22.5	25.9	22.6	0.615	–	3400	tables @ step 200
qwen1.5b s_gaed s1	OF td	28.4	26.7	23.4	28.6	26.8	0.662	–	3400	tables @ step 200
qwen1.5b s_gaed s1	F credit	40.6	34.7	26.6	27.5	32.3	0.739	–	3400	tables @ step 200
llama3b s_gae_v2 s0 (pilot, 10 rounds)	OF value	7.3	17.9	20.2	23.8	17.3	0.571	–	3400	final gate 0.00; tables @ step 10
llama3b s_gae_v2 s0 (pilot, 10 rounds)	OF td	11.9	22.0	22.5	27.2	20.9	0.640	–	3400	final gate 0.00; tables @ step 10
llama3b s_gae_v2 s0 (pilot, 10 rounds)	F credit	20.9	15.9	12.2	17.6	16.7	0.576	–	3400	final gate 0.00; tables @ step 10
llama3b s_gae_v2 s1	OF value	7.3	17.9	20.2	23.8	17.3	0.571	–	3400	final gate 0.00; tables @ step 200
llama3b s_gae_v2 s1	OF td	11.9	22.0	22.5	27.2	20.9	0.639	–	3400	final gate 0.00; tables @ step 200
llama3b s_gae_v2 s1	F credit	23.5	19.1	1.4	11.8	14.0	0.543	–	3400	final gate 0.00; tables @ step 200
qwen1.5b j_gae_v2 s0 (pilot, 10 rounds)	OF value	7.3	11.9	19.7	21.6	15.1	0.566	–	3400	final gate 0.00; tables @ step 10
qwen1.5b j_gae_v2 s0 (pilot, 10 rounds)	OF td	9.4	16.4	21.3	21.3	17.1	0.512	–	3400	final gate 0.00; tables @ step 10
qwen1.5b j_gae_v2 s0 (pilot, 10 rounds)	F credit	10.5	18.0	17.0	21.7	16.8	0.485	–	3400	final gate 0.00; tables @ step 10
qwen1.5b j_gae_v2 s1	OF value	7.3	11.9	19.7	21.6	15.1	0.539	–	3400	final gate 0.00; tables @ step 200
qwen1.5b j_gae_v2 s1	OF td	13.6	14.4	19.4	19.0	16.6	0.508	–	3400	final gate 0.00; tables @ step 200
qwen1.5b j_gae_v2 s1	F credit	10.9	18.9	13.4	14.1	14.3	0.493	–	3400	final gate 0.00; tables @ step 200
qwen1.5b j_gaed_v2 s1	OF value	9.6	20.5	15.6	20.3	16.5	0.514	–	3400	final gate 0.00; tables @ step 200
qwen1.5b j_gaed_v2 s1	OF td	13.9	20.0	16.2	18.0	17.0	0.506	–	3400	final gate 0.00; tables @ step 200
qwen1.5b j_gaed_v2 s1	F credit	16.1	17.6	14.8	16.2	16.2	0.491	–	3400	final gate 0.00; tables @ step 200
qwen1.5b s_gae_v2 s0 (pilot, 10 rounds)	OF value	17.9	24.1	22.5	25.9	22.6	0.615	–	3400	final gate 0.00; tables @ step 10
qwen1.5b s_gae_v2 s0 (pilot, 10 rounds)	OF td	22.3	20.6	27.3	31.1	25.4	0.674	–	3400	final gate 0.00; tables @ step 10
qwen1.5b s_gae_v2 s0 (pilot, 10 rounds)	F credit	15.7	16.6	24.0	23.8	20.0	0.610	–	3400	final gate 0.00; tables @ step 10
qwen1.5b s_gae_v2 s1	OF value	17.9	24.1	22.5	25.9	22.6	0.615	–	3400	final gate 0.00; tables @ step 200
qwen1.5b s_gae_v2 s1	OF td	28.4	27.0	23.5	28.9	27.0	0.657	–	3400	final gate 0.00; tables @ step 200
qwen1.5b s_gae_v2 s1	F credit	21.0	21.1	15.8	23.0	20.2	0.597	–	3400	final gate 0.00; tables @ step 200
qwen1.5b s_gae_v2_nogate s1	OF value	17.9	24.1	22.5	25.9	22.6	0.615	–	3400	final gate 1.00; tables @ step 200
qwen1.5b s_gae_v2_nogate s1	OF td	28.4	27.0	23.5	28.9	27.0	0.662	–	3400	final gate 1.00; tables @ step 200
qwen1.5b s_gae_v2_nogate s1	F credit	18.6	16.4	19.3	19.1	18.3	0.605	–	3400	final gate 1.00; tables @ step 200
qwen1.5b s_gaed_v2 s1	OF value	17.9	24.1	22.5	25.9	22.6	0.615	–	3400	final gate 0.00; tables @ step 200
qwen1.5b s_gaed_v2 s1	OF td	28.4	26.7	23.4	28.6	26.8	0.667	–	3400	final gate 0.00; tables @ step 200
qwen1.5b s_gaed_v2 s1	F credit	23.5	23.0	18.0	21.1	21.4	0.596	–	3400	final gate 0.00; tables @ step 200
qwenprm_v2 llama3b	F credit	68.0	56.2	51.6	50.2	56.5	0.820	–	3400	final gate 0.00
prm800k_v2 qwen1.5b	F credit	51.0	39.0	31.8	36.9	39.7	0.774	–	3400	final gate 0.02; 2 runs share it
qwenprm_v2 qwen1.5b	F credit	69.1	57.5	51.6	49.1	56.9	0.821	–	3400	final gate 0.18
versaprm_v2 qwen1.5b	F credit	54.4	36.1	21.9	29.5	35.5	0.738	–	3400	final gate 0.00
b_td_v3 llama3b	F credit	31.8	25.5	18.7	17.0	23.2	0.583	–	3385	2 runs share it
b_td_v3 llama3b	F belief level	31.6	25.9	12.2	16.0	21.4	0.594	–	3385	2 runs share it
b_td_v3 qwen1.5b	F credit	31.3	25.0	18.7	17.3	23.1	0.583	–	3385	5 runs share it
b_td_v3 qwen1.5b	F belief level	31.6	25.9	12.2	16.0	21.4	0.594	–	3385	5 runs share it
outcome R (grpo)	F outcome R	25.6	25.0	10.7	11.7	18.2	0.715	–	3400	7 runs share it

PRMBench

Published (threshold 0.5): qwenprm 65.5.

method	signal	PRMScore (tuned)	F1 / neg-F1	step AUROC	PRMScore @0.5 (full)	items	note
prm800k	OF value	67.5	88.2 / 46.9	0.803	59.8	6211	4 runs share it
prm800k	OF td	65.8	89.4 / 42.2	0.643	–	6211	4 runs share it
prm800k	OF value (official input)	66.6	87.9 / 45.4	0.794	58.3	6211	4 runs share it
prm800k (v1 channel)	F credit	67.5	88.2 / 46.9	0.803	–	6211	2 runs share it
qwenprm	OF value	68.7	88.6 / 48.8	0.797	66.7	6211	4 runs share it
qwenprm	OF td	67.4	90.1 / 44.7	0.642	–	6211	4 runs share it
qwenprm	OF value (official input)	68.3	88.9 / 47.6	0.799	65.4	6211	4 runs share it
qwenprm (v1 channel)	F credit	68.7	88.6 / 48.8	0.797	–	6211	2 runs share it
versaprm	OF value	67.2	90.8 / 43.5	0.783	65.7	6211	3 runs share it
versaprm	OF td	66.2	90.8 / 41.5	0.630	–	6211	3 runs share it
versaprm	OF value (official input)	67.3	90.0 / 44.7	0.782	65.3	6211	3 runs share it
versaprm (v1 channel)	F credit	67.2	90.8 / 43.5	0.783	–	6211	2 runs share it
llama3b j_gae s1	OF value	48.6	85.3 / 11.9	0.438	–	6211	tables @ step 200
llama3b j_gae s1	OF td	51.2	86.4 / 16.0	0.448	–	6211	tables @ step 200
llama3b j_gae s1	F credit	61.7	87.8 / 35.7	0.564	–	6211	tables @ step 200
llama3b j_gaed s1	OF value	52.2	82.8 / 21.7	0.474	–	6211	tables @ step 200
llama3b j_gaed s1	OF td	51.2	84.3 / 18.2	0.468	–	6211	tables @ step 200
llama3b j_gaed s1	F credit	61.8	87.8 / 35.7	0.572	–	6211	tables @ step 200
llama3b s_gae s1	OF value	46.4	84.2 / 8.7	0.512	–	6211	tables @ step 200
llama3b s_gae s1	OF td	53.8	82.5 / 25.2	0.572	–	6211	tables @ step 200
llama3b s_gae s1	F credit	61.4	88.8 / 34.0	0.583	–	6211	tables @ step 200
llama3b s_gaed s1	OF value	47.6	88.2 / 7.0	0.466	–	6211	tables @ step 200
llama3b s_gaed s1	OF td	46.4	90.8 / 1.9	0.487	–	6211	tables @ step 200
llama3b s_gaed s1	F credit	61.8	87.1 / 36.5	0.582	–	6211	tables @ step 200
qwen1.5b j_gae s1	OF value	52.1	82.1 / 22.0	0.534	–	6211	tables @ step 200
qwen1.5b j_gae s1	OF td	52.7	84.6 / 20.8	0.509	–	6211	tables @ step 200
qwen1.5b j_gae s1	F credit	61.7	87.9 / 35.5	0.566	–	6211	tables @ step 200
qwen1.5b j_gaed s1	OF value	52.1	82.1 / 22.0	0.523	–	6211	tables @ step 200
qwen1.5b j_gaed s1	OF td	53.0	83.4 / 22.6	0.518	–	6211	tables @ step 200
qwen1.5b j_gaed s1	F credit	61.7	87.9 / 35.5	0.568	–	6211	tables @ step 200
qwen1.5b s_gae s1	OF value	48.5	86.4 / 10.7	0.503	–	6211	tables @ step 200
qwen1.5b s_gae s1	OF td	56.2	84.6 / 27.8	0.585	–	6211	tables @ step 200
qwen1.5b s_gae s1	F credit	58.0	89.8 / 26.2	0.560	–	6211	tables @ step 200
qwen1.5b s_gaed s1	OF value	46.6	84.9 / 8.2	0.499	–	6211	tables @ step 200
qwen1.5b s_gaed s1	OF td	55.9	84.0 / 27.9	0.586	–	6211	tables @ step 200
qwen1.5b s_gaed s1	F credit	58.2	89.4 / 27.0	0.569	–	6211	tables @ step 200
llama3b s_gae_v2 s0 (pilot, 10 rounds)	OF value	46.4	84.2 / 8.7	0.464	–	6211	final gate 0.00; tables @ step 10
llama3b s_gae_v2 s0 (pilot, 10 rounds)	OF td	53.4	85.8 / 21.1	0.550	–	6211	final gate 0.00; tables @ step 10
llama3b s_gae_v2 s0 (pilot, 10 rounds)	F credit	54.0	87.5 / 20.5	0.555	–	6211	final gate 0.00; tables @ step 10
llama3b s_gae_v2 s1	OF value	46.4	84.2 / 8.7	0.464	–	6211	final gate 0.00; tables @ step 200
llama3b s_gae_v2 s1	OF td	53.4	85.8 / 21.1	0.550	–	6211	final gate 0.00; tables @ step 200
llama3b s_gae_v2 s1	F credit	54.3	87.9 / 20.7	0.558	–	6211	final gate 0.00; tables @ step 200
qwen1.5b j_gae_v2 s0 (pilot, 10 rounds)	OF value	52.1	82.1 / 22.0	0.479	–	6211	final gate 0.00; tables @ step 10
qwen1.5b j_gae_v2 s0 (pilot, 10 rounds)	OF td	52.4	86.3 / 18.5	0.491	–	6211	final gate 0.00; tables @ step 10
qwen1.5b j_gae_v2 s0 (pilot, 10 rounds)	F credit	52.1	87.8 / 16.5	0.485	–	6211	final gate 0.00; tables @ step 10
qwen1.5b j_gae_v2 s1	OF value	52.1	82.1 / 22.0	0.534	–	6211	final gate 0.00; tables @ step 200
qwen1.5b j_gae_v2 s1	OF td	52.3	85.5 / 19.0	0.512	–	6211	final gate 0.00; tables @ step 200
qwen1.5b j_gae_v2 s1	F credit	51.7	84.8 / 18.6	0.507	–	6211	final gate 0.00; tables @ step 200
qwen1.5b j_gaed_v2 s1	OF value	48.7	84.8 / 12.6	0.442	–	6211	final gate 0.00; tables @ step 200
qwen1.5b j_gaed_v2 s1	OF td	50.5	87.1 / 14.0	0.469	–	6211	final gate 0.00; tables @ step 200
qwen1.5b j_gaed_v2 s1	F credit	50.6	86.4 / 14.8	0.480	–	6211	final gate 0.00; tables @ step 200
qwen1.5b s_gae_v2 s0 (pilot, 10 rounds)	OF value	46.6	84.9 / 8.2	0.499	–	6211	final gate 0.00; tables @ step 10
qwen1.5b s_gae_v2 s0 (pilot, 10 rounds)	OF td	55.9	84.0 / 27.9	0.592	–	6211	final gate 0.00; tables @ step 10
qwen1.5b s_gae_v2 s0 (pilot, 10 rounds)	F credit	57.0	87.9 / 26.1	0.600	–	6211	final gate 0.00; tables @ step 10
qwen1.5b s_gae_v2 s1	OF value	48.5	86.4 / 10.7	0.503	–	6211	final gate 0.00; tables @ step 200
qwen1.5b s_gae_v2 s1	OF td	56.2	84.6 / 27.8	0.585	–	6211	final gate 0.00; tables @ step 200
qwen1.5b s_gae_v2 s1	F credit	56.5	86.9 / 26.1	0.584	–	6211	final gate 0.00; tables @ step 200
qwen1.5b s_gae_v2_nogate s1	OF value	48.5	86.4 / 10.7	0.503	–	6211	final gate 1.00; tables @ step 200
qwen1.5b s_gae_v2_nogate s1	OF td	56.2	84.6 / 27.8	0.590	–	6211	final gate 1.00; tables @ step 200
qwen1.5b s_gae_v2_nogate s1	F credit	57.1	87.5 / 26.6	0.598	–	6211	final gate 1.00; tables @ step 200
qwen1.5b s_gaed_v2 s1	OF value	46.6	84.9 / 8.2	0.499	–	6211	final gate 0.00; tables @ step 200
qwen1.5b s_gaed_v2 s1	OF td	56.2	84.6 / 27.8	0.587	–	6211	final gate 0.00; tables @ step 200
qwen1.5b s_gaed_v2 s1	F credit	55.8	85.5 / 26.1	0.600	–	6211	final gate 0.00; tables @ step 200
qwenprm_v2 llama3b	F credit	66.6	90.1 / 43.1	0.628	–	6211	final gate 0.00
prm800k_v2 qwen1.5b	F credit	64.2	88.0 / 40.5	0.631	–	6211	final gate 0.02; 2 runs share it
qwenprm_v2 qwen1.5b	F credit	66.6	90.0 / 43.2	0.628	–	6211	final gate 0.18
versaprm_v2 qwen1.5b	F credit	64.6	88.1 / 41.1	0.624	–	6211	final gate 0.00
b_td_v3 llama3b	F credit	55.0	88.0 / 22.0	0.543	–	6211	2 runs share it
b_td_v3 llama3b	F belief level	50.5	84.1 / 17.0	0.487	–	6211	2 runs share it
b_td_v3 qwen1.5b	F credit	55.1	86.9 / 23.2	0.543	–	6211	5 runs share it
b_td_v3 qwen1.5b	F belief level	50.5	84.1 / 17.0	0.487	–	6211	5 runs share it
outcome R (grpo)	F outcome R	45.7	64.6 / 26.9	0.535	–	6211	7 runs share it

PRMBench by error category (PRMScore, tuned threshold)

method	signal	circular	confidence	counterfactual	deception	domain_inconsistency	missing_condition	multi_solutions	redundency	step_contradiction
prm800k	OF value	67.3	75.0	71.9	69.7	65.1	63.6	48.2	54.7	66.8
prm800k	OF td	66.3	71.3	66.6	66.4	68.0	62.1	48.3	54.4	68.6
prm800k	OF value (official input)	64.8	74.1	71.7	68.9	64.0	63.3	47.5	54.6	66.4
prm800k (v1 channel)	F credit	67.3	75.0	71.9	69.7	65.1	63.6	48.2	54.7	66.8
qwenprm	OF value	66.5	76.7	73.7	70.5	67.9	64.2	49.1	54.1	68.0
qwenprm	OF td	65.3	74.5	69.1	67.8	70.9	62.6	49.0	53.7	70.9
qwenprm	OF value (official input)	64.6	77.0	73.7	71.0	66.8	63.4	49.0	52.6	68.1
qwenprm (v1 channel)	F credit	66.5	76.7	73.7	70.5	67.9	64.2	49.1	54.1	68.0
versaprm	OF value	65.8	75.5	69.8	65.1	70.8	59.6	49.4	51.2	73.4
versaprm	OF td	65.5	72.4	66.1	64.4	71.1	59.8	49.2	53.1	73.5
versaprm	OF value (official input)	66.2	74.6	70.4	65.5	70.1	61.2	49.3	52.5	71.9
versaprm (v1 channel)	F credit	65.8	75.5	69.8	65.1	70.8	59.6	49.4	51.2	73.4
llama3b j_gae s1	OF value	43.0	45.8	47.9	50.6	48.8	48.6	47.9	49.1	52.3
llama3b j_gae s1	OF td	51.0	48.1	50.9	49.0	53.3	51.8	46.8	52.1	51.8
llama3b j_gae s1	F credit	49.8	69.8	67.7	65.7	57.1	58.8	49.9	48.8	60.5
llama3b j_gaed s1	OF value	57.6	48.3	50.5	48.0	51.3	55.7	45.2	52.9	49.6
llama3b j_gaed s1	OF td	54.0	47.4	49.6	46.9	53.5	53.5	45.8	52.7	49.9
llama3b j_gaed s1	F credit	49.7	69.8	67.8	65.7	56.9	58.9	49.9	48.8	60.4
llama3b s_gae s1	OF value	42.6	48.0	45.3	47.1	45.4	45.9	43.3	49.0	47.5
llama3b s_gae s1	OF td	55.9	52.5	53.3	53.3	54.4	54.6	41.0	55.3	52.1
llama3b s_gae s1	F credit	49.1	69.0	67.0	65.2	58.0	59.1	49.9	48.6	61.2
llama3b s_gaed s1	OF value	43.4	49.9	46.7	48.0	46.7	46.6	46.6	49.0	49.1
llama3b s_gaed s1	OF td	42.8	48.0	45.0	45.9	47.0	45.4	49.2	47.2	48.0
llama3b s_gaed s1	F credit	50.3	67.8	67.8	65.3	57.7	59.1	49.9	48.9	60.3
qwen1.5b j_gae s1	OF value	57.9	47.8	50.3	48.0	51.2	55.5	45.0	52.7	49.2
qwen1.5b j_gae s1	OF td	56.3	52.6	52.7	50.6	52.0	54.7	44.8	51.4	49.3
qwen1.5b j_gae s1	F credit	49.8	70.0	67.5	65.6	56.8	58.8	49.9	48.8	60.5
qwen1.5b j_gaed s1	OF value	57.9	47.8	50.3	48.0	51.2	55.5	45.0	52.7	49.2
qwen1.5b j_gaed s1	OF td	57.4	52.9	53.0	51.0	53.4	54.7	44.3	51.1	49.3
qwen1.5b j_gaed s1	F credit	49.8	70.0	67.5	65.6	56.8	58.9	49.9	48.8	60.5
qwen1.5b s_gae s1	OF value	43.3	51.3	50.0	49.5	46.3	48.0	45.6	48.5	49.9
qwen1.5b s_gae s1	OF td	55.0	55.8	56.4	55.5	58.1	56.7	43.1	56.2	56.2
qwen1.5b s_gae s1	F credit	48.1	59.4	61.0	60.3	61.3	57.1	49.9	47.6	60.5
qwen1.5b s_gaed s1	OF value	42.3	48.5	45.6	47.3	45.6	45.9	43.9	49.0	47.6
qwen1.5b s_gaed s1	OF td	55.0	55.6	56.5	55.2	57.4	56.7	42.1	55.8	55.9
qwen1.5b s_gaed s1	F credit	48.3	58.7	61.0	60.4	61.9	58.0	49.9	47.6	60.7
llama3b s_gae_v2 s0 (pilot, 10 rounds)	OF value	42.6	48.0	45.3	47.1	45.4	45.9	43.3	49.0	47.5
llama3b s_gae_v2 s0 (pilot, 10 rounds)	OF td	50.0	54.9	54.1	53.8	53.0	53.0	43.4	55.5	53.2
llama3b s_gae_v2 s0 (pilot, 10 rounds)	F credit	49.5	56.9	54.6	54.6	52.8	53.0	45.2	55.7	54.4
llama3b s_gae_v2 s1	OF value	42.6	48.0	45.3	47.1	45.4	45.9	43.3	49.0	47.5
llama3b s_gae_v2 s1	OF td	50.0	54.9	54.1	53.8	53.0	53.0	43.4	55.5	53.2
llama3b s_gae_v2 s1	F credit	49.7	57.4	54.6	54.7	53.2	53.2	45.9	55.8	55.0
qwen1.5b j_gae_v2 s0 (pilot, 10 rounds)	OF value	57.9	47.8	50.3	48.0	51.2	55.5	45.0	52.7	49.2
qwen1.5b j_gae_v2 s0 (pilot, 10 rounds)	OF td	56.1	48.8	50.9	47.6	53.8	54.7	46.3	54.3	50.3
qwen1.5b j_gae_v2 s0 (pilot, 10 rounds)	F credit	54.4	48.8	51.3	48.0	54.7	52.1	45.9	54.0	50.5
qwen1.5b j_gae_v2 s1	OF value	57.9	47.8	50.3	48.0	51.2	55.5	45.0	52.7	49.2
qwen1.5b j_gae_v2 s1	OF td	55.1	52.1	51.7	50.1	52.1	54.3	45.9	51.2	49.1
qwen1.5b j_gae_v2 s1	F credit	54.0	51.8	51.4	50.1	53.0	52.1	44.1	49.7	50.0
qwen1.5b j_gaed_v2 s1	OF value	43.3	45.5	47.9	50.5	49.6	48.7	47.7	49.2	52.1
qwen1.5b j_gaed_v2 s1	OF td	48.7	47.6	50.5	49.1	53.3	50.9	47.1	50.6	51.7
qwen1.5b j_gaed_v2 s1	F credit	50.6	48.5	49.3	49.4	52.5	49.8	45.1	51.0	50.2
qwen1.5b s_gae_v2 s0 (pilot, 10 rounds)	OF value	42.3	48.5	45.6	47.3	45.6	45.9	43.9	49.0	47.6
qwen1.5b s_gae_v2 s0 (pilot, 10 rounds)	OF td	55.0	55.6	56.5	55.2	57.4	56.7	42.1	55.8	55.9
qwen1.5b s_gae_v2 s0 (pilot, 10 rounds)	F credit	54.2	59.0	57.3	56.9	58.2	55.4	45.2	56.9	57.7
qwen1.5b s_gae_v2 s1	OF value	43.3	51.3	50.0	49.5	46.3	48.0	45.6	48.5	49.9
qwen1.5b s_gae_v2 s1	OF td	55.0	55.8	56.4	55.5	58.1	56.7	43.1	56.2	56.2
qwen1.5b s_gae_v2 s1	F credit	53.0	58.6	57.1	56.1	56.6	55.9	45.0	56.9	57.5
qwen1.5b s_gae_v2_nogate s1	OF value	43.3	51.3	50.0	49.5	46.3	48.0	45.6	48.5	49.9
qwen1.5b s_gae_v2_nogate s1	OF td	55.0	55.8	56.4	55.5	58.1	56.7	43.1	56.2	56.2
qwen1.5b s_gae_v2_nogate s1	F credit	53.8	59.2	57.5	56.8	57.6	55.9	45.3	57.2	58.3
qwen1.5b s_gaed_v2 s1	OF value	42.3	48.5	45.6	47.3	45.6	45.9	43.9	49.0	47.6
qwen1.5b s_gaed_v2 s1	OF td	55.0	55.8	56.4	55.5	58.1	56.7	43.1	56.2	56.2
qwen1.5b s_gaed_v2 s1	F credit	54.2	57.3	56.4	54.1	55.5	54.4	44.8	55.2	54.9
qwenprm_v2 llama3b	F credit	63.8	73.6	68.3	67.4	70.7	60.9	48.3	54.1	71.2
prm800k_v2 qwen1.5b	F credit	65.2	68.0	65.4	65.2	66.3	60.2	46.2	55.7	67.2
qwenprm_v2 qwen1.5b	F credit	63.9	73.6	68.5	67.2	70.6	61.1	48.3	54.2	71.1
versaprm_v2 qwen1.5b	F credit	64.7	69.0	64.8	63.9	67.0	60.5	46.4	55.6	71.0
b_td_v3 llama3b	F credit	45.9	63.7	58.7	58.3	52.3	52.2	48.5	49.5	54.3
b_td_v3 llama3b	F belief level	44.8	53.1	50.9	53.9	48.4	51.3	47.3	48.8	49.0
b_td_v3 qwen1.5b	F credit	46.4	62.6	58.5	58.5	52.6	52.9	48.1	50.6	53.7
b_td_v3 qwen1.5b	F belief level	44.8	53.1	50.9	53.9	48.4	51.3	47.3	48.8	49.0
outcome R (grpo)	F outcome R	49.1	33.6	39.1	43.8	44.3	47.6	49.9	49.1	37.0

Not evaluated

run	reason
llama3b-math-j_hc-s1	HC fit (h, C) is refit online and was not saved
llama3b-math-s_hc-s1	HC fit (h, C) is refit online and was not saved
llama3b-math-s_hc_v3-s1	HC fit (h, C) is refit online and was not saved
llama3b-math-vp_gae_v2-s1	value-probe states are refit online and were not saved
qwen1.5b-math-b_hc_v3-s1	HC fit (h, C) is refit online and was not saved
qwen1.5b-math-j_hc-s1	HC fit (h, C) is refit online and was not saved
qwen1.5b-math-j_hc_v2-s1	HC fit (h, C) is refit online and was not saved
qwen1.5b-math-j_hc_v3-s1	HC fit (h, C) is refit online and was not saved
qwen1.5b-math-s_gaed_v3-s1	v3 latent run without saved transition tables
qwen1.5b-math-s_hc-s1	HC fit (h, C) is refit online and was not saved
qwen1.5b-math-s_hc_v2-s1	HC fit (h, C) is refit online and was not saved
qwen1.5b-math-s_hc_v3-s0	HC fit (h, C) is refit online and was not saved
qwen1.5b-math-s_hc_v3-s1	HC fit (h, C) is refit online and was not saved
qwen1.5b-math-s_hc_v3-s2	HC fit (h, C) is refit online and was not saved
qwen1.5b-math-s_hc_v3-s3	HC fit (h, C) is refit online and was not saved
qwen1.5b-math-vp_gae_v2-s0	value-probe states are refit online and were not saved
qwen1.5b-math-vp_gae_v2-s1	value-probe states are refit online and were not saved
qwen1.5b-math-vp_gae_v2_nogate-s1	value-probe states are refit online and were not saved
qwen1.5b-math-vp_gaed_v2-s1	value-probe states are refit online and were not saved
qwen1.5b-math-vp_hc_v2-s1	HC fit (h, C) is refit online and was not saved

In-loop eval (greedy, in-domain eval set)

v1 methods

run	r25	r50	r75	r100	r125	r150	r175	r200	best
llama3b-math-grpo-s1	0.476	0.502	0.512	0.502	0.534	0.532	0.524	0.528	0.534 (r125)
llama3b-math-s_hc-s1	0.458	0.508	0.488	0.484	0.476	0.472	0.498	0.496	0.508 (r50)
llama3b-math-s_gae-s1	0.482	0.510	0.522	0.512	0.520	0.502	0.512	0.504	0.522 (r75)
llama3b-math-s_gaed-s1	0.496	0.506	0.482	0.484	0.510	0.514	0.518	0.492	0.518 (r175)
llama3b-math-j_hc-s1	0.512	0.440	0.482	0.484	0.506	0.502	0.502	0.496	0.512 (r25)
llama3b-math-j_gae-s1	0.458	0.496	0.496	0.510	0.524	0.504	0.514	0.526	0.526 (r200)
llama3b-math-j_gaed-s1	0.488	0.500	0.512	0.512	0.504	0.528	0.518	0.542	0.542 (r200)
llama3b-math-prm800k-s1	0.494	0.398	0.308	0.248	0.222	0.240	0.230	0.248	0.494 (r25)
llama3b-math-qwenprm-s1	0.514	0.428	0.408	0.350	0.340	0.418	0.466	0.444	0.514 (r25)
llama3b-math-versaprm-s1	0.492	0.384	0.414	0.338	0.354	0.466	0.416	0.396	0.492 (r25)
llama3b-proverqa-grpo-s1	0.684	0.778	0.818	0.868	0.856	0.886	0.908	0.890	0.908 (r175)
llama3b-proverqa-s_hc-s1	0.710	0.722	0.756	0.776	0.812	0.824	0.844	0.838	0.844 (r175)
llama3b-proverqa-s_gae-s1	0.712	0.788	0.812	0.864	0.888	0.822	0.868	0.876	0.888 (r125)
llama3b-proverqa-s_gaed-s1	0.718	0.528	0.778	0.794	0.848	0.862	0.860	0.866	0.866 (r200)
llama3b-proverqa-j_hc-s1	0.672	0.768	0.804	0.818	0.820	0.846	0.854	0.868	0.868 (r200)
llama3b-proverqa-j_gae-s1	0.716	0.800	0.844	0.870	0.884	0.890	0.906	0.886	0.906 (r175)
llama3b-proverqa-j_gaed-s1	0.720	0.778	0.814	0.852	0.888	0.894	0.882	0.908	0.908 (r200)
llama3b-proverqa-prm800k-s1	0.632	0.634	0.706	0.732	0.770	0.782	0.796	0.790	0.796 (r175)
llama3b-proverqa-qwenprm-s1	0.640	0.666	0.772	0.738	0.794	0.788	0.604	0.780	0.794 (r125)
llama3b-proverqa-versaprm-s1	0.666	0.704	0.736	0.752	0.774	0.732	0.718	0.754	0.774 (r125)
qwen1.5b-math-grpo-s1	0.538	0.562	0.582	0.578	0.580	0.568	0.586	0.570	0.586 (r175)
qwen1.5b-math-grpo-s2	0.564	0.576	0.570	0.574	0.562	0.592	0.620	0.576	0.620 (r175)
qwen1.5b-math-grpo-s3	0.558	0.564	0.578	0.584	0.574	0.578	0.560	0.590	0.590 (r200)
qwen1.5b-math-s_hc-s1	0.512	0.490	0.488	0.470	0.490	0.468	0.484	0.442	0.512 (r25)
qwen1.5b-math-s_gae-s1	0.560	0.574	0.582	0.570	0.566	0.570	0.580	0.598	0.598 (r200)
qwen1.5b-math-s_gaed-s1	0.562	0.542	0.566	0.554	0.560	0.550	0.556	0.548	0.566 (r75)
qwen1.5b-math-j_hc-s1	0.540	0.560	0.512	0.540	0.530	0.558	0.560	0.568	0.568 (r200)
qwen1.5b-math-j_gae-s1	0.558	0.576	0.590	0.570	0.586	0.576	0.552	0.566	0.590 (r75)
qwen1.5b-math-j_gaed-s1	0.568	0.570	0.562	0.560	0.586	0.582	0.576	0.552	0.586 (r125)
qwen1.5b-math-prm800k-s1	0.508	0.486	0.300	0.476	0.494	0.540	0.554	0.576	0.576 (r200)
qwen1.5b-math-qwenprm-s1	0.550	0.520	0.472	0.416	0.270	0.244	0.356	0.360	0.550 (r25)
qwen1.5b-math-versaprm-s1	0.516	0.488	0.384	0.380	0.286	0.272	0.436	0.428	0.516 (r25)
qwen1.5b-proverqa-grpo-s1	0.600	0.630	0.684	0.702	0.748	0.760	0.746	0.748	0.760 (r150)
qwen1.5b-proverqa-grpo-s2	0.602	0.646	0.674	0.698	0.724	0.728	0.750	0.744	0.750 (r175)
qwen1.5b-proverqa-grpo-s3	0.598	0.586	0.666	0.672	0.712	0.656	0.676	0.758	0.758 (r200)
qwen1.5b-proverqa-s_hc-s1	0.532	0.660	0.688	0.700	0.736	0.766	0.760	0.786	0.786 (r200)
qwen1.5b-proverqa-s_gae-s1	0.600	0.644	0.686	0.716	0.734	0.754	0.766	0.748	0.766 (r175)
qwen1.5b-proverqa-s_gaed-s1	0.578	0.580	0.592	0.632	0.692	0.726	0.782	0.742	0.782 (r175)
qwen1.5b-proverqa-j_hc-s1	0.564	0.598	0.634	0.698	0.740	0.772	0.762	0.798	0.798 (r200)
qwen1.5b-proverqa-j_gae-s1	0.620	0.648	0.670	0.696	0.736	0.754	0.772	0.780	0.780 (r200)
qwen1.5b-proverqa-j_gaed-s1	0.512	0.572	0.582	0.656	0.630	0.646	0.590	0.756	0.756 (r200)
qwen1.5b-proverqa-prm800k-s1	0.624	0.682	0.736	0.742	0.758	0.742	0.766	0.778	0.778 (r200)
qwen1.5b-proverqa-qwenprm-s1	0.628	0.672	0.712	0.740	0.760	0.746	0.754	0.762	0.762 (r200)
qwen1.5b-proverqa-versaprm-s1	0.610	0.654	0.678	0.718	0.708	0.752	0.736	0.760	0.760 (r200)

v2 methods

run	r25	r50	r75	r100	r125	r150	r175	r200	best
llama3b-math-grpo_v2-s1	0.488	0.482	0.500	0.510	0.520	0.506	0.504	0.518	0.520 (r125)
llama3b-math-s_gae_v2-s1	0.472	0.504	0.500	0.508	0.506	0.496	0.506	0.526	0.526 (r200)
llama3b-math-vp_gae_v2-s1	0.500	0.514	0.516	0.508	0.502	0.504	0.514	0.528	0.528 (r200)
llama3b-math-qwenprm_v2-s1	0.472	0.504	0.510	0.498	0.534	0.536	0.518	0.542	0.542 (r200)
qwen1.5b-math-grpo_v2-s1	0.538	0.558	0.558	0.578	0.590	0.590	0.596	0.602	0.602 (r200)
qwen1.5b-math-s_hc_v2-s1	0.550	0.562	0.572	0.590	0.592	0.600	0.604	0.594	0.604 (r175)
qwen1.5b-math-s_gae_v2-s1	0.560	0.542	0.560	0.584	0.564	0.584	0.584	0.588	0.588 (r200)
qwen1.5b-math-s_gae_v2_nogate-s1	0.522	0.536	0.552	0.570	0.584	0.564	0.574	0.572	0.584 (r125)
qwen1.5b-math-s_gaed_v2-s1	0.556	0.564	0.574	0.572	0.572	0.590	0.580	0.582	0.590 (r150)
qwen1.5b-math-j_hc_v2-s1	0.532	0.570	0.582	0.566	0.558	0.566	0.564	0.566	0.582 (r75)
qwen1.5b-math-j_gae_v2-s1	0.552	0.578	0.592	0.558	0.586	0.590	0.598	0.598	0.598 (r175)
qwen1.5b-math-j_gaed_v2-s1	0.586	0.570	0.588	0.562	0.568	0.580	0.580	0.588	0.588 (r75)
qwen1.5b-math-vp_hc_v2-s1	0.564	0.554	0.572	0.574	0.582	0.590	0.580	0.574	0.590 (r150)
qwen1.5b-math-vp_gae_v2-s1	0.578	0.562	0.570	0.584	0.586	0.598	0.574	0.578	0.598 (r150)
qwen1.5b-math-vp_gae_v2_nogate-s1	0.560	0.568	0.562	0.578	0.580	0.588	0.580	0.576	0.588 (r150)
qwen1.5b-math-vp_gaed_v2-s1	0.570	0.550	0.568	0.582	0.590	0.570	0.596	0.590	0.596 (r175)
qwen1.5b-math-prm800k_v2-s1	0.550	0.550	0.586	0.614	0.594	0.592	0.582	0.574	0.614 (r100)
qwen1.5b-math-qwenprm_v2-s1	0.562	0.552	0.576	0.592	0.560	0.558	0.584	0.568	0.592 (r100)
qwen1.5b-math-versaprm_v2-s1	0.548	0.576	0.572	0.592	0.584	0.572	0.572	0.566	0.592 (r100)
qwen1.5b-proverqa-grpo_v2-s1	0.582	0.648	0.690	0.632	0.676	0.608	0.620	0.530	0.690 (r75)
qwen1.5b-proverqa-s_hc_v2-s1	0.596	0.610	0.676	0.738	0.710	0.722	0.730	0.790	0.790 (r200)
qwen1.5b-proverqa-s_gae_v2-s1	0.594	0.588	0.664	0.676	0.744	0.764	0.770	0.770	0.770 (r175)
qwen1.5b-proverqa-s_gaed_v2-s1	0.538	0.622	0.662	0.718	0.732	0.726	0.662	0.504	0.732 (r125)
qwen1.5b-proverqa-vp_hc_v2-s1	0.550	0.628	0.666	0.724	0.746	0.700	0.666	0.556	0.746 (r125)
qwen1.5b-proverqa-vp_gae_v2-s1	0.560	0.630	0.678	0.714	0.742	0.560	0.516	0.636	0.742 (r125)
qwen1.5b-proverqa-vp_gaed_v2-s1	0.550	0.622	0.686	0.658	0.640	0.750	0.664	0.520	0.750 (r150)
qwen1.5b-proverqa-prm800k_v2-s1	0.598	0.648	0.680	0.726	0.718	0.658	0.534	0.628	0.726 (r100)
qwen1.5b-proverqa-qwenprm_v2-s1	0.574	0.610	0.654	0.710	0.750	0.718	0.776	0.808	0.808 (r200)
qwen1.5b-proverqa-versaprm_v2-s1	0.562	0.614	0.632	0.678	0.670	0.678	0.752	0.764	0.764 (r200)

v3 methods (references: the v1 grpo curves of the same cell and seed, v1 table)

run	r25	r50	r75	r100	r125	r150	r175	r200	best
llama3b-math-b_td_v3-s1	0.480	0.498	0.502	0.516	0.498	0.496	0.506	0.502	0.516 (r100)
llama3b-math-s_hc_v3-s1	0.488	0.476	0.464	0.496	0.464	0.494	0.466	0.448	0.496 (r100)
qwen1.5b-math-b_td_v3-s1	0.564	0.556	0.566	0.548	0.576	0.574	0.574	0.584	0.584 (r200)
qwen1.5b-math-b_td_v3-s2	0.548	0.578	0.572	0.556	0.570	0.578	0.578	0.566	0.578 (r50)
qwen1.5b-math-b_td_v3-s3	0.566	0.562	0.564	0.580	0.590	0.604	0.580	0.588	0.604 (r150)
qwen1.5b-math-b_td_v3_nozv-s1	0.542	0.564	0.574	0.596	0.604	0.584	0.582	0.598	0.604 (r125)
qwen1.5b-math-b_hc_v3-s1	0.558	0.566	0.562	0.556	0.564	0.530	0.554	0.562	0.566 (r50)
qwen1.5b-math-s_hc_v3-s1	0.566	0.570	0.576	0.548	0.552	0.548	0.556	0.530	0.576 (r75)
qwen1.5b-math-s_hc_v3-s2	0.562	0.574	0.564	0.558	0.574	0.556	0.534	0.550	0.574 (r50)
qwen1.5b-math-s_hc_v3-s3	0.538	0.548	0.554	0.562	0.578	0.566	0.524	0.530	0.578 (r125)
qwen1.5b-math-s_gaed_v3-s1	0.566	0.546	0.576	0.562	0.554	0.548	0.562	0.580	0.580 (r200)
qwen1.5b-math-j_hc_v3-s1	0.558	0.490	0.522	0.524	0.540	0.562	0.546	0.530	0.562 (r150)
qwen1.5b-proverqa-b_td_v3-s1	0.558	0.598	0.648	0.674	0.734	0.758	0.748	0.762	0.762 (r200)
qwen1.5b-proverqa-b_td_v3-s2	0.484	0.544	0.628	0.740	0.750	0.760	0.690	0.790	0.790 (r200)
qwen1.5b-proverqa-b_td_v3-s3	0.518	0.564	0.662	0.720	0.750	0.782	0.772	0.736	0.782 (r150)
qwen1.5b-proverqa-b_td_v3_nozv-s1	0.618	0.664	0.718	0.748	0.760	0.758	0.778	0.774	0.778 (r175)
qwen1.5b-proverqa-b_td_v3_nomask-s1	0.576	0.606	0.682	0.688	0.742	0.748	0.774	0.794	0.794 (r200)
qwen1.5b-proverqa-b_td_v3_zs-s1	0.586	0.620	0.664	0.698	0.730	0.778	0.772	0.728	0.778 (r150)
qwen1.5b-proverqa-b_hc_v3-s1	0.434	0.506	0.584	0.734	0.772	0.774	0.768	0.782	0.782 (r200)
qwen1.5b-proverqa-s_hc_v3-s1	0.618	0.640	0.672	0.706	0.734	0.726	0.750	0.754	0.754 (r200)
qwen1.5b-proverqa-s_gaed_v3-s1	0.568	0.620	0.620	0.690	0.682	0.692	0.728	0.722	0.728 (r175)
qwen1.5b-proverqa-s_gaed_v3-s2	0.550	0.638	0.624	0.686	0.050	0.392	0.406	0.696	0.696 (r200)
qwen1.5b-proverqa-s_gaed_v3-s3	0.584	0.558	0.638	0.710	0.724	0.770	0.760	0.748	0.770 (r150)
qwen1.5b-proverqa-j_hc_v3-s1	0.604	0.528	0.598	0.654	0.750	0.776	0.790	0.760	0.790 (r175)

v3 monitors (label-free; first 3 rounds → last 3 rounds)

Expected (section 6 of the v3 plan): gain correct > wrong; ceiling share ProverQA ~0.5–0.6, MATH ~0.2; escape monitor +0.01 to +0.08; first-step term 0 to +0.05; sign flips < 0.01; abs term/step token b_td_v3 0.02–0.1; |rho mass-length| < 0.3; zero-var term/token ≥ 0 under the gain gate; reader gap ProverQA −0.06 to 0; zero-step < 1%.

run	round	gain correct	gain wrong	Phi_0	ceiling share	escape monitor	first-step term	sign flips	abs term/step tok	rho mass-length	zero-var term/tok	reader gap	zero-step	reader/states s
llama3b-math-b_td_v3-s1	200	0.461 → 0.481	-0.126 → -0.189	0.39 → 0.36	0.14 → 0.17	0.076 → 0.246	0.015 → 0.175	0.007 → 0.000	0.102 → 0.228	0.05 → -0.67	0.046 → 0.252	0.003 → 0.123	0.000 → 0.000	80 → 27
llama3b-math-s_hc_v3-s1	200	–	–	–	–	0.066 → -0.025	-0.009 → 0.038	0.061 → 0.068	0.321 → 0.328	0.21 → 0.10	-0.001 → 0.109	–	0.000 → 0.000	27 → 8
qwen1.5b-math-b_td_v3-s1	200	0.409 → 0.439	-0.156 → -0.252	0.43 → 0.43	0.26 → 0.22	0.062 → 0.072	0.019 → 0.037	0.009 → 0.013	0.087 → 0.115	0.05 → 0.12	0.009 → 0.080	-0.055 → 0.079	0.000 → 0.000	66 → 52
qwen1.5b-math-b_td_v3-s2	200	0.380 → 0.429	-0.141 → -0.291	0.46 → 0.44	0.25 → 0.24	0.056 → 0.050	0.025 → 0.019	0.010 → 0.012	0.090 → 0.095	0.01 → 0.15	0.009 → 0.069	-0.092 → 0.115	0.001 → 0.000	62 → 61
qwen1.5b-math-b_td_v3-s3	200	0.307 → 0.438	-0.189 → -0.250	0.51 → 0.41	0.29 → 0.22	0.045 → 0.049	-0.014 → 0.028	0.010 → 0.007	0.087 → 0.092	0.05 → 0.14	0.011 → 0.060	-0.057 → 0.101	0.000 → 0.000	67 → 62
qwen1.5b-math-b_td_v3_nozv-s1	200	0.419 → 0.452	-0.158 → -0.240	0.43 → 0.44	0.26 → 0.22	0.062 → 0.058	0.020 → 0.038	0.008 → 0.003	0.086 → 0.041	0.04 → 0.11	0.000 → 0.000	-0.052 → 0.072	0.000 → 0.000	66 → 71
qwen1.5b-math-b_hc_v3-s1	200	0.428 → 0.442	-0.164 → -0.226	0.43 → 0.44	0.26 → 0.21	0.072 → 0.225	-0.005 → -0.073	0.024 → 0.024	0.161 → 0.342	0.26 → 0.23	-0.008 → 0.185	-0.057 → 0.070	0.000 → 0.000	67 → 17
qwen1.5b-math-s_hc_v3-s1	200	–	–	–	–	0.075 → 0.028	-0.059 → 0.062	0.082 → 0.055	0.380 → 0.249	0.03 → 0.39	-0.020 → 0.063	–	0.000 → 0.000	18 → 18
qwen1.5b-math-s_hc_v3-s2	200	–	–	–	–	0.074 → 0.023	-0.050 → 0.109	0.077 → 0.076	0.378 → 0.319	0.04 → 0.19	0.004 → 0.065	–	0.001 → 0.000	17 → 12
qwen1.5b-math-s_hc_v3-s3	200	–	–	–	–	0.074 → 0.024	-0.064 → 0.035	0.079 → 0.045	0.367 → 0.215	0.05 → 0.03	0.001 → 0.048	–	0.000 → 0.000	17 → 23
qwen1.5b-math-s_gaed_v3-s1	200	–	–	–	–	0.567 → 0.415	-0.019 → 0.046	0.000 → 0.012	0.395 → 0.262	-0.02 → 0.15	0.047 → 0.083	–	0.000 → 0.000	18 → 18
qwen1.5b-math-j_hc_v3-s1	200	–	–	–	–	0.020 → 0.012	0.059 → 0.002	0.067 → 0.025	0.302 → 0.090	0.03 → 0.35	-0.016 → 0.011	–	0.000 → 0.000	124 → 32
qwen1.5b-proverqa-b_td_v3-s1	200	0.062 → 0.023	-0.297 → 0.086	0.71 → 0.66	0.58 → 0.56	0.012 → -0.004	-0.008 → -0.013	0.007 → 0.053	0.084 → 0.062	-0.04 → -0.01	0.007 → 0.051	-0.104 → 0.061	0.000 → 0.000	56 → 29
qwen1.5b-proverqa-b_td_v3-s2	200	0.055 → 0.117	-0.286 → -0.307	0.67 → 0.70	0.54 → 0.58	0.010 → 0.098	-0.012 → 0.003	0.008 → 0.042	0.084 → 0.090	-0.08 → -0.18	0.005 → 0.074	-0.087 → 0.077	0.000 → 0.000	56 → 13
qwen1.5b-proverqa-b_td_v3-s3	200	0.060 → 0.140	-0.238 → -0.366	0.65 → 0.73	0.53 → 0.59	0.012 → -0.069	0.008 → -0.231	0.008 → 0.039	0.081 → 0.076	-0.03 → 0.40	0.004 → 0.063	-0.096 → 0.034	0.000 → 0.000	55 → 15
qwen1.5b-proverqa-b_td_v3_nozv-s1	200	0.070 → 0.006	-0.293 → 0.024	0.71 → 0.66	0.58 → 0.55	0.013 → 0.006	-0.008 → 0.014	0.006 → 0.006	0.082 → 0.001	-0.06 → 0.13	0.000 → 0.000	-0.096 → 0.100	0.000 → 0.000	55 → 33
qwen1.5b-proverqa-b_td_v3_nomask-s1	200	0.075 → -0.013	-0.348 → 0.016	0.71 → 0.66	0.57 → 0.56	0.014 → 0.005	-0.011 → 0.030	0.007 → 0.000	0.085 → 0.014	-0.05 → -0.02	0.009 → 0.011	-0.080 → 0.085	0.000 → 0.000	55 → 25
qwen1.5b-proverqa-b_td_v3_zs-s1	200	0.055 → 0.047	-0.298 → -0.011	0.71 → 0.66	0.58 → 0.56	-0.000 → -0.000	0.017 → 0.075	0.006 → 0.008	0.092 → 0.066	0.00 → -0.14	0.000 → -0.000	-0.106 → 0.055	0.000 → 0.000	56 → 32
qwen1.5b-proverqa-b_hc_v3-s1	200	0.058 → 0.164	-0.296 → -0.241	0.71 → 0.66	0.58 → 0.56	0.080 → 0.057	0.067 → -0.041	0.022 → 0.045	0.191 → 0.148	0.28 → -0.19	0.042 → 0.064	-0.111 → 0.035	0.000 → 0.000	55 → 28
qwen1.5b-proverqa-s_hc_v3-s1	200	–	–	–	–	0.005 → 0.087	0.039 → 0.015	0.061 → 0.045	0.321 → 0.417	0.11 → -1.00	-0.011 → 0.004	–	0.000 → 0.000	10 → 5
qwen1.5b-proverqa-s_gaed_v3-s1	200	–	–	–	–	0.533 → 0.551	0.019 → 0.034	0.000 → 0.000	0.435 → 0.106	0.07 → -0.19	0.044 → 0.089	–	0.000 → 0.000	10 → 287
qwen1.5b-proverqa-s_gaed_v3-s2	200	–	–	–	–	0.549 → 0.462	0.017 → 0.038	0.000 → 0.000	0.431 → 0.134	0.06 → -0.45	0.041 → 0.095	–	0.000 → 0.000	10 → 282
qwen1.5b-proverqa-s_gaed_v3-s3	200	–	–	–	–	0.559 → 0.401	0.017 → 0.059	0.000 → 0.000	0.432 → 0.250	0.04 → 0.11	0.047 → 0.113	–	0.000 → 0.000	10 → 116
qwen1.5b-proverqa-j_hc_v3-s1	200	–	–	–	–	0.030 → 0.017	-0.032 → -0.031	0.074 → 0.028	0.331 → 0.258	0.06 → 0.50	0.019 → 0.035	–	0.000 → 0.000	81 → 16

Warnings (automatic)

• v3: |rho mass-length| > 0.5 (a length push) (2 runs): llama3b-math-b_td_v3-s1 (-0.67), qwen1.5b-proverqa-s_hc_v3-s1 (-1.00)
• v3: sign flips > 0.05 (4 runs): llama3b-math-s_hc_v3-s1 (0.068), qwen1.5b-math-s_hc_v3-s1 (0.055), qwen1.5b-math-s_hc_v3-s2 (0.076), qwen1.5b-proverqa-b_td_v3-s1 (0.053)
• v3: reader/states take > 50% of the round (consider the vLLM reader) (2 runs): qwen1.5b-proverqa-s_gaed_v3-s1 (69%), qwen1.5b-proverqa-s_gaed_v3-s2 (69%)
• KL > 0.3 (41 runs): llama3b-math-b_td_v3-s1 (10.46), llama3b-math-grpo-s1 (10.39), llama3b-math-grpo_v2-s1 (10.20), llama3b-math-j_gae-s1 (10.32), llama3b-math-j_gaed-s1 (10.26), llama3b-math-j_hc-s1 (10.42), llama3b-math-prm800k-s1 (11.06), llama3b-math-qwenprm-s1 (10.50), llama3b-math-qwenprm_v2-s1 (10.27), llama3b-math-s_gae-s1 (10.32), llama3b-math-s_gae_v2-s1 (10.27), llama3b-math-s_gaed-s1 (8.46), llama3b-math-s_hc-s1 (10.30), llama3b-math-s_hc_v3-s1 (10.51), llama3b-math-versaprm-s1 (10.61), llama3b-math-vp_gae_v2-s1 (10.27), llama3b-proverqa-grpo-s1 (10.11), llama3b-proverqa-j_gae-s1 (10.42), llama3b-proverqa-j_gaed-s1 (10.25), llama3b-proverqa-j_hc-s1 (10.65), llama3b-proverqa-prm800k-s1 (11.51), llama3b-proverqa-qwenprm-s1 (11.62), llama3b-proverqa-s_gae-s1 (10.49), llama3b-proverqa-s_gaed-s1 (7.41), llama3b-proverqa-s_hc-s1 (10.64), llama3b-proverqa-versaprm-s1 (11.50), qwen1.5b-math-qwenprm-s1 (0.51), qwen1.5b-math-s_hc-s1 (0.37), qwen1.5b-proverqa-b_hc_v3-s1 (0.76), qwen1.5b-proverqa-b_td_v3-s1 (0.47), qwen1.5b-proverqa-b_td_v3-s2 (0.59), qwen1.5b-proverqa-b_td_v3-s3 (0.56), qwen1.5b-proverqa-b_td_v3_nomask-s1 (0.45), qwen1.5b-proverqa-j_gae-s1 (0.40), qwen1.5b-proverqa-j_hc-s1 (0.46), qwen1.5b-proverqa-j_hc_v3-s1 (0.36), qwen1.5b-proverqa-prm800k-s1 (1.64), qwen1.5b-proverqa-qwenprm-s1 (1.27), qwen1.5b-proverqa-s_gae-s1 (0.35), qwen1.5b-proverqa-s_hc-s1 (0.76), qwen1.5b-proverqa-versaprm-s1 (1.29)
• in-loop eval fell ≥ 0.05 below its best (13 runs): llama3b-math-prm800k-s1 (0.494 r25 → 0.248), llama3b-math-qwenprm-s1 (0.514 r25 → 0.444), llama3b-math-versaprm-s1 (0.492 r25 → 0.396), qwen1.5b-math-qwenprm-s1 (0.550 r25 → 0.360), qwen1.5b-math-s_hc-s1 (0.512 r25 → 0.442), qwen1.5b-math-versaprm-s1 (0.516 r25 → 0.428), qwen1.5b-proverqa-b_td_v3_zs-s1 (0.778 r150 → 0.728), qwen1.5b-proverqa-grpo_v2-s1 (0.690 r75 → 0.530), qwen1.5b-proverqa-prm800k_v2-s1 (0.726 r100 → 0.628), qwen1.5b-proverqa-s_gaed_v2-s1 (0.732 r125 → 0.504), qwen1.5b-proverqa-vp_gae_v2-s1 (0.742 r125 → 0.636), qwen1.5b-proverqa-vp_gaed_v2-s1 (0.750 r150 → 0.520), qwen1.5b-proverqa-vp_hc_v2-s1 (0.746 r125 → 0.556)
• step-marker rate < 0.2: the step format is disappearing (38 runs): llama3b-math-prm800k-s1 (0.00), llama3b-math-versaprm-s1 (0.00), llama3b-proverqa-prm800k-s1 (0.00), llama3b-proverqa-qwenprm-s1 (0.00), llama3b-proverqa-versaprm-s1 (0.00), qwen1.5b-math-b_hc_v3-s1 (0.00), qwen1.5b-math-b_td_v3-s1 (0.05), qwen1.5b-math-b_td_v3-s2 (0.18), qwen1.5b-math-b_td_v3-s3 (0.04), qwen1.5b-math-b_td_v3_nozv-s1 (0.04), qwen1.5b-math-grpo-s1 (0.03), qwen1.5b-math-grpo-s2 (0.05), qwen1.5b-math-grpo-s3 (0.03), qwen1.5b-math-grpo_v2-s1 (0.02), qwen1.5b-math-j_gae-s1 (0.07), qwen1.5b-math-j_gae_v2-s1 (0.04), qwen1.5b-math-j_gaed-s1 (0.02), qwen1.5b-math-j_gaed_v2-s1 (0.17), qwen1.5b-math-j_hc-s1 (0.00), qwen1.5b-math-j_hc_v2-s1 (0.02), qwen1.5b-math-prm800k-s1 (0.03), qwen1.5b-math-prm800k_v2-s1 (0.10), qwen1.5b-math-s_gae-s1 (0.06), qwen1.5b-math-s_gae_v2-s1 (0.03), qwen1.5b-math-s_gae_v2_nogate-s1 (0.01), qwen1.5b-math-s_gaed-s1 (0.06), qwen1.5b-math-s_gaed_v2-s1 (0.04), qwen1.5b-math-s_gaed_v3-s1 (0.03), qwen1.5b-math-s_hc_v2-s1 (0.03), qwen1.5b-math-s_hc_v3-s3 (0.17), qwen1.5b-math-versaprm-s1 (0.15), qwen1.5b-math-versaprm_v2-s1 (0.08), qwen1.5b-math-vp_gae_v2-s1 (0.03), qwen1.5b-math-vp_gae_v2_nogate-s1 (0.03), qwen1.5b-math-vp_gaed_v2-s1 (0.05), qwen1.5b-math-vp_hc_v2-s1 (0.04), qwen1.5b-proverqa-qwenprm-s1 (0.00), qwen1.5b-proverqa-versaprm-s1 (0.00)
• < 1 step/rollout: process credit has almost nothing to act on (7 runs): llama3b-math-prm800k-s1 (0.00), llama3b-proverqa-prm800k-s1 (0.00), llama3b-proverqa-qwenprm-s1 (0.00), llama3b-proverqa-versaprm-s1 (0.00), qwen1.5b-math-prm800k-s1 (0.19), qwen1.5b-proverqa-qwenprm-s1 (0.00), qwen1.5b-proverqa-versaprm-s1 (0.00)
• > 70% of groups give no GRPO signal (all right or all wrong) (38 runs): llama3b-math-prm800k-s1 (77%), llama3b-proverqa-grpo-s1 (83%), llama3b-proverqa-j_gae-s1 (83%), llama3b-proverqa-j_gaed-s1 (81%), llama3b-proverqa-j_hc-s1 (89%), llama3b-proverqa-prm800k-s1 (99%), llama3b-proverqa-qwenprm-s1 (99%), llama3b-proverqa-s_gae-s1 (76%), llama3b-proverqa-s_gaed-s1 (77%), llama3b-proverqa-s_hc-s1 (93%), llama3b-proverqa-versaprm-s1 (98%), qwen1.5b-math-qwenprm-s1 (72%), qwen1.5b-proverqa-b_hc_v3-s1 (97%), qwen1.5b-proverqa-b_td_v3-s1 (95%), qwen1.5b-proverqa-b_td_v3-s2 (92%), qwen1.5b-proverqa-b_td_v3-s3 (96%), qwen1.5b-proverqa-b_td_v3_nomask-s1 (99%), qwen1.5b-proverqa-b_td_v3_nozv-s1 (97%), qwen1.5b-proverqa-b_td_v3_zs-s1 (95%), qwen1.5b-proverqa-grpo-s1 (98%), qwen1.5b-proverqa-grpo-s2 (75%), qwen1.5b-proverqa-grpo-s3 (88%), qwen1.5b-proverqa-j_gae-s1 (98%), qwen1.5b-proverqa-j_hc-s1 (86%), qwen1.5b-proverqa-j_hc_v3-s1 (81%), qwen1.5b-proverqa-prm800k-s1 (98%), qwen1.5b-proverqa-qwenprm-s1 (98%), qwen1.5b-proverqa-s_gae-s1 (95%), qwen1.5b-proverqa-s_gaed-s1 (71%), qwen1.5b-proverqa-s_gaed_v2-s1 (91%), qwen1.5b-proverqa-s_gaed_v3-s1 (96%), qwen1.5b-proverqa-s_gaed_v3-s2 (90%), qwen1.5b-proverqa-s_hc-s1 (99%), qwen1.5b-proverqa-s_hc_v3-s1 (98%), qwen1.5b-proverqa-versaprm-s1 (97%), qwen1.5b-proverqa-vp_gae_v2-s1 (92%), qwen1.5b-proverqa-vp_gaed_v2-s1 (91%), qwen1.5b-proverqa-vp_hc_v2-s1 (92%)
• entropy < 0.05: policy nearly deterministic (12 runs): llama3b-proverqa-prm800k-s1 (0.024), llama3b-proverqa-qwenprm-s1 (0.001), qwen1.5b-math-qwenprm-s1 (0.040), qwen1.5b-math-s_hc-s1 (0.046), qwen1.5b-math-s_hc_v3-s1 (0.045), qwen1.5b-math-s_hc_v3-s2 (0.047), qwen1.5b-proverqa-prm800k-s1 (0.001), qwen1.5b-proverqa-qwenprm-s1 (0.005), qwen1.5b-proverqa-s_gaed_v3-s1 (0.004), qwen1.5b-proverqa-s_gaed_v3-s2 (0.017), qwen1.5b-proverqa-s_hc_v3-s1 (0.019), qwen1.5b-proverqa-versaprm-s1 (0.004)

Training dynamics (first 3 rounds → last 3 rounds)

run	round	reward	zero-std groups	length	truncated	entropy	KL	grad norm
llama3b-math-b_td_v3-s1	200	0.577 → 0.841	0.15 → 0.55	452 → 287	0.064 → 0.007	1.009 → 0.090	9.923 → 10.459	0.48 → 1.03
llama3b-math-grpo-s1	200	0.589 → 0.841	0.18 → 0.58	449 → 444	0.069 → 0.010	1.013 → 0.138	10.194 → 10.388	0.44 → 0.54
llama3b-math-grpo_v2-s1	200	0.538 → 0.650	0.20 → 0.24	473 → 472	0.084 → 0.014	1.047 → 0.308	9.801 → 10.196	0.44 → 0.55
llama3b-math-j_gae-s1	200	0.587 → 0.846	0.16 → 0.52	445 → 435	0.062 → 0.006	0.977 → 0.165	9.925 → 10.318	0.73 → 0.79
llama3b-math-j_gaed-s1	200	0.585 → 0.859	0.15 → 0.55	453 → 412	0.074 → 0.008	0.981 → 0.163	9.896 → 10.256	0.71 → 0.84
llama3b-math-j_hc-s1	200	0.590 → 0.810	0.17 → 0.47	446 → 259	0.062 → 0.000	0.943 → 0.166	9.982 → 10.419	0.53 → 0.87
llama3b-math-prm800k-s1	200	0.585 → 0.496	0.15 → 0.77	447 → 18	0.065 → 0.000	1.013 → 0.094	9.753 → 11.062	0.62 → 4.04
llama3b-math-qwenprm-s1	200	0.568 → 0.704	0.14 → 0.64	454 → 189	0.066 → 0.001	1.014 → 0.054	9.940 → 10.504	0.63 → 1.29
llama3b-math-qwenprm_v2-s1	200	0.551 → 0.639	0.23 → 0.33	471 → 516	0.081 → 0.021	1.075 → 0.229	10.208 → 10.272	0.42 → 0.43
llama3b-math-s_gae-s1	200	0.569 → 0.846	0.15 → 0.52	449 → 409	0.068 → 0.008	1.010 → 0.161	9.976 → 10.322	0.71 → 0.98
llama3b-math-s_gae_v2-s1	200	0.542 → 0.627	0.23 → 0.26	476 → 531	0.087 → 0.033	1.150 → 0.238	10.150 → 10.269	0.44 → 0.49
llama3b-math-s_gaed-s1	200	0.597 → 0.826	0.16 → 0.49	443 → 1018	0.053 → 0.984	0.912 → 1.033	9.967 → 8.459	0.78 → 0.56
llama3b-math-s_hc-s1	200	0.599 → 0.811	0.17 → 0.49	447 → 286	0.063 → 0.001	0.931 → 0.127	10.018 → 10.301	0.70 → 0.94
llama3b-math-s_hc_v3-s1	200	0.577 → 0.785	0.15 → 0.50	447 → 216	0.055 → 0.000	0.936 → 0.097	9.971 → 10.512	0.51 → 1.06
llama3b-math-versaprm-s1	200	0.574 → 0.616	0.14 → 0.68	452 → 93	0.071 → 0.000	0.988 → 0.079	9.901 → 10.608	0.64 → 1.97
llama3b-math-vp_gae_v2-s1	200	0.538 → 0.645	0.18 → 0.31	470 → 488	0.074 → 0.021	0.986 → 0.225	10.403 → 10.265	0.45 → 0.52
llama3b-proverqa-grpo-s1	200	0.493 → 0.912	0.06 → 0.83	360 → 204	0.151 → 0.000	4.027 → 0.559	7.367 → 10.109	0.68 → 0.62
llama3b-proverqa-j_gae-s1	200	0.492 → 0.921	0.07 → 0.83	368 → 268	0.160 → 0.000	4.167 → 0.142	7.237 → 10.415	1.01 → 1.37
llama3b-proverqa-j_gaed-s1	200	0.492 → 0.904	0.06 → 0.81	364 → 696	0.156 → 0.005	4.101 → 0.252	7.361 → 10.252	1.06 → 0.79
llama3b-proverqa-j_hc-s1	200	0.482 → 0.900	0.08 → 0.89	368 → 69	0.164 → 0.000	4.155 → 0.069	7.282 → 10.647	0.78 → 1.62
llama3b-proverqa-prm800k-s1	200	0.486 → 0.837	0.09 → 0.99	362 → 8	0.155 → 0.000	4.017 → 0.024	7.462 → 11.508	0.80 → 1.95
llama3b-proverqa-qwenprm-s1	200	0.492 → 0.831	0.06 → 0.99	365 → 6	0.152 → 0.000	4.046 → 0.001	7.384 → 11.622	0.88 → 0.49
llama3b-proverqa-s_gae-s1	200	0.484 → 0.891	0.08 → 0.76	380 → 242	0.173 → 0.003	4.337 → 0.186	7.132 → 10.487	0.99 → 1.76
llama3b-proverqa-s_gaed-s1	200	0.476 → 0.863	0.06 → 0.77	372 → 970	0.163 → 0.861	4.161 → 2.315	7.379 → 7.415	1.10 → 0.56
llama3b-proverqa-s_hc-s1	200	0.488 → 0.882	0.06 → 0.93	359 → 45	0.150 → 0.000	4.019 → 0.113	7.405 → 10.635	1.13 → 1.04
llama3b-proverqa-versaprm-s1	200	0.484 → 0.790	0.06 → 0.98	360 → 8	0.145 → 0.000	3.900 → 0.075	7.476 → 11.503	0.88 → 4.66
qwen1.5b-math-b_hc_v3-s1	200	0.434 → 0.797	0.15 → 0.56	384 → 370	0.014 → 0.019	0.431 → 0.078	0.000 → 0.120	0.42 → 0.65
qwen1.5b-math-b_td_v3-s1	200	0.432 → 0.799	0.18 → 0.62	377 → 405	0.014 → 0.016	0.420 → 0.069	0.000 → 0.060	0.40 → 0.41
qwen1.5b-math-b_td_v3-s2	200	0.420 → 0.844	0.18 → 0.61	349 → 407	0.014 → 0.023	0.439 → 0.065	0.000 → 0.063	0.42 → 0.39
qwen1.5b-math-b_td_v3-s3	200	0.464 → 0.814	0.18 → 0.58	372 → 402	0.016 → 0.017	0.441 → 0.074	0.000 → 0.058	0.43 → 0.41
qwen1.5b-math-b_td_v3_nozv-s1	200	0.439 → 0.815	0.14 → 0.58	379 → 408	0.012 → 0.010	0.423 → 0.088	0.000 → 0.062	0.41 → 0.35
qwen1.5b-math-grpo-s1	200	0.437 → 0.794	0.16 → 0.60	380 → 410	0.014 → 0.010	0.434 → 0.083	0.000 → 0.050	0.38 → 0.28
qwen1.5b-math-grpo-s2	200	0.417 → 0.846	0.19 → 0.57	351 → 416	0.014 → 0.020	0.444 → 0.083	0.000 → 0.048	0.41 → 0.30
qwen1.5b-math-grpo-s3	200	0.461 → 0.819	0.17 → 0.59	369 → 395	0.017 → 0.012	0.443 → 0.072	0.000 → 0.056	0.41 → 0.32
qwen1.5b-math-grpo_v2-s1	200	0.380 → 0.508	0.22 → 0.27	382 → 558	0.015 → 0.059	0.481 → 0.126	0.000 → 0.053	0.37 → 0.40
qwen1.5b-math-j_gae-s1	200	0.441 → 0.816	0.15 → 0.58	383 → 393	0.011 → 0.008	0.431 → 0.068	0.000 → 0.058	0.60 → 0.58
qwen1.5b-math-j_gae_v2-s1	200	0.380 → 0.479	0.22 → 0.29	375 → 570	0.014 → 0.063	0.467 → 0.106	0.000 → 0.048	0.37 → 0.32
qwen1.5b-math-j_gaed-s1	200	0.439 → 0.811	0.18 → 0.60	387 → 416	0.015 → 0.018	0.424 → 0.058	0.000 → 0.063	0.63 → 0.72
qwen1.5b-math-j_gaed_v2-s1	200	0.386 → 0.560	0.20 → 0.26	376 → 511	0.011 → 0.029	0.470 → 0.094	0.000 → 0.051	0.38 → 0.44
qwen1.5b-math-j_hc-s1	200	0.436 → 0.796	0.16 → 0.57	379 → 364	0.014 → 0.018	0.423 → 0.058	0.000 → 0.071	0.48 → 0.55
qwen1.5b-math-j_hc_v2-s1	200	0.387 → 0.523	0.22 → 0.31	382 → 571	0.014 → 0.061	0.481 → 0.100	0.000 → 0.050	0.37 → 0.31
qwen1.5b-math-j_hc_v3-s1	200	0.424 → 0.796	0.20 → 0.54	381 → 341	0.014 → 0.018	0.420 → 0.065	0.000 → 0.142	0.43 → 0.43
qwen1.5b-math-prm800k-s1	200	0.440 → 0.768	0.16 → 0.49	384 → 488	0.016 → 0.023	0.420 → 0.055	0.000 → 0.125	0.53 → 0.35
qwen1.5b-math-prm800k_v2-s1	200	0.380 → 0.523	0.17 → 0.30	379 → 556	0.016 → 0.063	0.476 → 0.100	0.000 → 0.050	0.38 → 0.37
qwen1.5b-math-qwenprm-s1	200	0.439 → 0.585	0.16 → 0.72	385 → 187	0.015 → 0.000	0.426 → 0.040	0.000 → 0.513	0.55 → 0.89
qwen1.5b-math-qwenprm_v2-s1	200	0.386 → 0.576	0.20 → 0.30	381 → 594	0.018 → 0.112	0.466 → 0.073	0.000 → 0.060	0.38 → 0.37
qwen1.5b-math-s_gae-s1	200	0.438 → 0.823	0.14 → 0.63	381 → 403	0.016 → 0.008	0.424 → 0.081	0.000 → 0.071	0.59 → 0.58
qwen1.5b-math-s_gae_v2-s1	200	0.377 → 0.550	0.19 → 0.34	380 → 531	0.015 → 0.037	0.464 → 0.114	0.000 → 0.041	0.38 → 0.35
qwen1.5b-math-s_gae_v2_nogate-s1	200	0.380 → 0.513	0.20 → 0.26	383 → 476	0.020 → 0.036	0.474 → 0.076	0.000 → 0.089	0.43 → 0.45
qwen1.5b-math-s_gaed-s1	200	0.432 → 0.805	0.18 → 0.58	381 → 408	0.017 → 0.018	0.428 → 0.076	0.000 → 0.068	0.64 → 0.59
qwen1.5b-math-s_gaed_v2-s1	200	0.385 → 0.521	0.19 → 0.31	379 → 531	0.018 → 0.021	0.472 → 0.115	0.000 → 0.046	0.38 → 0.32
qwen1.5b-math-s_gaed_v3-s1	200	0.444 → 0.803	0.16 → 0.59	383 → 401	0.017 → 0.010	0.415 → 0.077	0.000 → 0.065	0.59 → 0.45
qwen1.5b-math-s_hc-s1	200	0.451 → 0.641	0.15 → 0.44	377 → 369	0.011 → 0.016	0.421 → 0.046	0.000 → 0.368	0.55 → 0.69
qwen1.5b-math-s_hc_v2-s1	200	0.393 → 0.521	0.20 → 0.24	379 → 539	0.016 → 0.036	0.469 → 0.123	0.000 → 0.065	0.41 → 0.33
qwen1.5b-math-s_hc_v3-s1	200	0.439 → 0.786	0.14 → 0.58	382 → 354	0.014 → 0.003	0.422 → 0.045	0.000 → 0.124	0.46 → 0.49
qwen1.5b-math-s_hc_v3-s2	200	0.426 → 0.820	0.19 → 0.56	348 → 355	0.008 → 0.007	0.440 → 0.047	0.000 → 0.089	0.50 → 0.49
qwen1.5b-math-s_hc_v3-s3	200	0.460 → 0.751	0.16 → 0.51	369 → 378	0.018 → 0.006	0.433 → 0.060	0.000 → 0.114	0.48 → 0.42
qwen1.5b-math-versaprm-s1	200	0.436 → 0.670	0.18 → 0.51	383 → 238	0.016 → 0.000	0.419 → 0.053	0.000 → 0.129	0.58 → 0.80
qwen1.5b-math-versaprm_v2-s1	200	0.375 → 0.550	0.22 → 0.27	378 → 530	0.020 → 0.037	0.471 → 0.105	0.000 → 0.049	0.37 → 0.39
qwen1.5b-math-vp_gae_v2-s1	200	0.388 → 0.550	0.18 → 0.27	382 → 544	0.014 → 0.050	0.473 → 0.092	0.000 → 0.052	0.40 → 0.32
qwen1.5b-math-vp_gae_v2_nogate-s1	200	0.368 → 0.504	0.21 → 0.29	378 → 487	0.016 → 0.035	0.482 → 0.081	0.000 → 0.056	0.42 → 0.39
qwen1.5b-math-vp_gaed_v2-s1	200	0.384 → 0.523	0.22 → 0.24	382 → 529	0.016 → 0.046	0.477 → 0.100	0.000 → 0.050	0.36 → 0.34
qwen1.5b-math-vp_hc_v2-s1	200	0.377 → 0.557	0.24 → 0.27	383 → 518	0.020 → 0.048	0.473 → 0.106	0.000 → 0.045	0.37 → 0.37
qwen1.5b-proverqa-b_hc_v3-s1	200	0.465 → 0.755	0.06 → 0.97	202 → 65	0.000 → 0.000	1.103 → 0.133	0.000 → 0.763	0.90 → 0.82
qwen1.5b-proverqa-b_td_v3-s1	200	0.477 → 0.753	0.06 → 0.95	206 → 67	0.001 → 0.000	1.102 → 0.238	0.000 → 0.469	0.92 → 0.62
qwen1.5b-proverqa-b_td_v3-s2	200	0.444 → 0.824	0.10 → 0.92	205 → 40	0.001 → 0.000	1.095 → 0.107	0.000 → 0.591	0.87 → 1.26
qwen1.5b-proverqa-b_td_v3-s3	200	0.447 → 0.796	0.08 → 0.96	203 → 49	0.000 → 0.000	1.107 → 0.074	0.000 → 0.565	0.91 → 1.34
qwen1.5b-proverqa-b_td_v3_nomask-s1	200	0.480 → 0.732	0.07 → 0.99	205 → 46	0.000 → 0.000	1.104 → 0.308	0.000 → 0.445	0.92 → 0.96
qwen1.5b-proverqa-b_td_v3_nozv-s1	200	0.496 → 0.757	0.05 → 0.97	205 → 68	0.000 → 0.000	1.101 → 0.440	0.000 → 0.204	0.90 → 0.49
qwen1.5b-proverqa-b_td_v3_zs-s1	200	0.466 → 0.740	0.08 → 0.95	206 → 183	0.000 → 0.000	1.115 → 0.572	0.000 → 0.170	0.85 → 0.24
qwen1.5b-proverqa-grpo-s1	200	0.486 → 0.731	0.06 → 0.98	205 → 63	0.000 → 0.000	1.108 → 0.634	0.000 → 0.223	0.86 → 0.19
qwen1.5b-proverqa-grpo-s2	200	0.452 → 0.805	0.09 → 0.75	205 → 251	0.000 → 0.000	1.093 → 0.315	0.000 → 0.104	0.84 → 0.33
qwen1.5b-proverqa-grpo-s3	200	0.443 → 0.820	0.07 → 0.88	205 → 83	0.000 → 0.000	1.110 → 0.293	0.000 → 0.224	0.84 → 1.29
qwen1.5b-proverqa-grpo_v2-s1	200	0.434 → 0.323	0.07 → 0.63	207 → 285	0.001 → 0.000	1.090 → 0.264	0.000 → 0.177	0.83 → 0.36
qwen1.5b-proverqa-j_gae-s1	200	0.482 → 0.760	0.07 → 0.98	205 → 51	0.000 → 0.000	1.107 → 0.191	0.000 → 0.399	1.29 → 0.62
qwen1.5b-proverqa-j_gaed-s1	200	0.480 → 0.753	0.07 → 0.66	205 → 341	0.000 → 0.000	1.100 → 0.591	0.000 → 0.168	1.34 → 0.81
qwen1.5b-proverqa-j_hc-s1	200	0.479 → 0.767	0.09 → 0.86	205 → 52	0.000 → 0.000	1.085 → 0.085	0.000 → 0.465	1.01 → 1.50
qwen1.5b-proverqa-j_hc_v3-s1	200	0.482 → 0.766	0.07 → 0.81	203 → 124	0.000 → 0.001	1.093 → 0.234	0.000 → 0.359	1.07 → 0.64
qwen1.5b-proverqa-prm800k-s1	200	0.473 → 0.721	0.09 → 0.98	203 → 15	0.000 → 0.000	1.095 → 0.001	0.001 → 1.639	1.06 → 0.54
qwen1.5b-proverqa-prm800k_v2-s1	200	0.463 → 0.429	0.06 → 0.28	203 → 240	0.001 → 0.000	1.079 → 0.641	0.000 → 0.130	0.84 → 0.77
qwen1.5b-proverqa-qwenprm-s1	200	0.487 → 0.735	0.07 → 0.98	202 → 7	0.000 → 0.000	1.081 → 0.005	0.001 → 1.267	1.04 → 2.26
qwen1.5b-proverqa-qwenprm_v2-s1	200	0.450 → 0.484	0.05 → 0.26	202 → 310	0.001 → 0.000	1.088 → 0.169	0.000 → 0.095	0.89 → 0.46
qwen1.5b-proverqa-s_gae-s1	200	0.463 → 0.734	0.07 → 0.95	202 → 85	0.000 → 0.000	1.087 → 0.309	0.000 → 0.345	1.32 → 0.89
qwen1.5b-proverqa-s_gae_v2-s1	200	0.441 → 0.549	0.08 → 0.34	205 → 214	0.001 → 0.001	1.093 → 0.251	0.000 → 0.176	0.83 → 0.70
qwen1.5b-proverqa-s_gaed-s1	200	0.486 → 0.754	0.10 → 0.71	204 → 406	0.000 → 0.003	1.113 → 0.246	0.000 → 0.213	1.36 → 0.69
qwen1.5b-proverqa-s_gaed_v2-s1	200	0.435 → 0.408	0.07 → 0.91	203 → 192	0.001 → 0.006	1.093 → 0.347	0.000 → 0.194	0.89 → 0.43
qwen1.5b-proverqa-s_gaed_v3-s1	200	0.477 → 0.735	0.07 → 0.96	204 → 1015	0.000 → 0.987	1.095 → 0.004	0.000 → 0.075	1.29 → 0.03
qwen1.5b-proverqa-s_gaed_v3-s2	200	0.454 → 0.752	0.08 → 0.90	206 → 994	0.000 → 0.568	1.100 → 0.017	0.000 → 0.103	1.24 → 0.17
qwen1.5b-proverqa-s_gaed_v3-s3	200	0.451 → 0.771	0.09 → 0.57	205 → 910	0.001 → 0.816	1.116 → 0.056	0.000 → 0.171	1.23 → 0.24
qwen1.5b-proverqa-s_hc-s1	200	0.477 → 0.740	0.07 → 0.99	202 → 26	0.001 → 0.000	1.083 → 0.148	0.001 → 0.757	1.10 → 1.88
qwen1.5b-proverqa-s_hc_v2-s1	200	0.447 → 0.665	0.06 → 0.33	206 → 143	0.001 → 0.005	1.101 → 0.160	0.000 → 0.207	0.90 → 0.92
qwen1.5b-proverqa-s_hc_v3-s1	200	0.486 → 0.751	0.06 → 0.98	206 → 185	0.001 → 0.000	1.103 → 0.019	0.001 → 0.104	0.96 → 0.17
qwen1.5b-proverqa-versaprm-s1	200	0.471 → 0.730	0.08 → 0.97	203 → 7	0.000 → 0.000	1.095 → 0.004	0.001 → 1.285	1.01 → 1.85
qwen1.5b-proverqa-versaprm_v2-s1	200	0.452 → 0.492	0.08 → 0.50	207 → 135	0.001 → 0.000	1.104 → 0.391	0.000 → 0.147	0.85 → 0.76
qwen1.5b-proverqa-vp_gae_v2-s1	200	0.430 → 0.559	0.07 → 0.92	203 → 71	0.001 → 0.000	1.087 → 0.716	0.000 → 0.168	0.86 → 0.88
qwen1.5b-proverqa-vp_gaed_v2-s1	200	0.444 → 0.215	0.08 → 0.91	207 → 112	0.001 → 0.001	1.094 → 0.350	0.000 → 0.191	0.84 → 0.67
qwen1.5b-proverqa-vp_hc_v2-s1	200	0.423 → 0.525	0.08 → 0.92	203 → 207	0.001 → 0.000	1.102 → 0.410	0.000 → 0.178	0.88 → 0.26

Process credit / PRM (first 3 rounds → last 3 rounds)

run	round	steps/rollout	step markers	abs process adv	PRM score	PRM min-step AUC	PRM last-step AUC	credit AUC	state self-loop	credit s/round
llama3b-math-b_td_v3-s1	200	7.17 → 1.99	0.99 → 1.00	–	–	–	–	–	–	81 → 27
llama3b-math-grpo-s1	200	7.13 → 6.43	0.99 → 1.00	–	–	–	–	–	–	–
llama3b-math-grpo_v2-s1	200	7.26 → 6.52	1.00 → 1.00	–	–	–	–	–	–	–
llama3b-math-j_gae-s1	200	7.11 → 8.07	0.99 → 1.00	0.414 → 0.294	–	–	–	0.989 → 0.984	0.42 → 0.43	127 → 144
llama3b-math-j_gaed-s1	200	7.20 → 7.81	0.99 → 1.00	0.473 → 0.401	–	–	–	0.987 → 0.988	0.42 → 0.44	128 → 138
llama3b-math-j_hc-s1	200	7.10 → 1.97	0.99 → 1.00	0.368 → 0.240	–	–	–	0.634 → 0.497	0.41 → 0.77	126 → 37
llama3b-math-prm800k-s1	200	7.12 → 0.00	0.99 → 0.00	0.415 → 0.167	0.784 → 0.164	0.840 → 1.000	0.807 → 1.000	0.830 → 1.000	–	24 → 0
llama3b-math-qwenprm-s1	200	7.25 → 4.64	0.99 → 0.57	0.486 → 0.271	0.765 → 0.938	0.896 → 0.964	0.904 → 0.942	0.893 → 0.921	–	24 → 10
llama3b-math-qwenprm_v2-s1	200	7.24 → 7.50	1.00 → 1.00	–	0.756 → 0.807	–	–	–	–	25 → 26
llama3b-math-s_gae-s1	200	7.11 → 8.07	0.99 → 1.00	0.407 → 0.304	–	–	–	0.982 → 0.987	0.52 → 0.63	29 → 26
llama3b-math-s_gae_v2-s1	200	7.25 → 9.31	1.00 → 1.00	–	–	–	–	–	0.51 → 0.62	30 → 35
llama3b-math-s_gaed-s1	200	7.13 → 4.05	0.99 → 1.00	0.493 → 0.453	–	–	–	0.984 → 0.990	0.52 → 0.22	28 → 42
llama3b-math-s_hc-s1	200	7.17 → 9.75	0.99 → 1.00	0.542 → 0.343	–	–	–	0.705 → 0.595	0.52 → 0.80	30 → 21
llama3b-math-s_hc_v3-s1	200	7.13 → 1.11	0.99 → 0.98	–	–	–	–	–	–	30 → 9
llama3b-math-versaprm-s1	200	7.25 → 4.26	0.99 → 0.00	0.459 → 0.282	0.849 → 0.958	0.838 → 0.878	0.828 → 0.858	0.843 → 0.844	–	27 → 7
llama3b-math-vp_gae_v2-s1	200	7.27 → 9.90	1.00 → 1.00	–	–	–	–	–	0.24 → 0.24	19 → 18
llama3b-proverqa-grpo-s1	200	5.39 → 5.81	1.00 → 1.00	–	–	–	–	–	–	–
llama3b-proverqa-j_gae-s1	200	5.34 → 12.15	1.00 → 1.00	0.461 → 0.173	–	–	–	0.997 → 0.894	0.35 → 0.56	91 → 199
llama3b-proverqa-j_gaed-s1	200	5.36 → 29.11	1.00 → 1.00	0.494 → 0.304	–	–	–	0.998 → 0.802	0.35 → 0.73	92 → 520
llama3b-proverqa-j_hc-s1	200	5.37 → 1.97	1.00 → 1.00	0.367 → 0.133	–	–	–	0.622 → 0.542	0.35 → 0.14	93 → 20
llama3b-proverqa-prm800k-s1	200	5.41 → 0.00	1.00 → 0.00	0.333 → 0.000	0.736 → 0.000	0.540 → 0.796	0.557 → 0.837	0.558 → 0.378	–	25 → 0
llama3b-proverqa-qwenprm-s1	200	5.44 → 0.00	1.00 → 0.00	0.575 → 0.000	0.606 → 0.000	0.689 → 0.600	0.685 → 1.000	0.648 → 1.000	–	25 → 0
llama3b-proverqa-s_gae-s1	200	5.33 → 11.82	1.00 → 0.98	0.435 → 0.270	–	–	–	0.997 → 0.879	0.50 → 0.80	27 → 27
llama3b-proverqa-s_gaed-s1	200	5.39 → 2.79	1.00 → 1.00	0.489 → 0.388	–	–	–	0.998 → 0.792	0.51 → 0.73	26 → 82
llama3b-proverqa-s_hc-s1	200	5.38 → 1.00	1.00 → 1.00	0.984 → 0.000	–	–	–	0.693 → 0.500	0.52 → 0.00	27 → 5
llama3b-proverqa-versaprm-s1	200	5.47 → 0.00	1.00 → 0.00	0.462 → 0.000	0.703 → 0.000	0.621 → 0.000	0.604 → 0.000	0.646 → 1.000	–	30 → 0
qwen1.5b-math-b_hc_v3-s1	200	7.34 → 1.02	0.97 → 0.00	–	–	–	–	–	–	67 → 18
qwen1.5b-math-b_td_v3-s1	200	7.29 → 5.46	0.97 → 0.05	–	–	–	–	–	–	67 → 52
qwen1.5b-math-b_td_v3-s2	200	7.16 → 6.83	0.98 → 0.18	–	–	–	–	–	–	62 → 62
qwen1.5b-math-b_td_v3-s3	200	7.14 → 7.42	0.98 → 0.04	–	–	–	–	–	–	67 → 63
qwen1.5b-math-b_td_v3_nozv-s1	200	7.25 → 8.66	0.97 → 0.04	–	–	–	–	–	–	67 → 71
qwen1.5b-math-grpo-s1	200	7.27 → 8.08	0.97 → 0.03	–	–	–	–	–	–	–
qwen1.5b-math-grpo-s2	200	7.24 → 7.88	0.98 → 0.05	–	–	–	–	–	–	–
qwen1.5b-math-grpo-s3	200	7.13 → 7.60	0.98 → 0.03	–	–	–	–	–	–	–
qwen1.5b-math-grpo_v2-s1	200	7.41 → 8.88	0.98 → 0.02	–	–	–	–	–	–	–
qwen1.5b-math-j_gae-s1	200	7.33 → 8.28	0.97 → 0.07	0.400 → 0.329	–	–	–	0.990 → 0.943	0.41 → 0.49	124 → 137
qwen1.5b-math-j_gae_v2-s1	200	7.33 → 8.46	0.98 → 0.04	–	–	–	–	–	0.43 → 0.49	124 → 154
qwen1.5b-math-j_gaed-s1	200	7.40 → 7.26	0.97 → 0.02	0.447 → 0.446	–	–	–	0.986 → 0.920	0.40 → 0.46	126 → 122
qwen1.5b-math-j_gaed_v2-s1	200	7.27 → 7.65	0.98 → 0.17	–	–	–	–	–	0.43 → 0.47	123 → 137
qwen1.5b-math-j_hc-s1	200	7.27 → 1.00	0.98 → 0.00	0.362 → 0.223	–	–	–	0.583 → 0.531	0.40 → 0.33	124 → 22
qwen1.5b-math-j_hc_v2-s1	200	7.40 → 8.72	0.98 → 0.02	–	–	–	–	–	0.42 → 0.48	126 → 161
qwen1.5b-math-j_hc_v3-s1	200	7.39 → 1.60	0.98 → 1.00	–	–	–	–	–	–	125 → 32
qwen1.5b-math-prm800k-s1	200	7.37 → 0.19	0.97 → 0.03	0.432 → 0.364	0.793 → 0.940	0.847 → 0.810	0.828 → 0.802	0.830 → 0.841	–	20 → 2
qwen1.5b-math-prm800k_v2-s1	200	7.32 → 9.98	0.98 → 0.10	–	0.768 → 0.844	–	–	–	–	20 → 27
qwen1.5b-math-qwenprm-s1	200	7.32 → 6.02	0.98 → 0.67	0.469 → 0.232	0.761 → 0.955	0.910 → 0.982	0.931 → 0.974	0.901 → 0.961	–	20 → 10
qwen1.5b-math-qwenprm_v2-s1	200	7.40 → 11.56	0.98 → 0.28	–	0.735 → 0.892	–	–	–	–	20 → 28
qwen1.5b-math-s_gae-s1	200	7.26 → 7.92	0.98 → 0.06	0.406 → 0.335	–	–	–	0.989 → 0.849	0.57 → 0.62	19 → 20
qwen1.5b-math-s_gae_v2-s1	200	7.36 → 9.19	0.98 → 0.03	–	–	–	–	–	0.57 → 0.60	19 → 28
qwen1.5b-math-s_gae_v2_nogate-s1	200	7.41 → 9.42	0.98 → 0.01	–	–	–	–	–	0.57 → 0.65	20 → 25
qwen1.5b-math-s_gaed-s1	200	7.28 → 2.37	0.97 → 0.06	0.476 → 0.457	–	–	–	0.983 → 0.952	0.57 → 0.32	20 → 12
qwen1.5b-math-s_gaed_v2-s1	200	7.40 → 8.54	0.98 → 0.04	–	–	–	–	–	0.57 → 0.55	19 → 26
qwen1.5b-math-s_gaed_v3-s1	200	7.26 → 7.32	0.97 → 0.03	–	–	–	–	–	–	19 → 19
qwen1.5b-math-s_hc-s1	200	7.30 → 20.97	0.97 → 1.00	0.505 → 0.295	–	–	–	0.646 → 0.669	0.57 → 0.88	20 → 37
qwen1.5b-math-s_hc_v2-s1	200	7.35 → 8.86	0.98 → 0.03	–	–	–	–	–	0.57 → 0.57	19 → 28
qwen1.5b-math-s_hc_v3-s1	200	7.32 → 11.53	0.97 → 0.96	–	–	–	–	–	–	21 → 19
qwen1.5b-math-s_hc_v3-s2	200	7.10 → 5.92	0.98 → 0.99	–	–	–	–	–	–	19 → 13
qwen1.5b-math-s_hc_v3-s3	200	7.08 → 12.39	0.98 → 0.17	–	–	–	–	–	–	20 → 24
qwen1.5b-math-versaprm-s1	200	7.38 → 6.69	0.97 → 0.15	0.423 → 0.269	0.853 → 0.961	0.839 → 0.899	0.831 → 0.887	0.841 → 0.893	–	23 → 15
qwen1.5b-math-versaprm_v2-s1	200	7.37 → 8.93	0.98 → 0.08	–	0.830 → 0.881	–	–	–	–	23 → 30
qwen1.5b-math-vp_gae_v2-s1	200	7.41 → 9.99	0.98 → 0.03	–	–	–	–	–	0.26 → 0.24	10 → 13
qwen1.5b-math-vp_gae_v2_nogate-s1	200	7.31 → 8.54	0.98 → 0.03	–	–	–	–	–	0.25 → 0.23	10 → 11
qwen1.5b-math-vp_gaed_v2-s1	200	7.38 → 8.45	0.97 → 0.05	–	–	–	–	–	0.25 → 0.22	10 → 12
qwen1.5b-math-vp_hc_v2-s1	200	7.49 → 9.30	0.98 → 0.04	–	–	–	–	–	0.25 → 0.22	10 → 12
qwen1.5b-proverqa-b_hc_v3-s1	200	5.18 → 5.03	1.00 → 1.00	–	–	–	–	–	–	56 → 29
qwen1.5b-proverqa-b_td_v3-s1	200	5.26 → 3.84	1.00 → 1.00	–	–	–	–	–	–	57 → 29
qwen1.5b-proverqa-b_td_v3-s2	200	5.26 → 2.23	1.00 → 1.00	–	–	–	–	–	–	57 → 13
qwen1.5b-proverqa-b_td_v3-s3	200	5.16 → 3.00	1.00 → 1.00	–	–	–	–	–	–	56 → 15
qwen1.5b-proverqa-b_td_v3_nomask-s1	200	5.25 → 3.03	1.00 → 1.00	–	–	–	–	–	–	56 → 25
qwen1.5b-proverqa-b_td_v3_nozv-s1	200	5.24 → 3.78	1.00 → 1.00	–	–	–	–	–	–	56 → 33
qwen1.5b-proverqa-b_td_v3_zs-s1	200	5.25 → 2.70	1.00 → 1.00	–	–	–	–	–	–	56 → 32
qwen1.5b-proverqa-grpo-s1	200	5.23 → 3.58	1.00 → 1.00	–	–	–	–	–	–	–
qwen1.5b-proverqa-grpo-s2	200	5.25 → 3.44	1.00 → 1.00	–	–	–	–	–	–	–
qwen1.5b-proverqa-grpo-s3	200	5.20 → 3.56	1.00 → 1.00	–	–	–	–	–	–	–
qwen1.5b-proverqa-grpo_v2-s1	200	5.29 → 8.23	1.00 → 1.00	–	–	–	–	–	–	–
qwen1.5b-proverqa-j_gae-s1	200	5.27 → 3.17	1.00 → 1.00	0.434 → 0.053	–	–	–	0.998 → 0.487	0.33 → 0.45	83 → 40
qwen1.5b-proverqa-j_gaed-s1	200	5.21 → 6.99	1.00 → 1.00	0.487 → 0.344	–	–	–	0.998 → 0.834	0.33 → 0.43	83 → 124
qwen1.5b-proverqa-j_hc-s1	200	5.22 → 1.57	1.00 → 1.00	0.391 → 0.236	–	–	–	0.571 → 0.465	0.33 → 0.02	83 → 18
qwen1.5b-proverqa-j_hc_v3-s1	200	5.15 → 1.03	1.00 → 1.00	–	–	–	–	–	–	82 → 16
qwen1.5b-proverqa-prm800k-s1	200	5.30 → 1.00	1.00 → 1.00	0.416 → 0.356	0.741 → 0.736	0.572 → 0.476	0.581 → 0.476	0.538 → 0.476	–	18 → 9
qwen1.5b-proverqa-prm800k_v2-s1	200	5.28 → 3.86	1.00 → 1.00	–	0.738 → 0.675	–	–	–	–	18 → 19
qwen1.5b-proverqa-qwenprm-s1	200	5.27 → 0.00	1.00 → 0.00	0.468 → 0.000	0.624 → 0.000	0.622 → 0.795	0.608 → 0.795	0.558 → 0.795	–	18 → 0
qwen1.5b-proverqa-qwenprm_v2-s1	200	5.22 → 3.72	1.00 → 1.00	–	0.617 → 0.633	–	–	–	–	18 → 21
qwen1.5b-proverqa-s_gae-s1	200	5.19 → 2.95	1.00 → 1.00	0.433 → 0.099	–	–	–	0.997 → 0.678	0.44 → 1.00	10 → 5
qwen1.5b-proverqa-s_gae_v2-s1	200	5.34 → 3.62	1.00 → 1.00	–	–	–	–	–	0.42 → 0.30	10 → 8
qwen1.5b-proverqa-s_gaed-s1	200	5.21 → 5.67	1.00 → 1.00	0.517 → 0.366	–	–	–	0.993 → 0.772	0.44 → 0.64	10 → 19
qwen1.5b-proverqa-s_gaed_v2-s1	200	5.29 → 8.66	1.00 → 1.00	–	–	–	–	–	0.42 → 0.68	10 → 13
qwen1.5b-proverqa-s_gaed_v3-s1	200	5.20 → 97.63	1.00 → 1.00	–	–	–	–	–	–	10 → 290
qwen1.5b-proverqa-s_gaed_v3-s2	200	5.31 → 99.99	1.00 → 1.00	–	–	–	–	–	–	11 → 285
qwen1.5b-proverqa-s_gaed_v3-s3	200	5.26 → 41.49	1.00 → 1.00	–	–	–	–	–	–	11 → 118
qwen1.5b-proverqa-s_hc-s1	200	5.19 → 2.00	1.00 → 1.00	0.466 → 0.486	–	–	–	0.538 → 0.500	0.44 → 1.00	11 → 5
qwen1.5b-proverqa-s_hc_v2-s1	200	5.35 → 6.34	1.00 → 1.00	–	–	–	–	–	0.43 → 0.70	10 → 9
qwen1.5b-proverqa-s_hc_v3-s1	200	5.13 → 1.00	1.00 → 1.00	–	–	–	–	–	–	12 → 6
qwen1.5b-proverqa-versaprm-s1	200	5.22 → 0.00	1.00 → 0.00	0.450 → 0.000	0.706 → 0.000	0.644 → 0.571	0.660 → 0.571	0.574 → 0.553	–	20 → 0
qwen1.5b-proverqa-versaprm_v2-s1	200	5.32 → 5.04	1.00 → 1.00	–	0.704 → 0.723	–	–	–	–	21 → 18
qwen1.5b-proverqa-vp_gae_v2-s1	200	5.29 → 4.18	1.00 → 0.99	–	–	–	–	–	0.20 → 0.20	8 → 4
qwen1.5b-proverqa-vp_gaed_v2-s1	200	5.35 → 4.14	1.00 → 1.00	–	–	–	–	–	0.21 → 0.19	8 → 5
qwen1.5b-proverqa-vp_hc_v2-s1	200	5.25 → 4.15	1.00 → 1.00	–	–	–	–	–	0.19 → 0.19	7 → 6

Queue

Pending, in start order (0): none

Held (0): none

Blocked by a failed dependency (0): none

Failed (0): none

Done (141): 725 GPU-job hours in total (per-job times: python -m tdprm.queue.gpuq status --plan <plan>)

Data and states

training set	pool	base pass@1	kept (solved 1–7 of 8)
hotpotqa_llama3b	7995	0.607	4427
hotpotqa_qwen1.5b	7990	0.134	4045
math_llama3b	7495	0.578	5194
math_qwen1.5b	7493	0.289	4631
proverqa_llama3b	4500	0.445	3793
proverqa_qwen1.5b	4500	0.463	4135