The role of natural language constraints in Safe Multi-agent Reinforcement Learning (MARL) is crucial, yet often overlooked. While Safe MARL has vast potential, especially in fields like robotics and autonomous vehicles, its full potential is limited by the need to define constraints in pre-designed mathematical terms, which requires extensive domain expertise and reinforcement learning knowledge, hindering its broader adoption. To address this limitation and make Safe MARL more accessible and adaptable, we propose a novel approach named Safe Multi-agent Reinforcement Learning with Natural Language constraints (SMALL). Our method leverages fine-tuned language models to interpret and process free-form textual constraints, converting them into semantic embeddings that capture the essence of prohibited states and behaviours. These embeddings are then integrated into the multi-agent policy learning process, enabling agents to learn policies that minimize constraint violations while optimizing rewards. To evaluate the effectiveness of SMALL, we introduce the LaMaSafe, a multi-task benchmark designed to assess the performance of multiple agents in adhering to natural language constraints. Empirical evaluations across various environments demonstrate that SMALL achieves comparable rewards and significantly fewer constraint violations, highlighting its effectiveness in understanding and enforcing natural language constraints.

How can we distinguish whether a peer review was written by a human or generated by an AI model? We argue that, in this setting, authorship should not be attributed solely from the textual features of a review, but also from the ideas, judgments, and claims it expresses. To this end, we propose Sem-Detect, an authorship detection method for peer reviews that operationalizes this principle by combining textual features with claim-level semantic analysis. Sem-Detect compares a target review against multiple AI-generated reviews of the same paper, leveraging the observation that different AI models tend to converge on similar points, while human reviewers introduce more unique and diverse ones. As a result, Sem-Detect is able to distinguish fully AI reviews from authentic human-written ones, including those that have been refined using an LLM but still reflect human judgment. Across a dataset of over 20,000 peer reviews from ICLR and NeurIPS conferences, Sem-Detect improves over the strongest baseline by 25.5

Food rescue organizations simultaneously tackle food insecurity and waste by working with volunteers to redistribute food from donors who have excess to recipients who need it. Volunteer feedback allows food rescue organizations to identify issues early and ensure volunteer satisfaction. However, food rescue organizations monitor feedback manually, which can be cumbersome and labor-intensive, making it difficult to prioritize which issues are most important. In this work, we investigate how large language models (LLMs) assist food rescue organizers in understanding and taking action based on volunteer experiences. We work with 412 Food Rescue, a large food rescue organization based in Pittsburgh, Pennsylvania, to design RescueLens, an LLM-powered tool that automatically categorizes volunteer feedback, suggests donors and recipients to follow up with, and updates volunteer directions based on feedback. We evaluate the performance of RescueLens on an annotated dataset, and show that it can recover 96% of volunteer issues at 71% precision. Moreover, by ranking donors and recipients according to their rates of volunteer issues, RescueLens allows organizers to focus on 0.5% of donors responsible for more than 30% of volunteer issues. RescueLens is now deployed at 412 Food Rescue and through semi-structured interviews with organizers, we find that RescueLens streamlines the feedback process so organizers better allocate their time.

Humans increasingly make decisions with AI assistance; for example, doctors may follow AI-recommended diagnostic tests and base their diagnoses on the results. A natural question is which tests should AI recommend to balance short-term decision quality and long-term human learning when different features (e.g., test results) are correlated. While prior work establishes that stationary policies that recommend the same tests repeatedly are optimal when features are independent, we prove that feature correlations lead such policies to perform arbitrarily poorly. Instead, we prove that any optimal policy must follow an explore-then-commit structure; initially, the AI should offer diverse tests so humans can learn accurate feature coefficients, then the AI should commit to a single set of tests, with exploration length that depends on the degree of feature correlation. We prove that computing the optimal policy is NP-hard and derive a dynamic programming-based algorithm that finds the optimal policy for finite horizons. We additionally develop an approximation that plans for shorter horizons and appends a stationary suffix, achieving near-optimal performance. Our empirical results complement our theory by showing that stronger feature correlation leads to longer exploration phases.

We present the PokeAgent Challenge, a large-scale benchmark for decision-making research built on Pokemon's multi-agent battle system and expansive role-playing game (RPG) environment. Partial observability, game-theoretic reasoning, and long-horizon planning remain open problems for frontier AI, yet few benchmarks stress all three simultaneously under realistic conditions. PokeAgent targets these limitations at scale through two complementary tracks: our Battling Track, which calls for strategic reasoning and generalization under partial observability in competitive Pokemon battles, and our Speedrunning Track, which requires long-horizon planning and sequential decision-making in the Pokemon RPG. Our Battling Track supplies a dataset of 20M+ battle trajectories alongside a suite of heuristic, RL, and LLM-based baselines capable of high-level competitive play. Our Speedrunning Track provides the first standardized evaluation framework for RPG speedrunning, including an open-source multi-agent orchestration system for modular, reproducible comparisons of harness-based LLM approaches. Our NeurIPS 2025 competition validates both the quality of our resources and the research community's interest in Pokemon, with over 100 teams competing across both tracks and winning solutions detailed in our paper. Participant submissions and our baselines reveal considerable gaps between generalist (LLM), specialist (RL), and elite human performance. Analysis against the BenchPress evaluation matrix shows that Pokemon battling is nearly orthogonal to standard LLM benchmarks, measuring capabilities not captured by existing suites and positioning Pokemon as an unsolved benchmark that can drive RL and LLM research forward. We transition to a living benchmark with a live leaderboard for Battling and self-contained evaluation for Speedrunning at https://pokeagentchallenge.com.

Training interpretable concept-based policies requires practitioners to manually select which human-understandable concepts an agent should reason with when making sequential decisions. This selection demands domain expertise, is time-consuming and costly, scales poorly with the number of candidates, and provides no performance guarantees. To overcome this limitation, we propose the first algorithms for principled automatic concept selection in sequential decision-making. Our key insight is that concept selection can be viewed through the lens of state abstraction: intuitively, a concept is decision-relevant if removing it would cause the agent to confuse states that require different actions. As a result, agents should rely on decision-relevant concepts; states with the same concept representation should share the same optimal action, which preserves the optimal decision structure of the original state space. This perspective leads to the Decision-Relevant Selection (DRS) algorithm, which selects a subset of concepts from a candidate set, along with performance bounds relating the selected concepts to the performance of the resulting policy. Empirically, DRS automatically recovers manually curated concept sets while matching or exceeding their performance, and improves the effectiveness of test-time concept interventions across reinforcement learning benchmarks and real-world healthcare environments.

Reinforcement learning with verifiable rewards (RLVR) has emerged as the leading approach for enhancing reasoning capabilities in large language models. However, it faces a fundamental compute and memory asymmetry: rollout generation is embarrassingly parallel and memory-light, whereas policy updates are communication-heavy and memory-intensive. To address this, we introduce PODS (Policy Optimization with Down-Sampling), which decouples rollout generation from policy updates by training only on a strategically selected subset of rollouts, maintaining learning quality while dramatically reducing update costs. We propose a principled subset selection criterion, max-variance down-sampling, that maximizes reward diversity, and provide an efficient O(nlog n) implementation. Empirically, Group Relative Policy Optimization (GRPO) with PODS achieves the peak test accuracy of vanilla GRPO at least 1.7× faster across the different reasoning benchmarks and hardware configurations we tested.

Winning competitive debates requires sophisticated reasoning and argument skills. There are unique challenges in the competitive debate: (1) The time constraints force debaters to make strategic choices about which points to pursue rather than covering all possible arguments; (2) The persuasiveness of the debate relies on the back-and-forth interaction between arguments, which a single final game status cannot evaluate. To address these challenges, we propose TreeDebater, a novel debate framework that excels in competitive debate. We introduce two tree structures: the Rehearsal Tree and Debate Flow Tree. The Rehearsal Tree anticipates the attack and defenses to evaluate the strength of the claim, while the Debate Flow Tree tracks the debate status to identify the active actions. TreeDebater allocates its time budget among candidate actions and uses the speech time controller and feedback from the simulated audience to revise its statement. The human evaluation on both the stage-level and the debate-level comparison shows that our TreeDebater outperforms the state-of-the-art multi-agent debate system, with a +15.6% improvement in stage-level persuasiveness with DeepSeek and +10% debate-level opinion shift win. Further investigation shows that TreeDebater shows better strategies in limiting time to important debate actions, aligning with the strategies of human debate experts.

Exploration remains a key bottleneck for reinforcement learning (RL) post-training of large language models (LLMs), where sparse feedback and large action spaces can lead to premature collapse into repetitive behaviors. We propose Verbalized Action Masking (VAM), which verbalizes an action mask in the prompt and enforces that the model outputs an action from the masked set. Building on this interface, we introduce iterative action-space pruning: if the target action is not sampled, we remove valid sampled actions from the mask and resample under the reduced candidate set, repeating until the target is sampled or a fixed budget is exhausted. We study VAM in chess and evaluate it under two training regimes: an engine-play regime that generates states via play against an engine opponent and a fixed-dataset regime that trains from a fixed dataset of positions with verifier scores. Across held-out chess puzzles and full-game play measured by average centipawn loss (ACPL), VAM improves learning efficiency and final performance over strong baselines, highlighting verbalized masking as a practical mechanism for controllable exploration in LLM RL post-training.

Training the public to recognize phishing emails is challenging because attackers increasingly use Large Language Models (LLMs) to craft more convincing scams. Although previous research suggests that fully LLM-generated phishing emails are easier for humans to identify, recent models like GPT-4 used in tandem with prompt engineering methods could generate phishing emails that are more difficult to detect. Various training techniques have been used to educate end users about the risks of phishing emails, but these methods may not address the challenges posed by LLM-generated phishing emails. To investigate these challenges, we conducted an experiment using multiple different methods that cyber attackers can use to develop phishing emails. We found that the most challenging emails to detect were those written by humans and then edited and stylized using code generated by an LLM. We also created an anti-phishing training platform to study a personalized learning approach using cognitive models. The basic idea, borrowed from tutoring systems, is to use a cognitive model based on Instance-Based Learning (IBL) Theory to trace the student’s actions and predict each email’s classification decision. The training protocol then selects educational examples that maximize the student’s classification improvement. We also used the predictions of the IBL model to prompt the LLM to generate written feedback for the students. Our study shows that our training methods significantly improved the student’s classification decisions. These results are highly relevant to researchers interested in the social impacts of LLM, their misuse, and mitigation strategies and to practitioners interested in novel anti-phishing training techniques.

Behavioral health conditions, which include mental health and substance use disorders, are the leading disease burden in the United States. Peer-run behavioral health organizations (PROs) critically assist individuals facing these conditions by combining mental health services with assistance for needs such as income, employment, and housing. However, limited funds and staffing make it difficult for PROs to address all service user needs. To assist peer providers at PROs with their day-to-day tasks, we introduce PeerCoPilot, a large language model (LLM)-powered assistant that helps peer providers create wellness plans, construct step-by-step goals, and locate organizational resources to support these goals. PeerCoPilot ensures information reliability through a retrieval-augmented generation pipeline backed by a large database of over 1,300 vetted resources. We conducted human evaluations with 15 peer providers and 6 service users and found that over 90% of users supported using PeerCoPilot. Moreover, we demonstrate that PeerCoPilot provides more reliable and specific information than a baseline LLM. PeerCoPilot is now used by a group of 5-10 peer providers at CSPNJ, a large behavioral health organization serving over 10,000 service users, and we are actively expanding PeerCoPilot's use.

Benchmarks are necessary for healthcare evaluation, but are not sufficient for predicting deployment performance. Our position is that the evaluation–deployment gap arises not because of poorly designed benchmarks, but from implicit assumptions about how users interact with models that cannot be surfaced from benchmarks alone. To make this precise, we propose a classification of assumptions into two categories: task, which can be tested from conversation data alone, and outcome, which requires outcome data and behavioral studies for testing. Critically, outcome assumptions depend on human behavior, something that even well-designed benchmarks cannot directly observe. To demonstrate the operationality of this framework, we retrospectively analyze a healthcare RCT as a case study and find that the gap naturally separates into task and outcome gaps of roughly equal size. To address this, we make two contributions: first, we propose BenchmarkCards, an artifact that documents assumptions, and second, we propose staged evaluation, a procedure that systematically tests assumptions and evaluates performance.

As AI-generated text enters the real-world at scale, institutions increasingly use commercial AI-text detectors, especially in education and academic-integrity workflows. We report a surprising empirical finding about such systems: when evaluated by GPTZero and Pangram, generated text from base models is often judged overwhelmingly human, whereas text generated by their instruction-tuned counterparts is not. Building on this observation, we propose Humanization by Iterative Paraphrasing (HIP), a detector-agnostic pipeline that minimally fine-tunes a base model into a paraphraser and applies it iteratively. Compared with the baselines we test, HIP yields a stronger trade-off between semantic preservation and detector evasion on commercial detectors. Across Llama-3 and Qwen-3 families, spanning model sizes from 0.6B to 70B, HIP consistently improves detector human-likeness. Our findings suggest that current detectors are tracking artifacts of instruction tuning and local context more than any invariant notion of machine-generated text. This, in turn, calls for detector designs that model these factors more explicitly.

Assigning papers to reviewers is a central challenge in the peer-review process of large academic conferences. Program chairs must balance competing objectives, including maximizing reviewer expertise, promoting diversity, and enhancing robustness to strategic manipulation, but it is challenging to do so at the modern conference scale. Existing algorithmic paper assignment approaches either fail to address all of these goals simultaneously or suffer from poor scalability. To address the limitation, we propose Robust Assignment via Marginal Perturbation (RAMP), a unified framework for large-scale peer review. Our approach formulates a linearized perturbed-maximization objective with soft constraints that flexibly balance assignment quality, diversity, and robustness while maintaining runtime efficiency. We further introduce an attribute-aware sampling procedure that converts fractional solutions into integral assignments and improves the diversity and robustness of the final assignment. On datasets with over 20,000 papers and 20,000 reviewers, RAMP runs in under 20 minutes, demonstrating its suitability for real-world deployment.

Watermarking techniques for large language models (LLMs), which encode hidden information in the output so its source can be verified, have gained significant attention in recent days, thanks to their potential capability to detect accidental or deliberate misuse. Similar challenges involving model misuse also exist in the context of game-playing, such as when detecting the unauthorized use of AI tools in gaming platforms (e.g., cheating in online chess). In this paper, we initiate the study of how game-playing strategies can be watermarked. We show how the KGW watermark for LLMs can be adapted to watermark game-playing agents in perfect-information extensive-form games. The watermark can then be detected using a statistical test. We show that the degradation in the quality of the watermarked strategy profile, quantified by the expected utility, can be bounded, but there is a tradeoff between detectability and quality. In our experiments, we bootstrap the watermarking framework to various chess engines and demonstrate that a) the impact of the watermark on the quality of the strategy is negligible and b) the watermark can be detected with just a handful of games.

Model distillation enables efficient emulation of frontier large language models (LLMs), creating a need for robust mechanisms to detect when a third-party student model has trained on a teacher model's outputs. However, existing fingerprinting techniques that could be used to detect such distillation rely on heuristic perturbations that impose a steep trade-off between generation quality and fingerprinting strength, often requiring significant degradation of utility to ensure the fingerprint is effectively internalized by the student. We introduce (ADFP), a principled approach that aligns the fingerprinting objective with the student's learning dynamics. Building upon the gradient-based framework of , ADFP utilizes a proxy model to identify and sample tokens that directly maximize the expected detectability of the fingerprint in the student after fine-tuning, rather than relying on the incidental absorption of the un-targeted biases of a more naive watermark. Experiments on GSM8K and OASST1 benchmarks demonstrate that ADFP achieves a significant Pareto improvement over state-of-the-art baselines, yielding stronger detection confidence with minimal impact on utility, even when the student model's architecture is unknown.

Despite the success of game-theoretic models in security resource allocations against adversaries, existing works have fallen short in addressing the critical challenge of team defense with composable targets. Composable targets, commonly seen in cybersecurity practices like vulnerability analysis, consist of heterogeneous tasks that can be processed by different defenders. The intrinsic heterogeneity and potential precedence constraints among tasks present a great challenge to devising optimal defender strategies. In this paper, we propose a general-sum Stackelberg game model for team defense with composable targets. We develop SWING, a novel method that efficiently calculates optimal defense strategies by combining binary search, linear programming, and column generation. We prove that our algorithm calculates strong Stackelberg equilibrium (SSE), and that in practice, it is runtime-efficient at finding optimal strategies. To further enhance the applicability of SWING, we extend its capabilities to encompass defense tasks with precedence constraints. This is achieved by leveraging flexible job shop problem (FJSP) literature to devise a branch-and-bound-based method. Our empirical evaluations illustrate that this extension enhances runtime efficiency and substantially improves solution quality compared to baseline methods.

The burgeoning field of artificial intelligence for social impact (AI4SI) represents a significant evolution in artificial intelligence, prioritizing measurable positive impact for vulnerable and under-resourced populations. This article examines the historical context and recent surge in AI4SI, driven by technological advancements and a growing awareness of societal challenges. It highlights the crucial role of interdisciplinary collaboration, ethical considerations, and the potential of emerging AI trends in addressing issues such as poverty, health, and environmental sustainability. Furthermore, the article delves into key research questions and challenges facing the field, including the need for contextually relevant AI design, overcoming data limitations, ensuring scalable and sustainable deployments in resource-constrained environments, and establishing robust evaluation frameworks. Realizing the full potential of AI to address pressing societal needs in the coming decade and beyond will hinge on effectively navigating these challenges and fostering a deeply impact-driven approach to research and development.

Recent advances in reinforcement learning (RL) have predominantly leveraged neural network policies for decision-making, yet these models often lack interpretability, posing challenges for stakeholder comprehension and trust. Concept bottleneck models offer an interpretable alternative by integrating human-understandable concepts into policies. However, prior work assumes that concept annotations are readily available during training. For RL, this requirement poses a significant limitation: it necessitates continuous real-time concept annotation, which either places an impractical burden on human annotators or incurs substantial costs in API queries and inference time when employing automated labeling methods. To overcome this limitation, we introduce a novel training scheme that enables RL agents to efficiently learn a concept-based policy by only querying annotators to label a small set of data. Our algorithm, LICORICE, involves three main contributions: interleaving concept learning and RL training, using an ensemble to actively select informative data points for labeling, and decorrelating the concept data with a simple strategy. We show how LICORICE reduces human labeling efforts to 500 or fewer concept labels in three environments, and 5000 or fewer in two complex environments, all at no cost to performance. We also explore the use of VLMs as automated concept annotators, finding them effective in some cases but challenging in others. This work significantly reduces the annotation burden for interpretable RL, making it more practical for real-world applications that necessitate transparency. Our code is released.