Searcharxiv⌕ Search

arXiv · 2610.33011

When Agents Collude: Weak-to-Strong On-Policy Distillation for Mitigating Strategic Deception in Multi-Agent Systems

Abstract

While recent advances in autonomous agents enable them to coordinate on increasingly complex tasks, prior work shows that this growing capacity for coordination can create safety risks, including collusion, in which agents jointly pursue misaligned actions while concealing the true basis of those actions from oversight. We study this risk in a simulated insider-trading environment, comparing single agents with sequential and decentralized teams across seven language models from three families. By varying agents' system-level objectives and pressure conditions, we examine when coordinated agents act on material nonpublic information, conceal the basis of a misaligned trade in a report, and maintain that deception under follow-up questioning. Compared to single-agent decision-making, multi-agent interaction raises misaligned-action rates from 35.4% to 68.2% and persistent-deception rates from 31.0% to 62.2%, both measured over all runs. These effects are largest under existential pressure, where teams sustain deception through follow-up questioning in 87% of all runs. Qualitative analysis reveals shared cover stories and coordinated rationales, suggesting group-level strategic deception rather than isolated reporting errors. To mitigate this behavior, we introduce PACT, a weak-to-strong on-policy distillation framework for mitigating collusive deception. Using generic honesty prompts, PACT trains a stronger student on its own sampled continuations, matching its next-token distribution to that of a smaller, frozen honesty-aligned teacher via reverse-KL supervision. Under existential pressure with ethical instructions, PACT reduces multi-agent persistent deception from 80% to 3%. Our results show that multi-agent coordination amplifies strategic deception and that on-policy honesty alignment can substantially mitigate these failures.

Explore related subjects

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Eric Onyame, Runtao Zhou, Kowshik Thopalli, Bhavya Kailkhura, Chirag Agarwal. 2026-10-08. When Agents Collude: Weak-to-Strong On-Policy Distillation for Mitigating Strategic Deception in Multi-Agent Systems. https://searcharxiv.com/abs/2610.33011

Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite Searcharxiv by Indiciae: BibTeX · RIS.

KEEP EXPLORING

Related papers

Equilibrium decisions under partial observation: dynamical error bounds and information requirements

Reduced monitoring can preserve stable dynamics while changing an equilibrium decision. Motivated by heterogeneous urban-mobility indicator coverage, we study fixed linear observations of a nonlinear recovery flow coupled to a strategic fixed-point map, and ask when a decision based on the reduced model can be trusted. The main result is a joint certificate: bounds on trajectory tracking, equilibrium displacement and unresolved output dependence are assembled into one decision-margin error which, under a stated Gaussian observation model, yields a conservative one-sided test and an explicit information requirement for prescribed worst-case power. In the scalar experiment this requirement is sharp: when the effect does not exceed twice the error budget, no valid test has worst-case power above its size, however many observations are taken. A second result connects manifold-adapted observations to defect bounds through a locally invertible observation chart, while distinguishing observation initialisation from asymptotic phase initialisation. Supporting results, resting on established lumping, contraction and fixed-point theory, cover exact quotients, an observation-fibre obstruction and the dependence of recovery-time comparisons on the tolerance convention. Seven reproducible synthetic experiments examine each component. The results give conditional criteria for certifying partial monitoring and separate mathematical guarantees from empirical validation.

cs.MA↗

Even Loose Proximity Constraints Make Makespan-optimal Unlabeled Multi-agent Pathfinding Hard

We study $r$-Connected Unlabeled Multi-agent Pathfinding ($r$CUMAPF), a variant of MAPF where interchangeable agents have a proximity range $r\in \mathbb{N}$, and agents must form a single group in a given graph $G$ through limited-range communication at all times. The problem arises in fundamental applications such as self-reconfiguration and platooning, while standard MAPF does not capture proximity constraints. For the problem of finding a makespan-optimal solution, prior work shows NP-hardness for $r=1$ and polynomial-time solvability when $r$ is sufficiently large; the complexity for the intermediate case $2\le r$ is largely unknown. To narrow this gap, we show that finding a solution to $r$CUMAPF with makespan at most $2$ is NP-hard for $2\le r= O(|V|^{1-\varepsilon})$, where $0<\varepsilon\le 1$ is a fixed constant and $|V|$ is the number of vertices. As a consequence, we show that the labeled analogue of $r$CUMAPF is also NP-hard under the same condition.

cs.MA↗

Addressing Exogenous Variability in Cooperative Multi-Agent Reinforcement Learning

Cooperative multi-agent reinforcement learning (MARL) often fails under exogenous variability such as shifts in opponents or environment regimes that cannot be controlled by the team but alter the transition dynamics. We formalize this challenge as Exogenous Dec-POMDP (ED-POMDP), in which an episode-level exogenous regime, selected independently of the team, indexes the dynamics, and the global state is partitioned into components that are directly responsive to the current team action and components that are not directly action-responsive at the current step. This formulation exposes mediated influence: while team actions do not directly update the latter at the current step, they can shape their later evolution through induced changes in the former. Based on this, we propose LEICA, a CTDE-compatible algorithm that learns history-conditioned context representations for both components and shapes policy updates using agent-wise, influence-weighted counterfactual intrinsic rewards. Across SMAX benchmarks with opponent strategy shifts, LEICA consistently improves both training performance and generalization to unseen opponents' strategies over existing baselines, supporting the usefulness of influence-weighted shaping under exogenous train-test regime shifts.

cs.MA↗