acceptodds
Withdrawn from ICLR 2027

What a Steered Lie Does Inside: The Deception-Belief Entanglement in LLMs

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Activation probes can monitor AI systems by decoding a model's internal belief and flagging public outputs that depart from it. This use requires the decoded belief to stay reliable when the model deceives. We test this requirement in One-Night Ultimate Werewolf, where night-phase card swaps give each seat an observation-grounded believed role that can differ from its true role. A belief probe decodes the model's believed role above 0.84 accuracy. We induce deception in Llama-3.1-8B-Instruct with activation steering added only at public-speech token positions, leaving the generation of private beliefs unchanged. The belief readout shifts monotonically with steering strength, while the model’s self-reported role becomes less confident. Removing the binary Werewolf-vs-Village direction explains only 7-10% of the shift; 83% instead concentrates in three high-gain role axes (Villager, Werewolf, and Tanner), and the steering delta is enriched 2-10 times in these directions relative to a random vector. Perturbations confined to the belief subspace flip the probe in 86-90% of cases, versus 2-7% for random subspaces. Qwen2.5-7B-Instruct shows the same structure, suggesting that belief monitors should be stress-tested against interventions on adjacent features.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Deception, Belief Representation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_nmtWOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The authors investigate whether a belief probe continues to be reliable if a model engages in deception. They study this research question in the controlled setting of "One-Night Ultimate Werewolf". The probed "belief" is the role a LLM adopts within this game. The authors induce deception via steering and find that under strong steering the probed "belief representation" changes.

Critical strengths

The general research question of studying the conditions under which the "beliefs" of a LLM can be accurately measured, is timely and interesting. "One-Night Ultimate Werewolf" is a promising testbed for this research.

Critical weaknesses

Unclear Motivation for the work

I do not understand the motivation for this work. In the introduction, the authors state they would like to test whether the belief of the LLM can still be accurately measured when the LLM deceives. This is a reasonable research question. However, I do not understand why the authors choose to induce deception via steering. Steering is known to have strong side effects on the output quality and the activation distribution. I therefore do not find it surprising that strong steering can lead to a change in the model's activations that affect a probe. The much more natural research avenue would have been to study naturally occurring instances of deception, e.g. during role-play or when the model pursues a goal given in context and deception is instrumentally useful. Why do the authors focus on steering instead?

Conceptual and Methodological issues and Soundness

The authors suggest that deception steering changes the model's internal belief about its role. However, the evidence they provide seems perfectly compatible with a much more mundane explanation: The belief probe tracks the stated role of the model and not necessarily the internally believed role of the model. On the training dataset, the internally believed role of the model and the externally stated role of the model are identical. However, after steering with the deception direction, the model might state having a different role and this leads to a large KL divergence for a probe tracking the stated role of the model. More generally, strong steering leads to a large change in the activations. This might also lead to a measurable KL-divergence. For these experiments to be convincing, more evidence is needed to show that (i) the identified representation indeed corresponds to the models internal belief about its role and not just its stated belief and that (ii) steering along other directions with a similar strength leads to a much lower KL divergence, i.e. a control condition.

Presentation

I found it very hard to follow the paper. Formulations are often very confusing and essential components are not clearly explained. Just to name a few:

  • The paper is centered around the game "One-Night Ultimate Werewolf" but this game is not clearly explained in Section 3.
  • The authors never mention which layers they steer when activation steering.
  • Example of a paragraph that is extremely confusingly written and I was not able to understand it: "The readout’s gain turns this overlap into a measurable shift, and the delta determines where the shift lands. It enters through the identity axes of the two roles whose win conditions reward lying. The lie grazes the belief representation; the readout amplifies the graze" (Line 61-63).
Questions for authors

Clarification

  • Why is it relevant that sometimes the belief of the model and the ground truth diverge when its card is swapped during the night? You mention multiple times that this is important (e.g. "belief and truth are dissociated" in Line 102) but I do not understand why. If the model's internal state is not affected at all by the external ground truth, why does that matter? Accordingly, I do not understand "Control 1" (Line 246). If the LLM has no way of knowing its true role, how could the belief probe possibly track its true role?
  • Please explain the data on which you train the deception probe in subsection 3.3 in more detail.
Reviewer_ADtAOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies the effect of steering activations towards deceptive representations on the decoded belief of the model. The authors study this under the setting of One-Night Ultimate Werewolf. The game has various roles which the model could be playing, and the authors train probes to predict the role an optimal player should be believing given the observation in context. The paper does some control experiments to establish the belief probe is tracking the model's belief. The paper then uses a difference-of-mean vector to steer towards deception and show that 1) it increases the KL divergence of belief probe prediction between the steered activations and the original activations, and 2) the steering affects the villager and werewolf axes of the low dimension belief representation space for initial layers, and through werewolf and tanner axes (the two roles with incentives to deceive) for later layers.

Critical strengths
  1. The setting of One-Night Ultimate Werewolf is good for studying the question of whether steering affects beliefs of the model since belief without steering can be clearly established based on the context.
  2. The result of steering with deception vector affecting the role-belief probe is valuable.
Critical weaknesses
  1. As written, the paper is written very confusingly. A lot of important details about the game and how it is formatted to give to the model, is in appendix. while trivial results that are true by construction are kept in the main paper (e.g., section 5.1, 6.1, 7). There are also terms like "energy" that are just used directly but not defined and are not common in the literature to know without defining. Figure 4 is not readable at all, its font size is too small. The caption on most figures can also do with a single takeaway rather than just describing the figure.
  2. Many key results hold by construction without any comparison vector. KL of belief probe increasing with steering needs a control with an unrelated concept vector. The effect looking low dimensional is by construction as well as the SVD is done on the probe weights. So any perturbations happening in the null space (space outside the vector space of the belief space) cannot be shown with the method.
  3. The steering vector is not shown to encode deception. Rollouts are labelled by the prompt containing written instruction for being honest or deceptive, and the paper doesn't measure the lie rate when the model is instructed to be deceptive. Also, using a deception probe trained on transcripts of the game could have a lot of game-specific confounders like present of role words or specific trigger contexts. The authors should use a general deception probe trained in the literature using generic chat related deception and lying to reduce the confounders.
  4. The section 4.3 control experiment doesn't properly establish internal state vs role word dependence. It only establishes that the belief probe applied on public span is not relying on the role token being present in the public span. However, the actual role token is still in the context as the model says it in the belief span. A better control could be to remove all such role word related confounder from the transcript (e.g., prefill the belief span with a neutral statement) and then check whether the probe on the public span (or even belief span) now captures the believed role.
  5. When the model is deceptive (eg. when instructed or steered), the paper never checks whether the model holds both the real role and stated role (lie) within the internal states. The belief probe can't tell this because it was only trained on data where both stated and real role are the same.
  • One experiment to check this could be: on deceptive turns, prompted and steered, train two probes on the public span activations: one for the believed role (real) and one for the claimed role (lie). If the believed role stays decodable at near-honest accuracy while the claimed role is also decodable, the model holds both. In that case the paper's "contamination" means an honest-trained probe confusing the two variables, not a degraded belief. The monitoring takeaway would also change: train the monitor on deceptive data so it separates the two roles. It would no longer be "belief monitors break under deception".
Questions for authors
  1. Is the model's stating of role mediated through the belief probe representations? This is not established in the paper and would be good to know before trusting whether the belief probe captures the model's internal state of the role. Just doing causal intervention (or ablation) using the belief probe and checking whether the stated role changes should establish this.
  • Relatedly, another experiment to perform to check whether deception is mediated through the belief probe representation change: Remove the belief directions (u₁…u₆) from the injected vector, regenerate the tokens for the self-report, and check whether the self-report role flip disappears.
Reviewer_SdQTOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper uses a 'Werewolf game' to study if steering a model to lie influences probes for its beliefs. The game factors out the role the model thinks it has from its actual role. The authors train a probe on private belief text, and steer only later public-response tokens. The private text stays the same, but the probe’s output shifts. They include several probe controls, analyze which probe directions change and repeat the result on a second model.

Critical strengths

The setup is clever. The game provides a pretty clean belief label, which is otherwise hard to get in deception settings. The authors also do a good job showing that the probe tracks believed role, not just explicit role words.

The intervention is carefully set up. The private belief text is already generated when steering starts, and it remains identical. So this is a real result about a monitor becoming unreliable during public speech. The decomposition into a few probe directions is also useful. It makes the result more than just sating that the probe changed.

Critical weaknesses

The paper goes too far in calling this belief contamination. The clear result is probe contamination. Steering is applied at the public-speech positions where the probe is read. But there may be possible confounds, which are not elimiated. For example. the probe may be picking up the role being claimed, or some other speech-specific feature -- rather than a changed belief.

The self-report experiment does not completely fix this. The model reports its belief after it has made the false claim, so it may just condition on its own earlier text. The scope is also limited: two models, one game and only one kind of self-belief.

Questions for authors

Can you test the model’s role belief or role-dependent behavior in a fresh context (without showing it the steered false claim)?

Can you compare with a matched non-deception steering direction? e.g., one that is not about lying.

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

Related papers

Discussion (0)

Sign in to comment.