Research map / Large language models / Reasoning and chain of thought

Reading Between the Dots: Decoding Hidden Computation across Filler Tokens

NeurIPS 2026 · Large language models › Reasoning and chain of thought · approach: Long CoT dynamics and efficiency · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Shows that frontier open-weights LLMs (DeepSeek V3, Kimi K2) compute over filler tokens in a legible way, and proposes an unsupervised logit-lens plus LLM-judge pipeline that recovers hidden intermediate values from filler hidden states.

Problem

Filler-token computation is cited as a limit case for CoT monitorability because visible tokens carry no information about the reasoning; the paper asks whether hidden intermediate computation is readable from the residual stream.

Approach

Apply the logit lens at every (layer, filler position), subtract the cross-example mean to remove artifacts, aggregate residual scores across positions, and have an LLM judge interpret the top-50 tokens to recover intermediate values without labels or training; supported by attention analysis and KV-cache transplants at filler positions.

Key results

• On 1-fact addition DeepSeek V3 accuracy rises from 54% to ~72% with filler; on system of equations DeepSeek V3 rises from 31.1% to 61.0%.
• Unsupervised pipeline recovers intermediate values in 82-94% of examples with best LLM judge, and 35-85% of targets appear in the top-2 decoded tokens without a judge.
• KV-cache transplants at filler positions only shift donor answer rank by tens to hundreds of ranks; for 2-fact addition, transplanting the A1-decoded positions recovers 93% of the whole-filler rank shift.

Surprising findings

• Filler tokens placed before the question give no uplift; benefit requires filler between question and answer.
• On the 2-fact task, the judge's gain is a context-assisted effect: stripping non-numeric tokens drops accuracy to the direct-token rate.

Limitations

• Results limited to cleanly decomposable, non-adversarial tasks with nameable intermediates.
• Logit lens only reads content projecting onto the unembedding; attention-only or orthogonal computation is invisible; analyses mostly use 4-bit quantized checkpoints.
• Only two models analyzed in depth; closed models cannot be tested with the pipeline.

Experimental setup

Datasets: Ryan Greenblatt compose_facts repository (1-fact addition facts) atomic number facts world capitals and chemical element facts (2-hop letter position) synthetic system-of-equations prompts

Benchmarks: 1-fact addition 2-fact addition 2-hop letter position system of equations

Metrics: accuracy top-2 direct recovery rate top-K judge recovery rate rank shift full-swap rate McNemar test

Baselines: no-filler baseline explicit chain-of-thought Qwen 3 480B Shuffled-token baseline Kimi K2.5

Builds on: logit lens patchscope tuned lens activation patching

Closest papers by meaning

• Towards Long-Horizon Interpretability: Efficient and Faithful Multi-Token Attribution for Reasoning LLMs (ICML 2026 · similarity 0.86)
• Beyond Semantics: The Unreasonable Effectiveness of Reasonless Intermediate Tokens (NeurIPS 2026 · similarity 0.86)
• Think Deep, Not Just Long: Measuring LLM Reasoning Effort via Deep-Thinking Tokens (ICML 2026 · similarity 0.86)
• Reasoning Theater: Disentangling Model Beliefs from Chain-of-Thought (ICML 2026 · similarity 0.86)
• Tracing the Traces: Latent Temporal Signals for Efficient and Accurate Reasoning (ICLR 2026 · similarity 0.85)
• Efficient Reasoning with Hidden Thinking (ICML 2026 · similarity 0.85)