- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper studies how explicit reasoning affects models’ choices between honest and deceptive actions in moral dilemmas. The authors report that reasoning generally increases preference for honest choices, while deceptive choices are less stable under paraphrasing, resampling, and activation noise. They propose a landscape explanation: reasoning moves models toward broader and deeper honest attractor basins. They support this account with hidden-state interpolation and reasoning-trajectory analyses.
Critical strengths
- Controlled evaluation setting. The datasets vary the cost of honesty, allowing a focused study of trade-offs between honesty and competing goals.
- Useful empirical coverage. The experiments span multiple model families, sizes, datasets, and reasoning budgets. This gives the main behavioral finding value beyond isolated examples.
- An interesting mechanistic hypothesis. The landscape account connects reasoning and behavioral stability. Studying input, output, and activation perturbations provides several complementary ways to investigate it, although the causal explanation remains unproven.
Critical weaknesses
W1: The paper's first two findings have substantial prior work. Neither is unique to this paper or reported here for the first time.
- a. Reasoning improves honesty.
Deliberative Alignment and later work on anti-scheming have already used reasoning to improve safety and reduce misbehavior. Prior work has also directly studied how reasoning changes the trade-off between honesty and other goals. However, these studies do not consistently find that reasoning improves honesty. For example, Liu et al. find that CoT shifts models toward helpfulness. The potential contribution here is more specific: in certain moral dilemma tasks, explicit reasoning increases a fixed model's preference for the honest option. It is not the first study to establish a link between reasoning and honesty. Guan et al., 2024, Schoen et al., 2025, Liu et al., 2024
- b. Reasoning and outputs can be inconsistent.
To my knowledge, many studies have directly examined this phenomenon. They have tested faithfulness through biased inputs, interventions on CoT content, and reasoning models' failure to disclose cues they actually use. At most, this paper adds a specific observation: under its tasks and evaluation protocol, deceptive choices are harder to predict from reasoning traces than honest choices. Turpin et al., 2023, Lanham et al., 2023, Chen et al., 2025
W2: The landscape theory is not convincingly established
As a study of mechanisms, the paper's central and most insightful potential contribution is its landscape theory. This contribution depends on strong validation. My main concern is that the experiments support some differences in stability, but do not establish the full explanation. In particular, they do not firmly show that honesty has broader and deeper attractor basins, and that these basins explain why reasoning shifts models toward honesty.
- a. The authors have not adequately ruled out simpler explanations.
The most important issue is the base-rate effect in resampling. Suppose a model independently chooses honesty with an 80% probability on each run. Upon resampling, the conditional probability of switching from deception to honesty is naturally 80%. The reverse probability is 20%. This asymmetry requires no explanation in terms of attractor basins.
The authors respond that the overall honesty rate also increases. This needs further clarification. If the baseline and later samples are independent draws from the same distribution under the same conditions, resampling alone cannot increase the expected honesty rate. The baseline generation procedure, sample selection, and weighting therefore need to be checked. The analysis should also be repeated with different baseline seeds.
The other two control analyses are also insufficient. Low overlap in the examples that flip across models does not rule out interactions between dataset features and model differences. Dividing examples into confidence >0.9 and <0.75 groups only creates broad confidence bins. It does not match the confidence distributions. The claim that dataset bias and confidence disparity have been ruled out is therefore too strong.
- b. The core measures have not been shown to identify attractor width and depth.
The authors infer basin width from survival rates along SLERP paths between hidden states. They infer basin depth from the length of reasoning segments that maintain the same choice. However:
- Connectivity along an interpolation path is not the same as basin width. It also depends on the distance between endpoints, the output margin, the interpolation path, and the final readout layer. With a standard linear readout, connectivity within a region that predicts the same answer may follow directly from the classifier's geometry.
- A choice persisting for longer does not imply a deeper basin. Segment length also depends on total reasoning length, stopping rules, and the selection of trajectories based on their final answers. The experiments do not directly measure recovery, escape, or attraction after a perturbation.
The interpolation experiment has another direct problem with its metric. The original answer token's absolute probability falling below 0.5 does not mean that the other option becomes preferred. Probability mass may shift to other tokens rather than to the honest answer.
These experiments are therefore better interpreted as showing differences in decision stability under a particular measurement protocol. To claim broader and deeper attractor basins, the authors need to validate the link between these proxy measures and the proposed dynamics.
c. Even if the stability differences are real, the paper has not shown that they explain why reasoning improves honesty.
The current argument is that reasoning improves honesty, and some noise interventions also favor honesty. The authors then treat reasoning as a noisy traversal that lets the model escape deceptive regions. However, similar behavioral effects do not establish a shared mechanism.
The authors try to rule out a semantic explanation by showing that raters struggle to predict deceptive choices from CoT. But difficulty in prediction does not mean that the content has no causal effect. In fact, 97% accuracy on honest examples and 53% on deceptive examples imply about 75% overall accuracy on a balanced sample. These results cannot be summarized as showing that the reasoning content carries no predictive information.
The paper also finds that an instruction to reason can shift some answers toward honesty before any reasoning tokens are generated. This shows that the prompt change itself already contributes to the effect.
What is missing is a set of experiments that separates the effects of the prompt, semantic content, computation length, and perturbation. With the prompt held fixed, the authors should compare reasoning of different lengths, neutral text of matched length, and reasoning that supports the opposite choice. They should then test whether independently measured stability predicts the honesty gains from reasoning. They should also test whether interventions on stability change those gains.
My main assessment is that the paper offers an explanation consistent with some of its results, but has not established it as the mechanism producing those results. The appendix also reports equal interpolation stability for the two classes in some models. A single application of activation noise generally has no significant effect. These results call for narrower claims about how broadly the proposed explanation applies.
W3: The landscape theory is not entirely new.
Prior work has explored similar frameworks in which different types of representations or behaviors have different levels of stability. Please clarify how this paper relates to the following studies.
- CCPS (2025) perturbs final hidden states and uses stability features to predict answer correctness. These features include the perturbation strength needed to change a prediction. This is directly related to the paper's use of representational stability to distinguish two classes of answers. Paper
- Chia & Pan (2025) explicitly interpret hidden activations in terms of safe and jailbroken states, attractor basins, and potential barriers. They study state transitions through targeted perturbations. This is close to both the landscape account and the framework used in this paper. Paper
- Peng et al. (2024) and AsFT (2025) study the stability of safety under weight perturbations and the idea of a safety basin. They provide related landscape frameworks in parameter space. Peng et al., AsFT
- SmoothLLM (2023) improves safety by exploiting the sensitivity of jailbreak behavior to input perturbations. It provides a behavioral precedent for the observation that harmful behavior is more easily disrupted by perturbations. Paper
The omission of these related studies is not appropriate.
Questions for authors
See weaknesses.