Reading Between the Dots: Decoding Hidden Computation across Filler Tokens

2026-07-03

AI Summary

This work demonstrates that large language models perform structured, interpretable implicit multi-step reasoning—even in regions without explicit chain-of-thought tokens such as ellipses—rendering their internal computations difficult to supervise. By analyzing hidden states, the study provides the first evidence that state-of-the-art models conduct coherent implicit reasoning in these regions and introduces an unsupervised decoding method that requires neither labels nor additional training to accurately reconstruct intermediate reasoning values directly from the residual stream. Integrating techniques including attention analysis, logit lens probing, and KV cache transplantation, the approach achieves 80–95% accuracy in recovering intermediate values across four distinct tasks on two open-source large language models, thereby establishing the feasibility of monitoring complete reasoning trajectories.

Abstract

Frontier LLMs can perform multi-step reasoning over content-free filler tokens like dots or counting sequences, producing correct answers with no visible chain-of-thought (CoT). This is a limit case for behavioral oversight, where surface tokens carry no information about the underlying reasoning. But hidden from the output is not the same as hidden from us. On four task families (fact retrieval, parallel numeric composition, string manipulation, and in-context computation), two open-weights frontier models (DeepSeek V3, Kimi K2) compute over filler tokens in a structured, legible way: attention routes the question through the filler region to the answer, logit-lens readouts show retrieved facts emerging early and their composition crystallizing in late layers, and KV-cache transplants at filler positions causally swap outputs between examples. We introduce an unsupervised decoding pipeline that takes only hidden states as input and recovers intermediate values with 80-95% accuracy (best LLM judge) across both models and all four tasks, without ground-truth labels or training. Hidden computation that defeats behavioral CoT monitoring is, on these tasks, directly readable from the residual stream, suggesting monitorability is a property of the model's full computational trace, not just its surface tokens.

Problem

• hidden computation
• filler tokens
• chain-of-thought
• behavioral oversight
• LLM interpretability

Innovation

• hidden computation
• filler tokens
• unsupervised decoding
• residual stream
• KV-cache transplant