JonesLin/next-jev-tetris-demo

next-jev Tetris demo (GPU, seed 42)

One timed Tetris game played by the next-jev choice-head model (checkpoint stage2-choice/checkpoint-00006056, based on Qwen3.5-0.8B). Seed 42, up to 1000 pieces, 1 loop.

Warning: the decoded reasoning text in reasoning/ is NOT reliable. It does not match the board or explain the model's choices (see Reasoning checks). It is here for analysis only. Do not present it as the model's reasoning.

Result

Pieces	1000 (game stopped at the limit, no top-out)
Lines cleared	398
Whole game	9.52 s
Hardware	NVIDIA H100 NVL; bf16, fused projections, torch.compile, CUDA graphs

Per-piece time (ms), 1000 model calls:

Part	mean	p50	p95
model: token ids to option scores (the decision)	3.75	3.76	3.80
server_preprocess: prompt to token ids	1.31	1.32	1.39
server_total	5.19	5.20	5.31
network_and_http: localhost HTTP overhead	2.96	2.73	4.53
step_total: whole piece, client and server	9.19	9.04	10.90

Reasoning decode was run after the game, so it is not in any of these times.

Files

File	Content
tetris-gpu-seed42-realtime.mp4	The game at real speed (1x, about 9.5 s). Board after each piece, decision time. No reasoning text.
game/tetris-gpu-seed42.jsonl	One row per piece: piece, next piece, number of candidates, chosen id, its heuristic rank, confidence, lines, all timings (ms), board after the move (20 rows, top first, . = empty).
game/tetris-gpu-seed42.csv	Same timings as a table.
game/tetris-gpu-seed42.summary.json	Mean, p50, p95, min, max of every timing.
game/tetris-gpu-seed42.requests.jsonl	The exact model input for each piece (state, questions), plus chosen and placement (x, y, rotation).
reasoning/loop1.jsonl	Decoded text, 1 loop (same as the game).
reasoning/loops3-last.jsonl	Decoded text, 3 loops, decoder reads loop 3 (as in training).
reasoning/loops3-concat.jsonl	Decoded text, 3 loops, decoder reads loops 1 to 3 concatenated (never seen in training).
reasoning/checks.txt	Output of the checks below.

Each reasoning row: step, chosen (game), replay_top (top option when re-run offline), probs (all options), reasoning.

How the reasoning text is made

The model does not write text when it decides. It reasons in 128 hidden "workspace" vectors. During training, a small side decoder (aux_decoder) learned to turn the last loop's workspace into a teacher's rationale. After the game, each saved request is re-run (eager, bf16 autocast). The replay picks the same option as the game in 1000 of 1000 steps. Then the text is greedy-decoded from the workspace.

This decoder is weak. With the correct previous words given, it predicts only about 40% of the next words (loss about 3.0, measured on 10 Tetris eval questions).

Reasoning checks

Rule checks against the model's own input: board facts in the first sentence, and whether the text names the chosen move. The table uses the 668 steps where the model is sure (p > 0.9).

Check	loop1	loops3-last	loops3-concat	Baseline
Stack height right	43%	61%	58%	always "very low": 51%
Surface right	61%	55%	63%	always "slightly uneven": 66%
Holes right	24%	23%	28%
All 3 right	9%	4%	12%
Names the chosen move's columns	9%	7%	16%	a random other option: 10% / 6% / 15%
"X is the only placement that clears a line" is right	4 of 125	1 of 26	6 of 101

In short, the text names the chosen move no more often than a random other move. So it does not explain the choice. Better text needs a better-trained decoder.

Newer model: stage 2 last-block decoder, step 3100

See stage2-last-block-3100/: the same game with the new decoder (366 pieces), a raw slow video with the decoded reasoning for every piece, and an LLM judge's verdict for every step (0 of 366 reasonable, 20 partly reasonable). The same model on the stage 2 eval set without Tetris (timing, scores, decoded reasoning) is in stage2-last-block-3100/eval-other-domains/.