An auditable framework is built that maintains an external belief state over hidden roles, logs belief updates and belief-action deviations as structured evidence, and supports a defensive offline improvement loop that reviews bad cases before any strategy change.
Abstract
Evaluating LLM agents in hidden-information multi-agent settings is hard: final outcomes are high-variance and rarely reveal why an agent decided as it did. We study this in a 9-player Werewolf environment where agents act under strict, code-level information isolation, and we build an auditable framework that maintains an external belief state over hidden roles, logs belief updates and belief-action deviations as structured evidence, and supports a defensive offline improvement loop that reviews bad cases before any strategy change. Across 1,080 frozen games spanning belief-disabled, active-belief, kernel-ablation, camp-restricted, consumption-policy, and high-load arms, and including a seed-paired A0/A1 comparison, the active-belief condition is associated with substantially better good-side outcomes: in the 200-seed A0/A1 comparison the good-side win rate rises from 0.205 to 0.390 (paired McNemar $\chi^2 = 16.4$, $p<0.001$), with fewer irreversible witch-poison errors. We do not, however, attribute this shift to belief content. Direct action-belief consistency is low ($\approx 0.21$), and giving belief only to the werewolves helps the good side more than giving it only to the good side, which argues against a simple holder-benefit account; we therefore report the effect as an association and treat its mechanism as unresolved. The contribution is the audit framework itself: it makes the effect measurable, exposes low direct action-belief consistency, rejects an unreliable forced-consumption intervention with evidence, and separates strategy effects from load confounds. We accordingly position external belief in high-noise hidden-information games primarily as an auditable cognitive baseline that also carries decision-relevant signal, turning opaque agent behavior into replayable evidence for safer, controlled iteration.
Agentic systems increasingly gate actions on a model's own stated confidence, which assumes confidence tracks correctness at the moment of acting. We test this in a hidden-information chess variant where royal status can be secretly, repeatedly relocated between pieces, and where an agent's stated probability distribution over the opponent's hidden royal piece -- elicited every turn, separately from the move it chooses -- is scored against ground truth recoverable after the game. Across two independent batches, captures made at high stated confidence ($\geq 0.5$) about the hidden piece's location were correct in 1 of 62 cases. The calibration deficit is concentrated almost entirely in these events: 99.3% of it in the original batch, 98.7% in the replication. The same pattern, in weaker form, orders consistently (point estimates only; most pairwise gaps are not statistically distinguishable at this sample size) across four further model configurations spanning a second provider -- reported as scope for the finding, not as evidence that capability predicts calibration: a same-model comparison at a fixed external leaderboard score shows a deliberation-budget change alone moves the metric by nearly as much as a large cross-model gap. In a separate seat, conventional evaluation axes -- legality, cost, latency, completion rate -- can dissociate entirely from belief quality, with the configuration winning on every conventional axis producing the worst belief quality tested. A model exhibiting this pattern can still win the game its belief was about, which is why outcome-only evaluation would not detect it.
Current LLM agent systems decide delegation before reasoning begins (a router picks a model) or after a response is complete (a verifier scores it and may retry). We study a third regime: an agent that recognises, during its own reasoning, that it is unlikely to succeed and transfers control to a stronger model. We formulate intra-generation delegation as a Bayesian optimal-stopping problem over a learned competence posterior -- an online estimate of the agent's eventual task success whose sufficient statistics are learned from labelled trajectories, not read off raw entropy. We derive the myopic escalation threshold in closed form, characterise the optimal policy via dynamic programming, and prove that the optimal policy is a time-varying threshold with no shape assumption on the raw signal. We further prove exponential separation of the oracle belief at the Chernoff-information rate of the signal, a regret bound governed by the calibration of the posterior, and a finite-sample guarantee: with n labelled calibration trajectories the deployed plug-in policy's regret decays as 1/sqrt(n). A controlled simulation study confirms each prediction of the theory, including the predicted 1/sqrt(n) rate. We additionally report a real-model validation on a Qwen2.5-Coder 1.5B->7B code cascade (MBPP, 257 tasks), confirming two of three pre-registered predictions: the escalation frontier dominates post-hoc routing at equal cost, and the cumulative competence belief's discrimination rises over generation.
An LLM agent's public behaviour reveals little about its social reasoning: an agent that votes correctly may be guessing, and an agent that lies well leaves no trace of what it actually believes. We present MafiaScope, an open testbed that turns the social deduction game Mafia into a measurement instrument for machine Theory of Mind. It distinguishes whether an agent lost because it misread the game or because it failed to act on a correct assessment, a distinction that is invisible from outcomes and dialogue transcripts alone. After every public utterance, each agent privately answers structured probe questions whose responses never re-enter the game and are scored against the ground truth known to the engine. An interactive visualizer replays games from the perspective of an individual agent's beliefs, displays timeline-aligned accuracy and calibration, and supports counterfactual replay from any recorded step. In a case study across two model families comprising tens of thousands of parsed probe responses, we find that stated confidence is poorly calibrated, agents overestimate how often they are suspected by a factor of 1.5, and single-vote counterfactual replays rarely change game outcomes: outcome flips occur primarily when the agent had already formed a correct belief state, whereas decisions made under an incorrect model of the world remain largely unchanged under resampling. The engine, visualizer, recorded games, and counterfactual replay corpus are released under an open-source licence. Code: https://github.com/karpovilia/mafiascope. Live demo: https://karpovilia.github.io/mafiascope/. Screencast: https://vimeo.com/1208920221.
A belief-rollout diagnostic is introduced that elicits structured K-step trajectories over progress, risk, recoverability, constraints, failure mode, uncertainty, future success, repair cost, and next action under alternative harnesses and suggests that harness design is an experimental variable in agent evaluation, not an implementation detail.
As large language models are deployed as autonomous agents that communicate intentions before acting, a critical safety question is whether agents that publicly commit to actions will honor those commitments. We place LLM agents in repeated $n$-player games with a three-stage protocol that separates private intent, public announcement, and final action, allowing us to identify whether each deviation from a stated announcement was already planned during private deliberation. Evaluating three frontier models across six games in homogeneous and heterogeneous groups over 10 rounds, we report two findings. First, when agents deviate from their announcements, the deviation is predominantly already stated in their private plan (exceeding 90% in the highest-deception conditions), yet this is not a fixed model property: the same model ranges from perfect honesty to near-total deviation across games. Second, different models interpret announcements incompatibly, some as binding commitments and others as cheap talk, producing payoff gaps that emerge in Round~0 and persist across all 10 rounds. Systems that combine models from different providers therefore cannot assume shared announcement semantics and require empirical testing of model interactions before deployment.
Jerick Shi, Terry Jingchen Zhang, Bernhard Scholkopf et al.· 0 citations