nla-thought-anchors-hash-rollouts

Sentence-level NLA rollout dataset — activations extracted at the #### token.

For each of 1319 GSM8K test examples, 40 NLA descriptions were sampled from kitft/nla-qwen2.5-7b-L20-av (temperature=0.7) using the layer-20 residual stream activation at the #### token. Each description was split into sentences and scored with kitft/nla-qwen2.5-7b-L20-ar.

See Realmbird/nla-thought-anchors-answer-rollouts for the answer-token variant.

Stats

Examples	1319
Rollouts per example	40
Total sentence rows	175,240
Mean sentences per rollout	~3.3

Schema

Column	Type	Description
example_idx	int32	Index in GSM8K test split
is_correct	bool	Whether model answer matched gold
gold_answer	string	Ground-truth answer
rollout_idx	int32	0–39
sentence_idx	int32	Sentence position within this rollout
sentence_text	string	The sentence text
sentence_embedding	list<float32> len=384	all-MiniLM-L6-v2 embedding
cos_sim_to_full_nla	float32	Cos-sim of sentence to its own rollout's full description
critic_cos_sim	float32	NLA critic reconstruction cos-sim vs original activation
critic_mse	float32	NLA critic reconstruction MSE vs original activation
contains_gold_answer	bool	Sentence contains the gold answer number
contains_pred_answer	bool	Sentence contains the predicted answer number

Key findings

Reconstruction quality by correctness:

Group	Mean critic cos-sim
Correct examples	0.604
Incorrect examples	0.416

Thought anchor analysis — sentences stable across all 40 rollouts (high cos-sim to cross-rollout centroid) reconstruct significantly better:

	Correct	Incorrect
Peripheral (Q1 anchor score)	0.503	0.366
Anchor (Q4 anchor score)	0.661	0.673

Anchor sentences from incorrect examples (0.673) reconstruct better than peripheral sentences from correct ones (0.503) — stable recurring content carries the activation signal regardless of correctness.

Computing anchor scores

Anchor score is not stored — compute post-hoc from sentence_embedding:

from datasets import load_dataset
import numpy as np
from sklearn.metrics.pairwise import cosine_similarity
import pandas as pd

df = load_dataset("Realmbird/nla-thought-anchors-hash-rollouts", split="train").to_pandas()
embs = np.stack(df["sentence_embedding"].values)

anchor_scores = np.zeros(len(df))
for ex_idx, group in df.groupby("example_idx"):
    idx = group.index
    centroid = embs[idx].mean(axis=0)
    centroid /= np.linalg.norm(centroid) + 1e-12
    anchor_scores[idx] = cosine_similarity(embs[idx], centroid.reshape(1,-1)).flatten()

df["anchor_score"] = anchor_scores

Related

• Answer-token variant: Realmbird/nla-thought-anchors-answer-rollouts
• Step 1 activations: Realmbird/nla-thought-anchors-hash
• Step 2 NLA descriptions: Realmbird/nla-thought-anchors-hash-step2
• NLA actor: kitft/nla-qwen2.5-7b-L20-av
• NLA critic: kitft/nla-qwen2.5-7b-L20-ar