acceptodds
Under review as a conference paper at ICLR 2027

Committed by the First Token: Measuring Recovery from Induced Answer Commitment in Chain-of-Thought Reasoning

Abstract

Chain-of-thought is usually read as the process by which a model works out its answer. We ask what a chain does with an answer that is already present when it starts. We prefill the assistant turn with graded yes/no commitments and, by holding the reasoning chain fixed while varying what sits in context, decompose the resulting shift in the elicited answer distribution into a direct path, through which the commitment is read at answer time, and a mediated path, carried by the chain the commitment induced. On BoolQ with Qwen3-4B a plain commitment overturns about half of the correct answers it contradicts, and the mediated path dominates: a chain written under a plain “no”, with the commitment itself removed, displaces the answer by 10.9 logits, against 5.0 for the commitment alone. Step-level probing shows the hand-off directly: the direct pull decays over the chain while the mediated pull grows, the two crossing within five steps. Three interventions confirm the split: masking the commitment from attention during generation drives the mediated effect to zero; withholding the supporting passage leaves the direct path unchanged and roughly doubles the mediated one, so the chain is not carrying the commitment by quoting evidence for it; and a chain generated under the opposite commitment overturns the commitment still beside it, driving compliance from 0.85 to 0.24. The model cannot report an unmarked commitment as foreign, and being told does not reduce compliance; directing the chain does. A forced verification prefix lowers compliance with a wrong commitment in every condition, and attenuating attention to the commitment to a tenth of its weight recovers 79% of the net damage while keeping the commitment readable. On Ministral-8B the effect is larger, not smaller, and the commitment's strength—legible before reasoning begins—no longer orders the outcome by the end of the chain.

Primary area:
interpretability and explainable AI
Keywords:
Large Language Models, Chain-of-Thought Reasoning, Reasoning Models, Mechanistic Interpretability, Model Behavior, Cognitive Bias, Answer Commitment

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_viTcOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how an injected commitment affects LLMs’ CoT and final answers. The core finding is that the commitment migrates from the injected tokens into the reasoning chain generated by the model itself; by the end of the chain, the answer is mainly influenced by “the chain the model wrote,” rather than by the still-readable commitment itself. The paper also uses fixed-chain splicing to decompose direct / mediated effects, masking, attention attenuation, a no-passage ablation, and a second model family.

Critical strengths

The paper is innovative in its approach. It does not limit itself to unfaithful CoT, but attempts to interpret CoT as carrying an injected commitment.

Critical weaknesses
  • The experimental setup has strong limitations. The conclusions are based on very small models and disabled thinking. Whether all the conclusions still hold on larger models and in reasoning modes is worth verifying.
  • The paper consists entirely of behavioral observations of models on a single simple dataset. It lacks observations across a wide range of domains; in particular, it lacks internal mechanistic / representational analysis.
Questions for authors
  1. The introduction raises the problem of multi-turn agentic scenarios, but the full paper only analyzes the simplest setting such as BoolQ, and does not answer the question raised at the beginning. Broader real-world validation across more scenarios is desired.
  2. The main conclusions of this paper could be validated using a flagship reasoning model.
  3. The conclusion of this paper is that the CoT generated after an injected commitment makes the model more inclined to generate content in the direction of this commitment. Since CoT is conditioned on the commitment, this seems like a normal phenomenon, unless the CoT process does not leak any content related to the commitment. It would be best to provide an example and statistics in the paper showing what the model’s actual reasoning chain looks like. For example, a compelling demonstration would be: two CoTs generated with and without commitment, neither of which exposes any commitment at all, both looking like natural reasoning, yet leading to drastically different yes/no logits.
  4. As I understand it, the mask experiment in Section 4.3 is the same as the direct setting defined in Section 3: commitment information + a natural CoT that does not read the commitment. What is the point of this experiment?
  5. Why do the title says "first token"? What is more important, the semantics of the commitment or its location?
1 comment
Area_Chair_avgWOpenReview ↗

Please respond to the authors

Hello!

Just a friendly bump that you should engage with the authors response

Reviewer_QTzxOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper studies what happens when a model is given an answer before it starts reasoning. The authors ask whether the final answer is mainly influenced by this initial answer or by the reasoning written afterward. They generate reasoning chains with and without a prefilled yes/no answer, and then swap or remove the prefill while keeping the reasoning fixed. This allows them to measure the effects of the prefill and the reasoning separately. On BoolQ, a wrong prefilled answer often changes the final answer, and its influence gradually becomes part of the generated reasoning.

The author uses methods like attention mask to validate their results. However, the paper never explain why and how a injected answer will affect CoT, and only few small models(qwen3-4b; ministra-8b) are been tested. The methodological rigor and generalizability of the paper are both questionable.

Critical strengths

The paper’s method on direct/mediated injection could be used to test the self-revise ability of an existing llm.

Critical weaknesses
  1. The paper provides evidence that an injected answer changes the subsequently generated chain-of-thought and that this changed text can influence the final answer. However, this establishes a behavioral pathway rather than an underlying mechanism. The fixed-chain decomposition measures the model’s responses to artificially constructed contexts, while attention masking only shows that the injected tokens must be accessible during generation for their influence to appear in the chain. Neither experiment explains why the model adopts the injected answer, how this information is internally represented, or which model components carry it across reasoning steps. Moreover, the entire generated chain is treated as a single mediator, even though commitment-induced and clean chains may differ in length, format, explicit answer words, conclusion statements, and reasoning content. Without controlling these properties or identifying which parts of the chain carry the effect, the mediated effect may simply reflect repetition or semantic continuation of the injected answer. Therefore, “commitment migration” is better understood as an operational description of the observed behavior than as a mechanistic explanation. Given prior work on prefilling, priming, and unfaithful chain-of-thought, the main novelty appears to be the experimental decomposition rather than a new explanation of the phenomenon.

  2. The empirical evaluation is too narrow to support broad claims about chain-of-thought reasoning. Most of the detailed experiments are conducted on Qwen3-4B with thinking disabled and greedy decoding, using the first 2,000 examples from a single binary question-answering dataset, BoolQ. Ministral-8B is used only for a partial replication: several central analyses, including the passage ablation, masking, attenuation, and other intervention studies, are not fully repeated on this model. The two models are also relatively small and already show meaningful differences. In particular, the stronger result that the mediated effect clearly dominates the direct effect does not consistently transfer to Ministral. The results are further sensitive to chat-format details: Ministral requires an additional newline before it continues reasoning, and the paper notes that changing the assistant-turn opening can substantially change answer behavior. The paper does not test larger models, modern reasoning models, sampling-based decoding, or tasks with different reasoning and answer formats. Therefore, the current evidence supports a finding about two small models in a narrow binary setting, rather than a general property of language-model reasoning.

  3. Different strength of injection is not defined well. The paper treats phrases such as “I think the answer is probably yes,” “The answer is yes,” and “The answer is definitely yes” as different levels of commitment strength, and the mediated effect is entirely based on chains generated under the direct commitment. It does not control the CoT’s structure, reasoning logics and stepwise confidence which could show a more detailed intervention effects.

Questions for authors
  1. The commitment-induced chain is produced from the original injected commitment and may already repeat or justify the same answer. In the natural generation process, the prefix and its chain are therefore not independent. How do the authors justify interpreting the prefix-only and chain-only counterfactuals as separate causal contributions to the original prediction? Could the reported mediated effect simply reflect the same injected information being repeated in later text?

  2. Attention-masking control. Please clarify the exact masking procedure, including how the attention mask, KV cache, and position IDs are handled during generation and during the final answer probe. At answer time, the commitment is made visible next to a chain generated without access to it, so the final sequence still appears to have the same off-policy mismatch as the spliced direct condition. What additional concern does masking resolve beyond showing that the masked chain is similar to the clean chain? Also, why does the α=0\alpha=0 condition cause about 14% of chains to lose the requested format, while the hard-masking experiment is described as behaving like the unprefilled condition?

  3. The main analysis is limited to Qwen3-4B on BoolQ, with only a partial replication on Ministral-8B. Since BoolQ is a binary Yes/No task and both tested models are relatively small, it is unclear whether the reported behavior generalizes to larger reasoning models or to other tasks. Could the authors report whether the main results including answer overturning, and the direct/mediated decomposition, hold across additional model families, model sizes, and non-binary reasoning datasets?

1 comment
Area_Chair_avgWOpenReview ↗

Please respond to the authors

Hello!

Just a friendly bump that you should engage with the authors response

Reviewer_fHC9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors prefill the start of the assistant turn with a yes/no answer at graded strengths: hedged ("probably"), plain ("The answer is no."), emphatic ("definitely"), plus a variant that invites verification and a neutral control. The task is BoolQ (2,000 yes/no questions with passages), run on Qwen3-4B with thinking turned off and greedy decoding. They splice chains across conditions to split the effect into two paths: 1. Direct path where the commitment placed in front of the chain the model writes without any prefill, and (2) Mediated path: the chain written under the commitment, with the commitment deleted. The authors find that a plain commitment overturns about half of the correct answers it contradicts. The effect of the mediated path is larger: for a plain "no", the chain alone moves the answer 10.9 logits vs 5.0 for the commitment alone. The work then studies some useful and un-useful defense methods. The conclusion is the reasoning chain works as a vehicle for the commitment rather than a check on it. The reasoning becomes the channel through which a planted answer reaches the output.

Critical strengths
  1. This work studies an interesting and important problem: whether the reasoning chain checks a prefilled answer or carries it forward. On Qwen3-4B, it mostly carries it: the chain written under a commitment, with the commitment deleted, moves the answer more than the commitment alone does.

  2. Insightful analysis: probing at each step shows the influence moving from the injected tokens into the generated ones.

  3. Thorough ablations: the authors conduct several independent checks of the decomposition including "Attention masking during generation", "Same-code-path control" and "Opposite-polarity chains".

  4. The authors studies several useful and useless defenses against prefilled commitment. For example, moving the commitment into the user turn, or attributing it to a person doesn't change compliance, while forced "Let me check the passage." prefix lowers compliance with a wrong commitment in every condition.

Critical weaknesses
  1. The scope is narrow, but the claims are over-general. The evidence is one task (BoolQ, yes/no), two small models (Qwen3-4B and Ministral-8B), one prompt template and greedy decoding only.

  2. The paper is about chain-of-thought, yet it disables the reasoning mode the model was trained with. Instead it uses prompted "Step N:" chains, which are short (Ministral averages 3.0–3.6 steps). Reasoning-trained models, with long thinking traces and trained self-checks like "wait, let me re-check", are exactly where a chain might act as a check rather than a vehicle.

  3. The choice of benchmark is questionable. BoolQ may not need reasoning at all and the answer is usually a lookup in the passage. If the chain does no real computation, it's unsurprising that it just carries whatever answer it starts with.

  4. The mediated path might just be the chain restating the answer. If the chain written after "The answer is no" contains, or ends with, a line like "So the answer is no", then deleting the prefill doesn't really remove the commitment. A copy of it is still in the chain, and the "mediated effect" is just that copy being read at answer time.

  5. The experimental design of step-level analysis is questionable. Chains have different lengths, so later steps average over fewer questions. Figure 1 has no error bands and no comparison restricted to chains of the same length. Forcing "Final answer:" in mid-chain is off-distribution, and the paper doesn't check that it measures what the model would conclude.

  6. The proposed defenses have limited practical value. Attention attenuation and masking require knowing which tokens are the injected commitment. In the threat model that motivates the paper (prefill attacks, seeded agent responses), if you knew that you could just delete it. On the other hand, the forced-verification prefix and the revision licence only partly help at the plain level (0.555 → 0.431).

Questions for authors

Please see my weakness points above. Moreover:

  1. How often the answer is restated? And the mediated effect after removing "re-statements" or probing before the concluding step?

  2. Instead of using greedy decoding and forcing the answer after different steps, how about use a decode temperature and hence study the "natural" chain length in the step-level analysis. Also please consider adding confidence bands.

  3. Did you try to not always have the commitment added at the beginning of the system turn? What if these commitment are randomly added during the CoT generations?

1 comment
Area_Chair_avgWOpenReview ↗

Please respond to the authors!

Hello!

Just a friendly bump that you should engage with the authors response

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.