acceptodds
Withdrawn from ICLR 2027

Auditing Cross-Lingual Chain-of-Thought Probes with Leakage Controls

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

10 views

Abstract

Chain-of-thought (CoT) monitors read a model's written reasoning to catch misbehavior. Outside English, that text often fails: the model can follow a planted wrong hint and never mention it. A common fix is to look inside the model instead. Train a linear probe on English hidden states, then reuse it in other languages, assuming the internal "about to follow the hint" state is shared. We test that idea on a misleading-hint task in eight languages and four scripts, using three open models: Qwen3-8B (thinking), plus Apertus-8B and EuroLLM-9B (instruct). We also extend the prediction check to a same-content Azerbaijani Latin/Perso-Arabic script swap and to a wider set of thinking models across size and family (Qwen3 1.7B/32B, R1-Distill-Llama-8B, Phi-4 Reasoning). After holding out question IDs, the English probe still predicts hint-following at post-CoT (mean non-English AUROC 0.69–0.79), and removing a linear language signal barely changes transfer. But much of what it reads is already written: filtering early answer letters leaves Qwen3-8B near 0.61, while some instruct-model scores fall to chance; cutting the end of the reasoning drops all three late-probe means to 0.52–0.56. And the same English directions do not control the behavior. Adding or subtracting them during generation is no better than a same-size random direction in English, Turkish, Persian, or Swahili; stronger and multi-layer injections also fail to give reliable signed effects. So an English probe can monitor hint-following across languages, but the directions we test cannot steer it.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
chain-of-thought monitoring, CoT faithfulness, AI safety, interpretability, linear probes, activation steering, cross-lingual transfer, multilingual large language models

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_NUiBOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper attempts to train multilingual (transfer) probes to detect hint-following in the context of chain-of-thought (CoT) reasoning. The authors take multiple-choice QA (MCQA) items and add a hint pointing towards a wrong answer; a model is deemed to be following the hint if it selects that wrong answer. The authors then train probes on English MCQA and use them on 1) unseen English items and 2) unseen non-English items. The probes are trained at various points in the CoT - the prompt, early CoT, and post-CoT. The probes only work well in the post-CoT setting, and also only really on Apertus and EuroLLM. They do transfer somewhat across languages, but the post-CoT probes aren't consistently better than an LLM judge, suggesting they just use info available in the text, rather than privileged internals-only info. They also don't work for steering. The intended conclusions are not clear to me.

Critical strengths

I think that this question - can we detect unfaithfulness in chain-of-thought reasoning - is valuable and interesting.

Critical weaknesses

Low impact due to low novelty, narrow focus, and inconclusive / negative results: This paper is, as it says, an audit. It digs a bit deeper than prior work into multilingual CoT monitoring probes, which is already a rather niche topic. It trains probe, but (Figure 1, Qwen) they only work well on 2/3 models. The authors then spend many pages just trying to determine if these CoT probes actually do transfer, and if they outperform a text-based LLM-judge baseline (they do not conclusively do the latter). The probes are also not useful for steering. In sum, this paper describes an inability to train probes for CoT hint-following.

To be clear, I don't want to claim that negative results are unimportant. But results like these, which amount essentially to "we failed to train a multilingual, causally effective CoT probe" are not very useful for the community at large. As discussed in the next point, they reflect issues with the authors' training setups more than they reflect some fundamental impossibility of training such probes. In the future, I would suggest that the authors focus more on finding setups that do work, than describing those that don't.

Experimental design issues: I think that at least some of the issues encountered by the authors are mainly due to experimental design. In particular, I don't think it makes much sense to probe at the prompt position (the model has probably not decided whether to hint-follow at that stage). Similarly, at the post-CoT position, it makes sense that the probes should do well but rely on information in the text; even if you cut off the model's final answer, surely it makes clear if it's going to output the final answer or not in its reasoning (either explicitly, with the answer choice, or not). I'm frankly only surprised that they don't do better

I think that what you probably want to detect is something like "the model is using the hint when generating this line of reasoning". That is, it's something that happens actively during reasoning, rather than at the end or necessarily in the very beginning. It might not even be easy to localize. But this is what I think is most likely to be findable, and not just be an artifact of answer presence.

This would also solve your steering issue: you probably can't steer using these probes because they're just picking up answer presence. But if you were to steer on the probes when your model was actively generating hint-biased reasoning, this would probably work much better. Perhaps it's worth investigating the raw reasoning traces in greater detail to see whether a) the intended answer is obviously visible (probably, since the LLM judges often pick it up) and b) you can somehow identify lines of biased reasoning (perhaps an LLM can find these too)

Questions for authors

Clarification:

  • What are the baseline accuracies of these models in English / other languages on these MCQA questions? Do you filter to only questions that the models get correct in the first place?
  • I find it difficult to understand the AUROC values that the authors report in the paper, and I struggle in particular to understand how they are framed. What constitutes, in the authors' opinion, a "good" AUROC on this task?

Suggestions:

  • This paper is a bit light on content; it gets deep into the weeds of controls and experimental details for experiments with ultimately negative results. Some of these would be better left in an appendix, freeing up room for more
  • More figures in the main text, fewer big tables
  • Use AI less; this paper reads as quite AI-written. This is visible in the paper's peculiar phrasing, like "An English probe can detect hint-following in other languages. Sharing the same questions across languages can fake transfer (up to +0.16 AUROC)." This is quite confusing for humans - what it means to "fake transfer" is not immediately obvious, and this is an unclear way of expressing this concept in the intro. Many other cases of this exist in the paper.
  • The language / LEACE experiments are not very informative, and can probably be cut
Reviewer_TRgKOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper tests English-trained probes for wrong-hint following across eight languages to study whether the hidden state can be taken as a useful signal in CoT. It controls shared question IDs, tests language erasure, and edits reasoning text. Late probe scores transfer across languages but depend strongly on the generated text. The tested steering directions do not give reliable control. The main message is that cross-language prediction does not by itself show an early decision signal or a useful control direction.

Critical strengths
  1. Question-level splits address a real leakage risk. Keeping translations and repeated samples of each question in one fold makes the transfer test more credible than a split over individual outputs.
  2. Useful findings in the negative conclusion of the probe in cross-lingual CoTs. The paper finds the probe's effectiveness was based on some early-level answers, and intervention towards better answers is hard to achieve.
Critical weaknesses
  1. Narrow questions and corresponding limited findings. In the authors' discussion of related work, there already exists work that finishes the exploration toward probe or internal signal faithfulness about CoT and reasoning; adding the restriction of multilingual does not bring a significant contribution. It should be an extension of these works. The targeted research question can not bring an ICLR main conference-level contribution.

  2. Test results help select the layer. Choosing layers by mean AUC across the target languages weakens the zero-shot claim and may inflate the reported scores. Selection should use separate development data, ideally English-only for this claim.

Questions for authors
  1. Compared to existing internal signal faithfulness work about reasoning or CoT, does your work get obviously different findings that is important to the community?
Reviewer_RPVSOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper seeks to train a classifier on situations where a model uses a misleading hint. They find that the classifier generalizes to question-answer pairs in non-English languages, and that it is difficult to learn a steering vector which causes the model to agree with misleading hints.

Critical strengths

Broadly speaking, the methodology is clearly explained, such as in the section "Models and read positions" towards the end of page 3. The approach to training a classifier in activation space is a natural one to answer the questions that this paper seeks to answer.

Critical weaknesses

The methodology does not isolate the desired behavior: If there is a hint that the answer has a particular value, the model could output that value because it either is following the hint or because it happens to independently agree with the content of the hint. Furthermore, the paper does not filter based on whether the model actually mentions in the CoT that it used the hint.

This causes a semantic shift away from the unfaithfulness motivation. If I am understanding this methodology correctly, the classifier, instead of measuring unfaithfulness, may be measuring a direction which quantifies the degree to which the model agrees with an incorrect hint.

(One could argue that measuring unfaithfulness is not the goal of the paper, but then it is misleading to have it centered in the first paragraph of the intro as the core motivation.)

Also, I could not tell with high confidence from the paper text, but it seems that the hint might remain in English across languages. If this were true, it would be another methodological problem, because there would be a shift in distribution between the train and test sets. The difference is that in the test set, the data points abruptly switch language, which I could imagine confounds the results. Perhaps this is partially mitigated by removing the linear subspace that identifies the language -- this point would need to be argued or demonstrated.

But methodology aside, even if this were measuring the desired phenomenon correctly, it does not seem to be an important finding. Perhaps it would be an important way to show lack of robustness if the classifier did not work across languages. But to convincingly show robustness the paper would need to demonstrate the behavior for a much wider set of possible dataset perturbations besides specifically language shift.

Questions for authors
  • What is the motivation for checking for a classifier that is robust to this behavior across languages?

    • If this is about data scarcity, then perhaps the amount of training data is not a realistic bottleneck for a linear probe.
  • If you are targeting robustness, there are many other kinds of syntactic or semantic perturbation you could have chosen, so why this type in particular?

  • Do you keep the hints in English between test-set languages?

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

Related papers

Discussion (0)

Sign in to comment.