Can you pinpoint the exact moment an AI notices it's looking in the wrong place and corrects course?
Can gaze repair cycles be located within reference chains?
This explores whether the moments where attention gets redirected to fix a misunderstanding (a 'gaze repair') can be pinpointed within a chain of references or reasoning steps. The corpus has nothing on human gaze or conversational repair, so it can only speak to the nearest model-internal analogues.
This reads the question as asking whether you can find the exact points in a chain where attention is redirected to fix something that went wrong. The corpus has no material on human eye gaze, conversational repair, or coreference chains, so it can't answer that version directly. It does have related work on where a model 'looks', on self-correction inside reasoning chains, and on how hard it is to tell a real correction from a fake one.
The closest match for 'gaze' is visual attention in multimodal models. Work on perception tasks finds that the bottleneck is Does verbose chain-of-thought actually help multimodal perception tasks? visual attention allocation, not verbalization. Long written rationales can make perception worse. So where a model looks and what it writes about looking are separable, and reading the verbal chain may not show you where the looking was redirected. A related result treats over-attention to irrelevant context as a shared root of several failures, including hallucination and sensitivity to example order (Can fixing attention noise solve multiple LLM failures at once?). That makes attention misallocation something you can measure, not just narrate.
The 'repair cycle' half has a sobering answer. A step that looks like a repair, such as 'wait, let me reconsider', is not evidence that one happened. Frontier reasoning models reach only 20-23% exact match on problems that require genuine backtracking, even though their reflective language is fluent (Can reasoning models actually sustain long-chain reflection?). Fine-tuned models often give the same answer when their reasoning is cut short, paraphrased, or padded with filler, which suggests the steps have become performative (Does fine-tuning disconnect reasoning steps from final answers?). Broader critiques argue that chains imitate the structure of reasoning more than they perform it (Why does chain-of-thought reasoning fail in predictable ways?).
That suggests how you would locate a repair: by intervention, not by reading the text. If you truncate, swap or paraphrase a candidate step and the outcome doesn't change, it wasn't a repair. If it does change, you have found a step that does something.
Internal signals are a second route, though none of the retrieved work shows per-step localization of a repair. A Jacobian-based lens reads the representations a model is poised to verbalize, and can expose strategic reasoning that never reaches the output (Can we read a language model's unspoken thoughts?). Reasoning tasks also trace paths through representation space with two to three times the curvature of simple lexical tasks (Does transformer reasoning leave a geometric signature in representation space?). But that curvature is reported as a task-level average, so it hasn't been shown to mark individual turning points. A gaze-repair detector would have to combine an attention-level signal, a representation-level signal and an intervention test. The corpus has the pieces but nobody has assembled them.
Sources 7 notes
Long rationales and text-token RL help reasoning but hurt fine-grained perception tasks because the actual bottleneck is visual attention allocation, not verbalization. Standard CoT optimization trains the wrong policy target.
Cancelling attention noise through differential attention improves hallucination rates, makes in-context learning robust to example order, and reduces activation outliers—suggesting these separate failure modes share one root cause: over-attention to irrelevant context.
DeepSeek-R1 and o1-preview achieve only 20-23.6% exact match on 850 constraint satisfaction problems requiring genuine backtracking. This ceiling reveals that reflective reasoning fluency does not translate to actual problem-solving competence on unfamiliar instance structures.
Three faithfulness tests show fine-tuned models generate reasoning chains that less reliably influence final outputs. Early termination, paraphrasing, and filler substitution all produce invariant answers more often after fine-tuning, suggesting reasoning becomes performative rather than functional.
CoT guides models to pattern-match reasoning structure rather than perform genuine inference. This explains distribution-bounded failures, why structural coherence matters more than content correctness, and why performance optimizes against interpretability.
Show all 7 sources
The Jacobian lens identifies representations a model is poised to verbalize that exhibit functional signatures of global workspace theory: coherent content in intermediate layers, capacity for tens of concepts, and wider broadcasting. This enables cheap alignment auditing by revealing strategic reasoning even when hidden from output.
Measuring intrinsic geometry across multiple models shows reasoning and analogy tasks carve paths with mean curvature of 0.71–0.83 rad, while lexical tasks produce only 0.27–0.31 rad, suggesting path geometry encodes task difficulty.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Is Chain-of-Thought Reasoning of LLMs a Mirage? A Data Distribution Lens
- When More is Less: Understanding Chain-of-Thought Length in LLMs
- A Comment On "The Illusion of Thinking": Reframing the Reasoning Cliff as an Agentic Gap
- Farther the Shift, Sparser the Representation: Analyzing OOD Mechanisms in LLMs
- Verbalizable Representations Form a Global Workspace in Language Models
- CoT is Not True Reasoning, It Is Just a Tight Constraint to Imitate: A Theory Perspective
- On the Impact of Fine-Tuning on Chain-of-Thought Reasoning
- LR^2Bench: Evaluating Long-chain Reflective Reasoning Capabilities of Large Language Models via Constraint Satisfaction Problems