Searcharxiv⌕ Search

arXiv subjects

Andrew Perrault

Publications and source records attributed to Andrew Perrault.

At least 19 recordsLinked to original sources

AfterDeleteBench: Auditing Post-Deletion Information Persistence in LLM Agents

Many LLM agents maintain external persistent memory and derive additional state from past interactions, such as summaries. Deleting an original memory record therefore need not remove information carried by its descendants. We introduce AfterDeleteBench, a controlled behavioral audit for this failure mode in fixed-weight LLM agents. For each synthetic secret, we compare an exposed-then-deleted agent with an otherwise matched agent that was never exposed to the secret, allowing post-deletion recovery to be attributed to the original exposure. In our primary setting, a single query recovers the deleted secret from Qwen2.5-7B-Instruct in roughly two thirds of cases, while matched never-exposed controls have zero recovery. Crucially, recovery does not require a surviving verbatim copy: when the complete secret string is absent and only separately stored components remain under a public composition rule, exact reconstruction succeeds in 62-99% of cases across Qwen, Mistral, and Phi-3. We also show that the response interface is part of the measurement system: target-bearing replies can be scored as failures when the requested format and parser are misaligned. Finally, we separate reconstruction from write-back and fresh-session recall. Provenance-aware closure eliminates observed recovery in the tested positive-baseline settings, whereas an exact-literal write filter blocks reinsertion but cannot erase information already present. Together, these results show that source deletion should be evaluated behaviorally at the level of the agent and its memory system, not only as a storage operation.

cs.CR↗

RealityBench: Simulated Environments to Evaluate a Vision Language Model's Ability to Document Human Physical Work

Augmenting worker capabilities in the physical world through guidance has been a decades-long goal for AI agents. One key supporting capability is to understand what the worker is doing, seeing through the human's camera or smart glasses, and assist them in documentation. Despite advances in vision language models, their capabilities in these settings remain uncertain and are challenging to benchmark because most physical work data is proprietary, and involves too much variability to study in a principled manner. Unlike structured settings such as kitchens and warehouses that have predictable lighting, geometry, and objects, field-facing tasks such as rooftop solar installation and vehicle repair are unconstrained and high variance. Here, we model real-world physical work in Blender across eleven environments, incorporating variants in both image acceptability and documentation. The acceptability task asks whether a particular view would be suitable proof-of-work and documentation asks whether an agent can understand the task-relevant state of the environment from that view. Across 26 vision-language models, we find that top models largely solve documentation when the answer is text in frame, but are weak at reading measurements and at judging small parts and spatial relations, such as whether a tape measure leans. Acceptability is harder to pin down because human raters disagree with each other. Measured against a scorer that agrees with the raters' majority about as often as a single rater does, the best models cannot be distinguished from a single non-author rater, at a cost within the range of a human pass. In the three environments where we can trace acceptability errors, most wrong rejections come from misreading the scene rather than the requirement. On the three documentation questions for which we have real images, model rankings on rendered frames transfer to real images.

cs.CV↗

Identity, Incentives, and Cooperative Behavior for AI: Reverse-Engineering Multiagent Cooperation

The behavior of AI systems increasingly depends on the extent to which agents cooperate or coordinate with one another, in ways that may be desirable or not depending on the situation. Conventional explanations for cooperative behavior distinguish between purely other-regarding preferences ("altruism") and instrumental reasons for self-interested agents to cooperate. The way that agents should enact these mechanisms for cooperation (or not) is complicated by the fact that there is not a clear or unambiguous locus of identity or selfhood for AI agents. For example, it is unclear whether agents involved in instances of misaligned collective behavior are motivated by goals at the level of the collective or by various forms of instrumental reasoning. We develop a framework for reverse-engineering why a given set of agents engage in a particular set of cooperation-related behaviors. We argue that many different mechanisms can often produce similar observed behaviors and introduce a set of experimental designs along with conditions under which they can successfully disentangle competing explanations for multiagent behavior. We apply this framework to interactions from a set of frontier models in a simulated environment and demonstrate that the behavior of different models arises from substantively different mechanisms.

cs.AI↗

Generating Synthetic Doctor-Patient Conversations for Long-form Audio Summarization

Long-context audio reasoning is underserved in both training data and evaluation. Existing benchmarks target short-context tasks, and the open-ended generation tasks most relevant to long-context reasoning pose well-known challenges for automatic evaluation. We propose a synthetic data generation pipeline designed to serve both as a training resource and as a controlled evaluation environment, and instantiate it for first-visit doctor-patient conversations with SOAP note generation as the task. The pipeline has three stages, persona-driven dialogue generation, multi-speaker audio synthesis with overlap/pause modeling, room acoustics, and sound events, and LLM-based reference SOAP note production, built entirely on open-weight models. We release 8,800 synthetic conversations with 1.3k hours of corresponding audio and reference notes. Evaluating current open-weight systems, we find that cascaded approaches still substantially outperform end-to-end models.

cs.SD↗

Many Preferences, Few Policies: Compact Portfolios for Multi-Objective LLM Alignment

Aligning large language models (LLMs) requires balancing competing objectives such as helpfulness, harmlessness, and conciseness. The appropriate balance varies across users and applications, yet training, evaluating, and deploying many policies across different reward weights is costly. We study how to identify a small portfolio of LLMs that preserves near-optimal performance across all reward weightings. We propose PALM (Portfolio of Aligned LLMs), an algorithm that combines a structured grid of weight vectors, a lazy search that optimizes policies only where needed, and pruning. Given target approximation tolerances, PALM returns a portfolio that provably contains a near-optimal policy for every weight vector, with an explicit upper bound on portfolio size. Such portfolios can support scalable personalization, reward-weight exploration during model development, and compact decoding-time configurations. Experiments show that PALM generally achieves smaller approximation gaps than same-size portfolios built from uniformly spaced or randomly sampled weights. We further demonstrate that PALM scales effectively to higher-dimensional reward spaces through efficient search and sparse preference structure.

cs.CL↗

Pathwise Individual Rationality in Federated Learning: A Mechanism-Architecture Co-Design

Participation in federated learning (FL) comes at a cost. Clients trade off privacy, communication, and compute costs for potentially greater gains in model efficacy. This paper explores this tradeoff under the aegis of individual rationality (IR) versus autarky, the basic game-theoretic requirement that the federation provide utility no worse than local training. Using the above as the design target, we examine pathwise performance of FL, as a per-round bound on cumulative surplus rather than only an asymptotic equilibrium guarantee under different degrees of client data heterogeneity. Along the learning path, clients can remain below their local-training baseline for hundreds of rounds. The natural remedy is to cap each client's per-round contribution so that this shortfall stays bounded. We prove that such a cap drives all contributions to zero, with high probability, when its tolerance is small relative to the warm-up deficit, and we observe learning collapse under it even at low heterogeneity. We then propose a design that combines a short-term participation guarantee, enforced after an announced grace window, with personalized model evaluation, while preserving the incentive properties of the underlying mechanism. We provide a theoretical basis for this approach and empirically demonstrate that, after the grace window, clients meet IR on nearly all rounds without harming overall performance under low to moderate heterogeneity; under severe heterogeneity, the design shows promising outcomes for clients compared to their local baseline at some cost in accuracy.

cs.LG↗

Recovering Physical Dynamics from Discrete Observations via Intrinsic Differential Consistency

Recovering continuous-time dynamics from discrete observations is difficult because local supervision loses fidelity as the observation interval grows. We replace local supervision with a global structural constraint: any flow representing autonomous dynamics must satisfy the semi-group property under time translation. We train a time-conditioned secant velocity field whose deviation from this property---which we call Symmetry Rupture---serves two roles: as a training regularizer it confines the hypothesis space to flows that compose consistently across temporal scales; as an inference oracle it guides an adaptive solver to select the largest step size that preserves internal consistency, without relying on local truncation error estimates. On the diffusion-reaction and Navier-Stokes benchmarks, our method achieves the lowest rollout RMSE among all evaluated methods using $\leq 3$ function evaluations per step. On the near-conservative shallow water benchmark, our method matches or exceeds competing flow-based methods at far lower computational cost; the strongest Neural ODE baseline achieves lower absolute RMSE but at the expense of $>100$ function evaluations per step. In the direct auto-regressive setting, our solver maintains stable long-horizon rollouts where flow-matching baselines diverge and Neural ODE requires up to $12\times$ more function evaluations. The core contribution is a model that internalizes temporal consistency, removing the need for high-order external solvers to compensate for structural bias.

cs.LG↗

Training Small LLMs as Spatial Multi-Agent Policies

Training LLM-based multi-agent systems with multi-agent reinforcement learning is rapidly gaining traction, and a parallel line of work argues that such systems should be judged by their behavior, not only their reward. We take up both threads in spatial cooperative games, where small frozen LLMs prompted with low-level actions fail outright, earning zero reward. Guided by the options/semi-MDP framework---and, because option execution is asynchronous across agents, its multi-agent extension in macro-action Dec-POMDPs---we equip each game with a library of symbolic \emph{options}: typed, state-feasible, short-horizon behaviors executed by a symbolic planner. Each library is drafted by a frontier coding model from the game's source code; the feasibility guards that filter each menu are then synthesized mechanically from cheap random-policy burn-in rollouts---a guard is adopted only if it explains repeated execution failures while hiding no logged success---so no guard is authored, selected, or reward-tuned by hand. Each agent's LLM acts as its policy over options, with a private per-agent LoRA adapter trained by a per-agent variant of multi-agent GRPO (PA-MAGRPO); this lifts frozen bases from zero reward to competent play across three games and four small backbones. Behavioral audits then reveal that reward and cooperation decouple: a rising reward curve may simply mean that one agent has learned to run the entire task alone while its partner idles---cooperation emerges only when the task makes it necessary. Reward alone is thus an unreliable readout of cooperation; behavioral evaluation must sit alongside it.

cs.MA↗

Optimizing Resource-Constrained Non-Pharmaceutical Interventions for Multi-Cluster Outbreak Control Using Hierarchical Reinforcement Learning

Non-pharmaceutical interventions (NPIs), such as diagnostic testing and quarantine, are crucial for controlling infectious disease outbreaks but are often constrained by limited resources, particularly in early outbreak stages. In real-world public health settings, resources must be allocated across multiple outbreak clusters that emerge asynchronously, vary in size and risk, and compete for a shared resource budget. Here, a cluster corresponds to a group of close contacts generated by a single infected index case. Thus, decisions must be made under uncertainty and heterogeneous demands, while respecting operational constraints. We formulate this problem as a constrained restless multi-armed bandit and propose a hierarchical reinforcement learning framework. A global controller learns a continuous action cost multiplier that adjusts global resource demand, while a generalized local policy estimates the marginal value of allocating resources to individuals within each cluster. We evaluate the proposed framework in a realistic agent-based simulator of SARS-CoV-2 with dynamically arriving clusters. Across a wide range of system scales and testing budgets, our method consistently outperforms RMAB-inspired and heuristic baselines, improving outbreak control effectiveness by 20%-30%. Experiments on up to 40 concurrently active clusters further demonstrate that the hierarchical framework is highly scalable and enables faster decision-making than the RMAB-inspired method.

cs.LG↗

VISTA: Verification In Sequential Turn-based Assessment

Hallucination--defined here as generating statements unsupported or contradicted by available evidence or conversational context--remains a major obstacle to deploying conversational AI systems in settings that demand factual reliability. Existing metrics either evaluate isolated responses or treat unverifiable content as errors, limiting their use for multi-turn dialogue. We introduce VISTA (Verification In Sequential Turn-based Assessment), a framework for evaluating conversational factuality through claim-level verification and sequential consistency tracking. VISTA decomposes each assistant turn into atomic factual claims, verifies them against trusted sources and dialogue history, and categorizes unverifiable statements (subjective, contradicted, lacking evidence, or abstaining). Across eight large language models and four dialogue factuality benchmarks (AIS, BEGIN, FAITHDIAL, and FADE), VISTA substantially improves hallucination detection over FACTSCORE and LLM-as-Judge baselines. Human evaluation confirms that VISTA's decomposition improves annotator agreement and reveals inconsistencies in existing benchmarks. By modeling factuality as a dynamic property of conversation, VISTA offers a more transparent, human-aligned measure of truthfulness in dialogue systems.

cs.CL↗

Beyond Length: Context-Aware Expansion and Independence as Developmentally Sensitive Evaluation in Child Utterances

Evaluating the quality of children's utterances in adult-child dialogue remains challenging due to insufficient context-sensitive metrics. Common proxies such as Mean Length of Utterance (MLU), lexical diversity (vocd-D), and readability indices (Flesch-Kincaid Grade Level, Gunning Fog Index) are dominated by length and ignore conversational context, missing aspects of response quality such as reasoning depth, topic maintenance, and discourse planning. We introduce an LLM-as-a-judge framework that first classifies the Previous Adult Utterance Type and then scores the child's response along two axes: Expansion (contextual elaboration and inferential depth) and Independence (the child's contribution to advancing the discourse). These axes reflect fundamental dimensions in child language development, where Expansion captures elaboration, clause combining, and causal and contrastive connectives. Independence captures initiative, topic control, decreasing reliance on adult scaffolding through growing self-regulation, and audience design. We establish developmental validity by showing age-related patterns and demonstrate predictive value by improving age estimation over common baselines. We further confirm semantic sensitivity by detecting differences tied to discourse relations. Our metrics align with human judgments, enabling large-scale evaluation. This shifts child utterance assessment from simply measuring length to evaluating how meaningfully the child's speech contributes to and advances the conversation within its context.

cs.CL↗

Learning Provably Correct Distributed Protocols Without Human Knowledge

Provably correct distributed protocols, which are a critical component of modern distributed systems, are highly challenging to design and have often required decades of human effort. These protocols allow multiple agents to coordinate to come to a common agreement in an environment with uncertainty and failures. We formulate protocol design as a search problem over strategies in a game with imperfect information, and the desired correctness conditions are specified in Satisfiability Modulo Theories (SMT). However, standard methods for solving multi-agent games fail to learn correct protocols in this setting, even when the number of agents is small. We propose a learning framework, GGMS, which integrates a specialized variant of Monte Carlo Tree Search with a transformer-based action encoder, a global depth-first search to break out of local minima, and repeated feedback from a model checker. Protocols output by GGMS are verified correct via exhaustive model checking for all executions within the bounded setting. We further prove that, under mild assumptions, the search process is complete: if a correct protocol exists, GGMS will eventually find it. In experiments, we show that GGMS can learn correct protocols for larger settings than existing methods.

cs.AI↗

Dreaming Falcon: Physics-Informed Model-Based Reinforcement Learning for Quadcopters

Current control algorithms for aerial robots struggle with robustness in dynamic environments and adverse conditions. Model-based reinforcement learning (RL) has shown strong potential in handling these challenges while remaining sample-efficient. Additionally, Dreamer has demonstrated that online model-based RL can be achieved using a recurrent world model trained on replay buffer data. However, applying Dreamer to aerial systems has been quite challenging due to its sample inefficiency and poor generalization of dynamics models. Our work explores a physics-informed approach to world model learning and improves policy performance. The world model treats the quadcopter as a free-body system and predicts the net forces and moments acting on it, which are then passed through a 6-DOF Runge-Kutta integrator (RK4) to predict future state rollouts. In this paper, we compare this physics-informed method to a standard RNN-based world model. Although both models perform well on the training data, we observed that they fail to generalize to new trajectories, leading to rapid divergence in state rollouts, preventing policy convergence.

cs.RO↗

Optimizing Urban Service Allocation with Time-Constrained Restless Bandits

Municipal inspections are an important part of maintaining the quality of goods and services. In this paper, we approach the problem of intelligently scheduling service inspections to maximize their impact, using the case of food establishment inspections in Chicago as a case study. The Chicago Department of Public Health (CDPH) inspects thousands of establishments each year, with a substantial fail rate (over 3,000 failed inspection reports in 2023). To balance the objectives of ensuring adherence to guidelines, minimizing disruption to establishments, and minimizing inspection costs, CDPH assigns each establishment an inspection window every year and guarantees that they will be inspected exactly once during that window. Meanwhile, CDPH also promises surprise public health inspections for unexpected food safety emergencies or complaints. These constraints create a challenge for a restless multi-armed bandit (RMAB) approach, for which there are no existing methods. We develop an extension to Whittle index-based systems for RMABs that can guarantee action window constraints and frequencies, and furthermore can be leveraged to optimize action window assignments themselves. Briefly, we combine MDP reformulation and integer programming-based lookahead to maximize the impact of inspections subject to constraints. A neural network-based supervised learning model is developed to model state transitions of real Chicago establishments using public CDPH inspection records, which demonstrates 10% AUC improvements compared with directly predicting establishments' failures. Our experiments not only show up to 24% (in simulation) or 33% (on real data) objective improvements resulting from our approach and robustness to surprise inspections, but also give insight into the impact of scheduling constraints.

cs.LG↗

Do Audio LLMs Really LISTEN, or Just Transcribe? Measuring Lexical vs. Acoustic Emotion Cues Reliance

Understanding emotion from speech requires sensitivity to both lexical and acoustic cues. However, it remains unclear whether large audio language models (LALMs) genuinely process acoustic information or rely primarily on lexical content. We present LISTEN (Lexical vs. Acoustic Speech Test for Emotion in Narratives), a controlled benchmark designed to disentangle lexical reliance from acoustic sensitivity in emotion understanding. Across evaluations of six state-of-the-art LALMs, we observe a consistent lexical dominance. Models predict "neutral" when lexical cues are neutral or absent, show limited gains under cue alignment, and fail to classify distinct emotions under cue conflict. In paralinguistic settings, performance approaches chance. These results indicate that current LALMs largely "transcribe" rather than "listen," relying heavily on lexical semantics while underutilizing acoustic cues. LISTEN offers a principled framework for assessing emotion understanding in multimodal models.

cs.CL↗

Fine-Tuning Text-to-Speech Diffusion Models Using Reinforcement Learning with Human Feedback

Diffusion models produce high-fidelity speech but are inefficient for real-time use due to long denoising steps and challenges in modeling intonation and rhythm. To improve this, we propose Diffusion Loss-Guided Policy Optimization (DLPO), an RLHF framework for TTS diffusion models. DLPO integrates the original training loss into the reward function, preserving generative capabilities while reducing inefficiencies. Using naturalness scores as feedback, DLPO aligns reward optimization with the diffusion model's structure, improving speech quality. We evaluate DLPO on WaveGrad 2, a non-autoregressive diffusion-based TTS model. Results show significant improvements in objective metrics (UTMOS 3.65, NISQA 4.02) and subjective evaluations, with DLPO audio preferred 67\% of the time. These findings demonstrate DLPO's potential for efficient, high-quality diffusion TTS in real-time, resource-limited settings.

cs.SD↗

Normality-Guided Distributional Reinforcement Learning for Continuous Control

Learning a predictive model of the mean return, or value function, plays a critical role in many reinforcement learning algorithms. Distributional reinforcement learning (DRL) has been shown to improve performance by modeling the value distribution, not just the mean. We study the value distribution in several continuous control tasks and find that the learned value distribution is empirically quite close to normal. We design a method that exploits this property, employing variances predicted from a variance network, along with returns, to analytically compute target quantile bars representing a normal for our distributional value function. In addition, we propose a policy update strategy based on the correctness as measured by structural characteristics of the value distribution not present in the standard value function. The approach we outline is compatible with many DRL structures. We use two representative on-policy algorithms, PPO and TRPO, as testbeds. Our method yields statistically significant improvements in 10 out of 16 continuous task settings, while utilizing a reduced number of weights and achieving faster training time compared to an ensemble-based method for quantifying value distribution uncertainty.

cs.LG↗

Symmetric Reinforcement Learning Loss for Robust Learning on Diverse Tasks and Model Scales

Reinforcement learning (RL) training is inherently unstable due to factors such as moving targets and high gradient variance. Reinforcement Learning from Human Feedback (RLHF) and Reinforcement Learning from AI Feedback (RLAIF) can introduce additional difficulty. Differing preferences can complicate the alignment process, and prediction errors in a trained reward model can become more severe as the LLM generates unseen outputs. To enhance training robustness, RL has adopted techniques from supervised learning, such as ensembles and layer normalization. In this work, we improve the stability of RL training by adapting the reverse cross entropy (RCE) from supervised learning for noisy data to define a symmetric RL loss. We demonstrate performance improvements across various tasks and scales. We conduct experiments in discrete action tasks (Atari games) and continuous action space tasks (MuJoCo benchmark and Box2D) using Symmetric A2C (SA2C) and Symmetric PPO (SPPO), with and without added noise with especially notable performance in SPPO across different hyperparameters. Furthermore, we validate the benefits of the symmetric RL loss when using SPPO for large language models through improved performance in RLHF tasks, such as IMDB positive sentiment sentiment and TL;DR summarization tasks.

cs.LG↗