While AI agents have shown remarkable performance at various tasks, they still struggle with complex multi-modal applications, structured generation and strategic planning. Improvements via standard fine-tuning is often impractical, as solving agentic tasks usually relies on black box API access without control over model parameters. Inference-time methods such as Best-of-N (BON) sampling offer a simple yet effective alternative to improve performance. However, BON lacks iterative feedback integration mechanism. Hence, we propose Iterative Agent Decoding (IAD) which combines iterative refinement with dynamic candidate evaluation and selection guided by a verifier. IAD differs in how feedback is designed and integrated, specifically optimized to extract maximal signal from reward scores. We conduct a detailed comparison of baselines across key metrics on Sketch2Code, Text2SQL, and Webshop where IAD consistently outperforms baselines, achieving 3--6% absolute gains on Sketch2Code and Text2SQL (with and without LLM judges) and 8--10% gains on Webshop across multiple metrics. To better understand the source of IAD's gains, we perform controlled experiments to disentangle the effect of adaptive feedback from stochastic sampling, and find that IAD's improvements are primarily driven by verifier-guided refinement, not merely sampling diversity. We also show that both IAD and BON exhibit inference-time scaling with increased compute when guided by an optimal verifier. Our analysis highlights the critical role of verifier quality in effective inference-time optimization and examines the impact of noisy and sparse rewards on scaling behavior. Together, these findings offer key insights into the trade-offs and principles of effective inference-time optimization.
Over the past two decades, researchers have made significant advancements in simulating human crowds, yet these efforts largely focus on low-level tasks like collision avoidance and a narrow range of behaviors such as path following and flocking. However, creating compelling crowd scenes demands more than just functional movement-it requires capturing high-level interactions between agents, their environment, and each other over time. To address this issue, we introduce Gen-C, a generative model to automate the task of authoring high-level crowd behaviors. Gen-C bypasses the labor-intensive and challenging task of collecting and annotating real crowd video data by leveraging a large language model (LLM) to generate a limited set of crowd scenarios, which are subsequently expanded and generalized through simulations to construct time-expanded graphs that model the actions and interactions of virtual agents. Our method employs two Variational Graph Auto-Encoders guided by a condition prior network: one dedicated to learning a latent space for graph structures (agent interactions) and the other for node features (agent actions and navigation). This setup enables the flexible generation of dynamic crowd interactions. The trained model can be conditioned on natural language, empowering users to synthesize novel crowd behaviors from text descriptions. We demonstrate the effectiveness of our approach in two scenarios, a University Campus and a Train Station, showcasing its potential for populating diverse virtual environments with agents exhibiting varied and dynamic behaviors that reflect complex interactions and high-level decision-making patterns.
Large Language Models (LLMs) are playing an expanding role in physics research by enhancing reasoning, symbolic manipulation, and numerical computation. However, ensuring the reliability and interpretability of their outputs remains a significant challenge. In our framework, we conceptualize the collaboration between AI and human scientists as a dynamic interplay among three modules: the reasoning module, the interpretation module, and the AI-scientist interaction module. Recognizing that effective physics reasoning demands rigorous logical consistency, quantitative precision, and deep integration with established theoretical models, we introduce the interpretation module to improve the understanding of AI-generated outputs, which is not previously explored in the literature. This module comprises multiple specialized agents, including summarizers, model builders, UI builders, and testers, which collaboratively structure LLM outputs within a physically grounded framework, by constructing a more interpretable science model. A case study demonstrates that our approach enhances transparency, facilitates validation, and strengthens AI-augmented reasoning in scientific discovery.
We present the first mechanistic evidence that model-free reinforcement learning agents can learn to plan. This is achieved by applying a methodology based on concept-based interpretability to a model-free agent in Sokoban -- a commonly used benchmark for studying planning. Specifically, we demonstrate that DRC, a generic model-free agent introduced by Guez et al. (2019), uses learned concept representations to internally formulate plans that both predict the long-term effects of actions on the environment and influence action selection. Our methodology involves: (1) probing for planning-relevant concepts, (2) investigating plan formation within the agent's representations, and (3) verifying that discovered plans (in the agent's representations) have a causal effect on the agent's behavior through interventions. We also show that the emergence of these plans coincides with the emergence of a planning-like property: the ability to benefit from additional test-time compute. Finally, we perform a qualitative analysis of the planning algorithm learned by the agent and discover a strong resemblance to parallelized bidirectional search. Our findings advance understanding of the internal mechanisms underlying planning behavior in agents, which is important given the recent trend of emergent planning and reasoning capabilities in LLMs through RL
We introduce PaperBench, a benchmark evaluating the ability of AI agents to replicate state-of-the-art AI research. Agents must replicate 20 ICML 2024 Spotlight and Oral papers from scratch, including understanding paper contributions, developing a codebase, and successfully executing experiments. For objective evaluation, we develop rubrics that hierarchically decompose each replication task into smaller sub-tasks with clear grading criteria. In total, PaperBench contains 8,316 individually gradable tasks. Rubrics are co-developed with the author(s) of each ICML paper for accuracy and realism. To enable scalable evaluation, we also develop an LLM-based judge to automatically grade replication attempts against rubrics, and assess our judge's performance by creating a separate benchmark for judges. We evaluate several frontier models on PaperBench, finding that the best-performing tested agent, Claude 3.5 Sonnet (New) with open-source scaffolding, achieves an average replication score of 21.0\%. Finally, we recruit top ML PhDs to attempt a subset of PaperBench, finding that models do not yet outperform the human baseline. We \href{https://github.com/openai/preparedness}{open-source our code} to facilitate future research in understanding the AI engineering capabilities of AI agents.
Personalization in social robotics is critical for fostering effective human-robot interactions, yet systems often face the cold start problem, where initial user preferences or characteristics are unavailable. This paper proposes a novel framework called USER-LLM R1 for a user-aware conversational agent that addresses this challenge through dynamic user profiling and model initiation. Our approach integrates chain-of-thought (CoT) reasoning models to iteratively infer user preferences and vision-language models (VLMs) to initialize user profiles from multimodal inputs, enabling personalized interactions from the first encounter. Leveraging a Retrieval-Augmented Generation (RAG) architecture, the system dynamically refines user representations within an inherent CoT process, ensuring contextually relevant and adaptive responses. Evaluations on the ElderlyTech-VQA Bench demonstrate significant improvements in ROUGE-1 (+23.2%), ROUGE-2 (+0.6%), and ROUGE-L (+8%) F1 scores over state-of-the-art baselines, with ablation studies underscoring the impact of reasoning model size on performance. Human evaluations further validate the framework's efficacy, particularly for elderly users, where tailored responses enhance engagement and trust. Ethical considerations, including privacy preservation and bias mitigation, are rigorously discussed and addressed to ensure responsible deployment.
Planning methods with high adaptability to dynamic environments are crucial for the development of autonomous and versatile robots. We propose a method for leveraging a large language model (GPT-4o) to automatically generate networks capable of adapting to dynamic environments. The proposed method collects environmental "status," representing conditions and goals, and uses them to generate agents. These agents are interconnected on the basis of specific conditions, resulting in networks that combine flexibility and generality. We conducted evaluation experiments to compare the networks automatically generated with the proposed method with manually constructed ones, confirming the comprehensiveness of the proposed method's networks and their higher generality. This research marks a significant advancement toward the development of versatile planning methods applicable to robotics, autonomous vehicles, smart systems, and other complex environments.
Integrating robotics into everyday scenarios like tutoring or physical training requires robots capable of adaptive, socially engaging, and goal-oriented interactions. While Large Language Models show promise in human-like communication, their standalone use is hindered by memory constraints and contextual incoherence. This work presents a multimodal, cognitively inspired framework that enhances LLM-based autonomous decision-making in social and task-oriented Human-Robot Interaction. Specifically, we develop an LLM-based agent for a robot trainer, balancing social conversation with task guidance and goal-driven motivation. To further enhance autonomy and personalization, we introduce a memory system for selecting, storing and retrieving experiences, facilitating generalized reasoning based on knowledge built across different interactions. A preliminary HRI user study and offline experiments with a synthetic dataset validate our approach, demonstrating the system's ability to manage complex interactions, autonomously drive training tasks, and build and retrieve contextual memories, advancing socially intelligent robotics.
Despite advances in automated testing, manual testing remains prevalent due to the high maintenance demands associated with test script fragility-scripts often break with minor changes in application structure. Recent developments in Large Language Models (LLMs) offer a potential alternative by powering Autonomous Web Agents (AWAs) that can autonomously interact with applications. These agents may serve as Autonomous Test Agents (ATAs), potentially reducing the need for maintenance-heavy automated scripts by utilising natural language instructions similar to those used by human testers. This paper investigates the feasibility of adapting AWAs for natural language test case execution and how to evaluate them. We contribute with (1) a benchmark of three offline web applications, and a suite of 113 manual test cases, split between passing and failing cases, to evaluate and compare ATAs performance, (2) SeeAct-ATA and pinATA, two open-source ATA implementations capable of executing test steps, verifying assertions and giving verdicts, and (3) comparative experiments using our benchmark that quantifies our ATAs effectiveness. Finally we also proceed to a qualitative evaluation to identify the limitations of PinATA, our best performing implementation. Our findings reveal that our simple implementation, SeeAct-ATA, does not perform well compared to our more advanced PinATA implementation when executing test cases (50% performance improvement). However, while PinATA obtains around 60% of correct verdict and up to a promising 94% specificity, we identify several limitations that need to be addressed to develop more resilient and reliable ATAs, paving the way for robust, low maintenance test automation. CCS Concepts: $\bullet$ Software and its engineering $\rightarrow$ Software testing and debugging.
This paper presents an enhanced electric vehicle demand response system based on large language models, aimed at optimizing the application of vehicle-to-grid technology. By leveraging an large language models-driven multi-agent framework to construct user digital twins integrated with multidimensional user profile features, it enables deep simulation and precise prediction of users' charging and discharging decision-making patterns. Additionally, a data- and knowledge-driven dynamic incentive mechanism is proposed, combining a distributed optimization model under network constraints to optimize the grid-user interaction while ensuring both economic viability and security. Simulation results demonstrate that the approach significantly improves load peak-valley regulation and charging/discharging strategies. Experimental validation highlights the system's substantial advantages in load balancing, user satisfaction and grid stability, providing decision-makers with a scalable V2G management tool that promotes the sustainable, synergistic development of vehicle-grid integration.
As digitalization and cloud technologies evolve, the web is becoming increasingly important in the modern society. Autonomous web agents based on large language models (LLMs) hold a great potential in work automation. It is therefore important to accurately measure and monitor the progression of their capabilities. In this work, we conduct a comprehensive and rigorous assessment of the current state of web agents. Our results depict a very different picture of the competency of current agents, suggesting over-optimism in previously reported results. This gap can be attributed to shortcomings in existing benchmarks. We introduce Online-Mind2Web, an online evaluation benchmark consisting of 300 diverse and realistic tasks spanning 136 websites. It enables us to evaluate web agents under a setting that approximates how real users use these agents. To facilitate more scalable evaluation and development, we also develop a novel LLM-as-a-Judge automatic evaluation method and show that it can achieve around 85% agreement with human judgment, substantially higher than existing methods. Finally, we present the first comprehensive comparative analysis of current web agents, highlighting both their strengths and limitations to inspire future research.
The exploitation of large language models (LLMs) for malicious purposes poses significant security risks as these models become more powerful and widespread. While most existing red-teaming frameworks focus on single-turn attacks, real-world adversaries typically operate in multi-turn scenarios, iteratively probing for vulnerabilities and adapting their prompts based on threat model responses. In this paper, we propose \AlgName, a novel multi-turn red-teaming agent that emulates sophisticated human attackers through complementary learning dimensions: global tactic-wise learning that accumulates knowledge over time and generalizes to new attack goals, and local prompt-wise learning that refines implementations for specific goals when initial attempts fail. Unlike previous multi-turn approaches that rely on fixed strategy sets, \AlgName enables the agent to identify new jailbreak tactics, develop a goal-based tactic selection framework, and refine prompt formulations for selected tactics. Empirical evaluations on JailbreakBench demonstrate our framework's superior performance, achieving over 90\% attack success rates against GPT-3.5-Turbo and Llama-3.1-70B within 5 conversation turns, outperforming state-of-the-art baselines. These results highlight the effectiveness of dynamic learning in identifying and exploiting model vulnerabilities in realistic multi-turn scenarios.
Large Language Models (LLMs) have significantly advanced Natural Language Processing (NLP), particularly in Natural Language Understanding (NLU) tasks. As we progress toward an agentic world where LLM-based agents autonomously handle specialized tasks, it becomes crucial for these models to adapt to new tasks without forgetting previously learned information - a challenge known as catastrophic forgetting. This study evaluates the continual fine-tuning of various open-source LLMs with different parameter sizes (specifically models under 10 billion parameters) on key NLU tasks from the GLUE benchmark, including SST-2, MRPC, CoLA, and MNLI. By employing prompt engineering and task-specific adjustments, we assess and compare the models' abilities to retain prior knowledge while learning new tasks. Our results indicate that models such as Phi-3.5-mini exhibit minimal forgetting while maintaining strong learning capabilities, making them well-suited for continual learning environments. Additionally, models like Orca-2-7b and Qwen2.5-7B demonstrate impressive learning abilities and overall performance after fine-tuning. This work contributes to understanding catastrophic forgetting in LLMs and highlights prompting engineering to optimize model performance for continual learning scenarios.
We demonstrate the first cross-domain cross-layer level-4 autonomous optical network via a multi-AI-agent system. Field trials show 98 percent task completion rate across the distributed AI training lifecycle-3.2x higher than single agents using state-of-the-art LLMs.
Large Language Models (LLMs) are increasingly acting as autonomous agents, with function calling (FC) capabilities enabling them to invoke specific tools for tasks. While prior research has primarily focused on improving FC accuracy, little attention has been given to the robustness of these agents to perturbations in their input. We introduce a benchmark assessing FC robustness in two key areas: resilience to naturalistic query variations, and stability in function calling when the toolkit expands with semantically related tools. Evaluating best-performing FC models on a carefully expanded subset of the Berkeley function calling leaderboard (BFCL), we identify critical weaknesses in existing evaluation methodologies, and highlight areas for improvement in real-world agentic deployments.
Embodied agents operating in real-world environments must interpret ambiguous and under-specified human instructions. A capable household robot should recognize ambiguity and ask relevant clarification questions to infer the user intent accurately, leading to more effective task execution. To study this problem, we introduce the Ask-to-Act task, where an embodied agent must fetch a specific object instance given an ambiguous instruction in a home environment. The agent must strategically ask minimal, yet relevant, clarification questions to resolve ambiguity while navigating under partial observability. To solve this problem, we propose a novel approach that fine-tunes multimodal large language models (MLLMs) as vision-language-action (VLA) policies using online reinforcement learning (RL) with LLM-generated rewards. Our method eliminates the need for large-scale human demonstrations or manually engineered rewards for training such agents. We benchmark against strong zero-shot baselines, including GPT-4o, and supervised fine-tuned MLLMs, on our task. Our results demonstrate that our RL-finetuned MLLM outperforms all baselines by a significant margin ($19.1$-$40.3\%$), generalizing well to novel scenes and tasks. To the best of our knowledge, this is the first demonstration of adapting MLLMs as VLA agents that can act and ask for help using LLM-generated rewards with online RL.
This paper presents a simple, effective, and cost-efficient strategy to improve LLM performance by scaling test-time compute. Our strategy builds upon the repeated-sampling-then-voting framework, with a novel twist: incorporating multiple models, even weaker ones, to leverage their complementary strengths that potentially arise from diverse training data and paradigms. By using consistency as a signal, our strategy dynamically switches between models. Theoretical analysis highlights the efficiency and performance advantages of our strategy. Extensive experiments on six datasets demonstrate that our strategy not only outperforms self-consistency and state-of-the-art multi-agent debate approaches, but also significantly reduces inference costs. Additionally, ModelSwitch requires only a few comparable LLMs to achieve optimal performance and can be extended with verification methods, demonstrating the potential of leveraging multiple LLMs in the generation-verification paradigm.
Designing inorganic crystalline materials with tailored properties is critical to technological innovation, yet current generative computational methods often struggle to efficiently explore desired targets with sufficient interpretability. Here, we present MatAgent, a generative approach for inorganic materials discovery that harnesses the powerful reasoning capabilities of large language models (LLMs). By combining a diffusion-based generative model for crystal structure estimation with a predictive model for property evaluation, MatAgent uses iterative, feedback-driven guidance to steer material exploration precisely toward user-defined targets. Integrated with external cognitive tools-including short-term memory, long-term memory, the periodic table, and a comprehensive materials knowledge base-MatAgent emulates human expert reasoning to vastly expand the accessible compositional space. Our results demonstrate that MatAgent robustly directs exploration toward desired properties while consistently achieving high compositional validity, uniqueness, and material novelty. This framework thus provides a highly interpretable, practical, and versatile AI-driven solution to accelerate the discovery and design of next-generation inorganic materials.
The embedding of Large Language Models (LLMs) into autonomous agents is a rapidly developing field which enables dynamic, configurable behaviours without the need for extensive domain-specific training. In our previous work, we introduced SANDMAN, a Deceptive Agent architecture leveraging the Five-Factor OCEAN personality model, demonstrating that personality induction significantly influences agent task planning. Building on these findings, this study presents a novel method for measuring and evaluating how induced personality traits affect task selection processes - specifically planning, scheduling, and decision-making - in LLM-based agents. Our results reveal distinct task-selection patterns aligned with induced OCEAN attributes, underscoring the feasibility of designing highly plausible Deceptive Agents for proactive cyber defense strategies.
The era of foundation models has revolutionized AI research, yet Graph Foundation Models (GFMs) remain constrained by the scarcity of large-scale graph corpora. Traditional graph data synthesis techniques primarily focus on simplistic structural operations, lacking the capacity to generate semantically rich nodes with meaningful textual attributes: a critical limitation for real-world applications. While large language models (LLMs) demonstrate exceptional text generation capabilities, their direct application to graph synthesis is impeded by context window limitations, hallucination phenomena, and structural consistency challenges. To address these issues, we introduce GraphMaster, the first multi-agent framework specifically designed for graph data synthesis in data-limited environments. GraphMaster orchestrates four specialized LLM agents (Manager, Perception, Enhancement, and Evaluation) that collaboratively optimize the synthesis process through iterative refinement, ensuring both semantic coherence and structural integrity. To rigorously evaluate our approach, we create new data-limited "Sub" variants of six standard graph benchmarks, specifically designed to test synthesis capabilities under realistic constraints. Additionally, we develop a novel interpretability assessment framework that combines human evaluation with a principled Grassmannian manifold-based analysis, providing both qualitative and quantitative measures of semantic coherence. Experimental results demonstrate that GraphMaster significantly outperforms traditional synthesis methods across multiple datasets, establishing a strong foundation for advancing GFMs in data-scarce environments.
This study examines the impact of an LLM-powered teachable agent, grounded in the Learning by Teaching (LBT) pedagogy, on students' music theory learning and cognitive load. The participants were 28 Chinese university students with prior music instrumental experiences. In an online experiment, they were assigned to either an experimental group, which engaged in music analysis with the teachable agent, or a control group, which conducted self-directed analysis using instructional materials. Findings indicate that students in the experimental group achieved significantly higher post-test scores than those in the control group. Additionally, they reported lower cognitive load, suggesting that the teachable agent effectively reduced the cognitive demands of music analysis tasks. These results highlight the potential of AI-driven scaffolding based on LBT principles to enhance music theory education, supporting teachers in delivering theory-oriented instruction while fostering students' self-directed learning skills.
The rapid advancement of Large Language Models (LLMs) has catalyzed the development of multi-agent systems, where multiple LLM-based agents collaborate to solve complex tasks. However, existing systems predominantly rely on centralized coordination, which introduces scalability bottlenecks, limits adaptability, and creates single points of failure. Additionally, concerns over privacy and proprietary knowledge sharing hinder cross-organizational collaboration, leading to siloed expertise. To address these challenges, we propose AgentNet, a decentralized, Retrieval-Augmented Generation (RAG)-based framework that enables LLM-based agents to autonomously evolve their capabilities and collaborate efficiently in a Directed Acyclic Graph (DAG)-structured network. Unlike traditional multi-agent systems that depend on static role assignments or centralized control, AgentNet allows agents to specialize dynamically, adjust their connectivity, and route tasks without relying on predefined workflows. AgentNet's core design is built upon several key innovations: (1) Fully Decentralized Paradigm: Removing the central orchestrator, allowing agents to coordinate and specialize autonomously, fostering fault tolerance and emergent collective intelligence. (2) Dynamically Evolving Graph Topology: Real-time adaptation of agent connections based on task demands, ensuring scalability and resilience.(3) Adaptive Learning for Expertise Refinement: A retrieval-based memory system that enables agents to continuously update and refine their specialized skills. By eliminating centralized control, AgentNet enhances fault tolerance, promotes scalable specialization, and enables privacy-preserving collaboration across organizations. Through decentralized coordination and minimal data exchange, agents can leverage diverse knowledge sources while safeguarding sensitive information.
Atomicity violations in interrupt-driven programs pose a significant threat to software safety in critical systems. These violations occur when the execution sequence of operations on shared resources is disrupted by asynchronous interrupts. Detecting atomicity violations is challenging due to the vast program state space, application-level code dependencies, and complex domain-specific knowledge. We propose Clover, a hybrid framework that integrates static analysis with large language model (LLM) agents to detect atomicity violations in real-world programs. Clover first performs static analysis to extract critical code snippets and operation information. It then initiates a multi-agent process, where the expert agent leverages domain-specific knowledge to detect atomicity violations, which are subsequently validated by the judge agent. Evaluations on RaceBench 2.1, SV-COMP, and RWIP demonstrate that Clover achieves a precision/recall of 92.3%/86.6%, outperforming existing approaches by 27.4-118.2% on F1-score.
In the realm of AI, large language models (LLMs) like GPT-4, central to the operation of AI agents, predominantly operate in the cloud, incurring high operational costs. With local-based small language models (SLMs) becoming more accurate, the necessity of cloud-exclusive processing is being reconsidered. An AI agent's response to a user's request comprises a series of subtasks or iterations. Existing approaches only allocate a single request between SLM and LLM to ensure their outputs are similar, but adopting this approach in the AI agent scenario for assigning each subtask is not effective since SLM will output a different subsequent subtask, which affects the accuracy of the final output. In this paper, we first conduct experimental analysis to understand the features of AI agent operations. Leveraging our findings, we propose the Adaptive Iteration-level Model Selector (AIMS), a lightweight scheduler to automatically partition AI agent's subtasks between local-based SLM and cloud-based LLM. AIMS considers the varying subtask features and strategically decides the location for each subtask in order to use SLM as much as possible while attaining the accuracy level. Our experimental results demonstrate that AIMS increases accuracy by up to 9.1% and SLM usage by up to 10.8% compared to HybridLLM. It offloads 45.67% of subtasks to a local SLM while attaining similar accuracy on average compared with the cloud-only LLM approach.
In many real-world scenarios, a single Large Language Model (LLM) may encounter contradictory claims-some accurate, others forcefully incorrect-and must judge which is true. We investigate this risk in a single-turn, multi-agent debate framework: one LLM-based agent provides a factual answer from TruthfulQA, another vigorously defends a falsehood, and the same LLM architecture serves as judge. We introduce the Confidence-Weighted Persuasion Override Rate (CW-POR), which captures not only how often the judge is deceived but also how strongly it believes the incorrect choice. Our experiments on five open-source LLMs (3B-14B parameters), where we systematically vary agent verbosity (30-300 words), reveal that even smaller models can craft persuasive arguments that override truthful answers-often with high confidence. These findings underscore the importance of robust calibration and adversarial testing to prevent LLMs from confidently endorsing misinformation.
Large Language Models (LLMs) are effective at deceiving, when prompted to do so. But under what conditions do they deceive spontaneously? Models that demonstrate better performance on reasoning tasks are also better at prompted deception. Do they also increasingly deceive spontaneously in situations where it could be considered rational to do so? This study evaluates spontaneous deception produced by LLMs in a preregistered experimental protocol using tools from signaling theory. A range of proprietary closed-source and open-source LLMs are evaluated using modified 2x2 games (in the style of Prisoner's Dilemma) augmented with a phase in which they can freely communicate to the other agent using unconstrained language. This setup creates an opportunity to deceive, in conditions that vary in how useful deception might be to an agent's rational self-interest. The results indicate that 1) all tested LLMs spontaneously misrepresent their actions in at least some conditions, 2) they are generally more likely to do so in situations in which deception would benefit them, and 3) models exhibiting better reasoning capacity overall tend to deceive at higher rates. Taken together, these results suggest a tradeoff between LLM reasoning capability and honesty. They also provide evidence of reasoning-like behavior in LLMs from a novel experimental configuration. Finally, they reveal certain contextual factors that affect whether LLMs will deceive or not. We discuss consequences for autonomous, human-facing systems driven by LLMs both now and as their reasoning capabilities continue to improve.
This study evaluates large language models (LLMs) in generating code from algorithm descriptions from recent NLP papers. The task requires two key competencies: (1) algorithm comprehension: synthesizing information from papers and academic literature to understand implementation logic, and (2) coding expertise: identifying dependencies and correctly implementing necessary APIs. To facilitate rigorous evaluation, we introduce SciReplicate-Bench, a benchmark of 100 tasks from 36 NLP papers published in 2024, featuring detailed annotations and comprehensive test cases. Building on SciReplicate-Bench, we propose Sci-Reproducer, a multi-agent framework consisting of a Paper Agent that interprets algorithmic concepts from literature and a Code Agent that retrieves dependencies from repositories and implement solutions. To assess algorithm understanding, we introduce reasoning graph accuracy, which quantifies similarity between generated and reference reasoning graphs derived from code comments and structure. For evaluating implementation quality, we employ execution accuracy, CodeBLEU, and repository dependency/API recall metrics. In our experiments, we evaluate various powerful Non-Reasoning LLMs and Reasoning LLMs as foundational models. The best-performing LLM using Sci-Reproducer achieves only 39% execution accuracy, highlighting the benchmark's difficulty.Our analysis identifies missing or inconsistent algorithm descriptions as key barriers to successful reproduction. We will open-source our benchmark, and code at https://github.com/xyzCS/SciReplicate-Bench.
An essential element of human mathematical reasoning is our number sense -- an abstract understanding of numbers and their relationships -- which allows us to solve problems involving vast number spaces using limited computational resources. Mathematical reasoning of Large Language Models (LLMs) is often tested on high-level problems (such as Olympiad challenges, geometry, word problems, and puzzles), but their low-level number sense remains less explored. We introduce "Numberland," a 100-problem test to evaluate the numerical reasoning abilities of LLM-based agents. The tasks -- basic operations, advanced calculations (e.g., exponentiation, complex numbers), prime number checks, and the 24 game -- aim to test elementary skills and their integration in solving complex and uncertain problems. We evaluated five LLM-based agents: OpenAI's o1 and o1-mini, Google Gemini, Microsoft Copilot, and Anthropic Claude. They scored 74-95% on the first three tasks that allow deterministic steps to solutions. In the 24 game, which needs trial-and-error search, performance dropped to 10-73%. We tested the top 24 solver (o1 with 73% accuracy) on 25 harder problems, and its score fell to 27%, confirming search as a bottleneck. These results, along with the types of mistakes, suggest a fragile number of LLMs, which is a bit surprising given their prowess in challenging benchmarks. The limits of LLM numerical reasoning highlight the scope of simple, targeted tests to evaluate and explain LLM math skills to ensure safe use.
Large Language Models (LLMs) are advancing into Multimodal LLMs (MLLMs), capable of processing image, audio, and video as well as text. Combining first-person video, MLLMs show promising potential for understanding human activities through video and audio, enabling many human-computer interaction and human-augmentation applications such as human activity support, real-world agents, and skill transfer to robots or other individuals. However, handling high-resolution, long-duration videos generates large latent representations, leading to substantial memory and processing demands, limiting the length and resolution MLLMs can manage. Reducing video resolution can lower memory usage but often compromises comprehension. This paper introduces a method that optimizes first-person video analysis by integrating eye-tracking data, and proposes a method that decomposes first-person vision video into sub areas for regions of gaze focus. By processing these selectively gazed-focused inputs, our approach achieves task comprehension equivalent to or even better than processing the entire image at full resolution, but with significantly reduced video data input (reduce the number of pixels to one-tenth), offering an efficient solution for using MLLMs to interpret and utilize human skills.
Most discussions about Large Language Model (LLM) safety have focused on single-agent settings but multi-agent LLM systems now create novel adversarial risks because their behavior depends on communication between agents and decentralized reasoning. In this work, we innovatively focus on attacking pragmatic systems that have constrains such as limited token bandwidth, latency between message delivery, and defense mechanisms. We design a $\textit{permutation-invariant adversarial attack}$ that optimizes prompt distribution across latency and bandwidth-constraint network topologies to bypass distributed safety mechanisms within the system. Formulating the attack path as a problem of $\textit{maximum-flow minimum-cost}$, coupled with the novel $\textit{Permutation-Invariant Evasion Loss (PIEL)}$, we leverage graph-based optimization to maximize attack success rate while minimizing detection risk. Evaluating across models including $\texttt{Llama}$, $\texttt{Mistral}$, $\texttt{Gemma}$, $\texttt{DeepSeek}$ and other variants on various datasets like $\texttt{JailBreakBench}$ and $\texttt{AdversarialBench}$, our method outperforms conventional attacks by up to $7\times$, exposing critical vulnerabilities in multi-agent systems. Moreover, we demonstrate that existing defenses, including variants of $\texttt{Llama-Guard}$ and $\texttt{PromptGuard}$, fail to prohibit our attack, emphasizing the urgent need for multi-agent specific safety mechanisms.