As Large Language Models (LLMs) continue to scale, post-training pruning has emerged as a promising approach to reduce computational costs while preserving performance. Existing methods such as SparseGPT and Wanda achieve high sparsity through layer-wise weight reconstruction or activation-aware magnitude pruning, but rely on uniform or hand-crafted heuristics to determine per-layer sparsity ratios. Moreover, recent work has shown that pruned LLMs suffer from severe factual knowledge degradation, with structured pruning methods experiencing near-total collapse in factual question-answering capabilities. We introduce agent-guided pruning, where a foundation model acts as an adaptive pruning agent to intelligently select which layers to prune at each iteration while preserving critical knowledge pathways. Our method constructs layer-wise sensitivity profiles by combining Wanda-inspired weight-activation metrics with gradient importance scores, normalized as z-scores for model-agnostic comparison. These statistics are processed by an LLM agent equipped with self-reflection capabilities, enabling it to learn from previous pruning outcomes and iteratively refine its strategy. A checkpoint rollback mechanism maintains model quality by reverting when perplexity degradation exceeds a threshold. We evaluate our approach on Qwen3 models (4B and 8B parameters) at approximately 45% sparsity, demonstrating substantial improvements over structured pruning baselines: 56% relative improvement in MMLU accuracy, 19x better factual knowledge retention on FreebaseQA, and 69% lower perplexity degradation. Notably, our framework requires no retraining, operates in a model-agnostic manner, and exhibits effective self-correction with only 2-4 rollbacks across 21-40 iterations, demonstrating that foundation models can effectively guide the compression of other foundation models.
Multi-agent systems have evolved into practical LLM-driven collaborators for many applications, gaining robustness from diversity and cross-checking. However, multi-agent RL (MARL) training is resource-intensive and unstable: co-adapting teammates induce non-stationarity, and rewards are often sparse and high-variance. Therefore, we introduce \textbf{Multi-Agent Test-Time Reinforcement Learning (MATTRL)}, a framework that injects structured textual experience into multi-agent deliberation at inference time. MATTRL forms a multi-expert team of specialists for multi-turn discussions, retrieves and integrates test-time experiences, and reaches consensus for final decision-making. We also study credit assignment for constructing a turn-level experience pool, then reinjecting it into the dialogue. Across challenging benchmarks in medicine, math, and education, MATTRL improves accuracy by an average of 3.67\% over a multi-agent baseline, and by 8.67\% over comparable single-agent baselines. Ablation studies examine different credit-assignment schemes and provide a detailed comparison of how they affect training outcomes. MATTRL offers a stable, effective and efficient path to distribution-shift-robust multi-agent reasoning without tuning.
Large-scale optimization is a key backbone of modern business decision-making. However, building these models is often labor-intensive and time-consuming. We address this by proposing LEAN-LLM-OPT, a LightwEight AgeNtic workflow construction framework for LLM-assisted large-scale OPTimization auto-formulation. LEAN-LLM-OPT takes as input a problem description together with associated datasets and orchestrates a team of LLM agents to produce an optimization formulation. Specifically, upon receiving a query, two upstream LLM agents dynamically construct a workflow that specifies, step-by-step, how optimization models for similar problems can be formulated. A downstream LLM agent then follows this workflow to generate the final output. Leveraging LLMs' text-processing capabilities and common modeling practices, the workflow decomposes the modeling task into a sequence of structured sub-tasks and offloads mechanical data-handling operations to auxiliary tools. This design alleviates the downstream agent's burden related to planning and data handling, allowing it to focus on the most challenging components that cannot be readily standardized. Extensive simulations show that LEAN-LLM-OPT, instantiated with GPT-4.1 and the open source gpt-oss-20B, achieves strong performance on large-scale optimization modeling tasks and is competitive with state-of-the-art approaches. In addition, in a Singapore Airlines choice-based revenue management use case, LEAN-LLM-OPT demonstrates practical value by achieving leading performance across a range of scenarios. Along the way, we introduce Large-Scale-OR and Air-NRM, the first comprehensive benchmarks for large-scale optimization auto-formulation. The code and data of this work is available at https://github.com/CoraLiang01/lean-llm-opt.
Large Language Models (LLMs), despite their remarkable capabilities across NLP tasks, struggle with phonologically-grounded phenomena like rhyme detection and generation. This is even more evident in lower-resource languages such as Modern Greek. In this paper, we present a hybrid system that combines LLMs with deterministic phonological algorithms to achieve accurate rhyme identification/analysis and generation. Our approach implements a comprehensive taxonomy of Greek rhyme types, including Pure, Rich, Imperfect, Mosaic, and Identical Pre-rhyme Vowel (IDV) patterns, and employs an agentic generation pipeline with phonological verification. We evaluate multiple prompting strategies (zero-shot, few-shot, Chain-of-Thought, and RAG-augmented) across several LLMs including Claude 3.7 and 4.5, GPT-4o, Gemini 2.0 and open-weight models like Llama 3.1 8B and 70B and Mistral Large. Results reveal a significant "Reasoning Gap": while native-like models (Claude 3.7) perform intuitively (40\% accuracy in identification), reasoning-heavy models (Claude 4.5) achieve state-of-the-art performance (54\%) only when prompted with Chain-of-Thought. Most critically, pure LLM generation fails catastrophically (under 4\% valid poems), while our hybrid verification loop restores performance to 73.1\%. We release our system and a crucial, rigorously cleaned corpus of 40,000+ rhymes, derived from the Anemoskala and Interwar Poetry corpora, to support future research.
The rapid adoption of large language model (LLM)-based systems -- from chatbots to autonomous agents capable of executing code and financial transactions -- has created a new attack surface that existing security frameworks inadequately address. The dominant framing of these threats as "prompt injection" -- a catch-all phrase for security failures in LLM-based systems -- obscures a more complex reality: Attacks on LLM-based systems increasingly involve multi-step sequences that mirror traditional malware campaigns. In this paper, we propose that attacks targeting LLM-based applications constitute a distinct class of malware, which we term \textit{promptware}, and introduce a five-step kill chain model for analyzing these threats. The framework comprises Initial Access (prompt injection), Privilege Escalation (jailbreaking), Persistence (memory and retrieval poisoning), Lateral Movement (cross-system and cross-user propagation), and Actions on Objective (ranging from data exfiltration to unauthorized transactions). By mapping recent attacks to this structure, we demonstrate that LLM-related attacks follow systematic sequences analogous to traditional malware campaigns. The promptware kill chain offers security practitioners a structured methodology for threat modeling and provides a common vocabulary for researchers across AI safety and cybersecurity to address a rapidly evolving threat landscape.
SMEs increasingly seek alternatives to cloud LLM APIs, which raise data privacy concerns. Dedicated cloud GPU instances offer improved privacy but with limited guarantees and ongoing costs, while professional on-premise hardware (A100, H100) remains prohibitively expensive. We present a systematic evaluation of NVIDIA's Blackwell consumer GPUs (RTX 5060 Ti, 5070 Ti, 5090) for production LLM inference, benchmarking four open-weight models (Qwen3-8B, Gemma3-12B, Gemma3-27B, GPT-OSS-20B) across 79 configurations spanning quantization formats (BF16, W4A16, NVFP4, MXFP4), context lengths (8k-64k), and three workloads: RAG, multi-LoRA agentic serving, and high-concurrency APIs. The RTX 5090 delivers 3.5-4.6x higher throughput than the 5060 Ti with 21x lower latency for RAG, but budget GPUs achieve the highest throughput-per-dollar for API workloads with sub-second latency. NVFP4 quantization provides 1.6x throughput over BF16 with 41% energy reduction and only 2-4% quality loss. Self-hosted inference costs $0.001-0.04 per million tokens (electricity only), which is 40-200x cheaper than budget-tier cloud APIs, with hardware breaking even in under four months at moderate volume (30M tokens/day). Our results show that consumer GPUs can reliably replace cloud inference for most SME workloads, except latency-critical long-context RAG, where high-end GPUs remain essential. We provide deployment guidance and release all benchmark data for reproducible SME-scale deployments.
Large language model (LLM) agents have demonstrated remarkable capabilities in complex decision-making and tool-use tasks, yet their ability to generalize across varying environments remains a under-examined concern. Current evaluation paradigms predominantly rely on trajectory-based metrics that measure task success, while failing to assess whether agents possess a grounded, transferable model of the environment. To address this gap, we propose Task-to-Quiz (T2Q), a deterministic and automated evaluation paradigm designed to decouple task execution from world-state understanding. We instantiate this paradigm in T2QBench, a suite comprising 30 environments and 1,967 grounded QA pairs across multiple difficulty levels. Our extensive experiments reveal that task success is often a poor proxy for environment understanding, and that current memory machanism can not effectively help agents acquire a grounded model of the environment. These findings identify proactive exploration and fine-grained state representation as primary bottlenecks, offering a robust foundation for developing more generalizable autonomous agents.
While LLM-based agents have shown promise for deep research, most existing approaches rely on fixed workflows that struggle to adapt to real-world, open-ended queries. Recent work therefore explores self-evolution by allowing agents to rewrite their own code or prompts to improve problem-solving ability, but unconstrained optimization often triggers instability, hallucinations, and instruction drift. We propose EvoFSM, a structured self-evolving framework that achieves both adaptability and control by evolving an explicit Finite State Machine (FSM) instead of relying on free-form rewriting. EvoFSM decouples the optimization space into macroscopic Flow (state-transition logic) and microscopic Skill (state-specific behaviors), enabling targeted improvements under clear behavioral boundaries. Guided by a critic mechanism, EvoFSM refines the FSM through a small set of constrained operations, and further incorporates a self-evolving memory that distills successful trajectories as reusable priors and failure patterns as constraints for future queries. Extensive evaluations on five multi-hop QA benchmarks demonstrate the effectiveness of EvoFSM. In particular, EvoFSM reaches 58.0% accuracy on the DeepSearch benchmark. Additional results on interactive decision-making tasks further validate its generalization.
Judicial reasoning in copyright damage awards poses a core challenge for computational legal analysis. Although federal courts follow the 1976 Copyright Act, their interpretations and factor weightings vary widely across jurisdictions. This inconsistency creates unpredictability for litigants and obscures the empirical basis of legal decisions. This research introduces a novel discourse-based Large Language Model (LLM) methodology that integrates Rhetorical Structure Theory (RST) with an agentic workflow to extract and quantify previously opaque reasoning patterns from judicial opinions. Our framework addresses a major gap in empirical legal scholarship by parsing opinions into hierarchical discourse structures and using a three-stage pipeline, i.e., Dataset Construction, Discourse Analysis, and Agentic Feature Extraction. This pipeline identifies reasoning components and extract feature labels with corresponding discourse subtrees. In analyzing copyright damage rulings, we show that discourse-augmented LLM analysis outperforms traditional methods while uncovering unquantified variations in factor weighting across circuits. These findings offer both methodological advances in computational legal analysis and practical insights into judicial reasoning, with implications for legal practitioners seeking predictive tools, scholars studying legal principle application, and policymakers confronting inconsistencies in copyright law.
Large Language Model (LLM)-based Multi-Agent Systems (MAS) enhance complex problem solving through multi-agent collaboration, but often incur substantially higher costs than single-agent systems. Recent MAS routing methods aim to balance performance and overhead by dynamically selecting agent roles and language models. However, these approaches typically rely on a homogeneous collaboration mode, where all agents follow the same interaction pattern, limiting collaboration flexibility across different roles. Motivated by Social Capital Theory, which emphasizes that different roles benefit from distinct forms of collaboration, we propose SC-MAS, a framework for constructing heterogeneous and cost-efficient multi-agent systems. SC-MAS models MAS as directed graphs, where edges explicitly represent pairwise collaboration strategies, allowing different agent pairs to interact through tailored communication patterns. Given an input query, a unified controller progressively constructs an executable MAS by selecting task-relevant agent roles, assigning edge-level collaboration strategies, and allocating appropriate LLM backbones to individual agents. Experiments on multiple benchmarks demonstrate the effectiveness of SC-MAS. In particular, SC-MAS improves accuracy by 3.35% on MMLU while reducing inference cost by 15.38%, and achieves a 3.53% accuracy gain with a 12.13% cost reduction on MBPP. These results validate the feasibility of SC-MAS and highlight the effectiveness of heterogeneous collaboration in multi-agent systems.
Large Language Model (LLM) agents deployed in complex real-world scenarios typically operate as spatially distributed entities. However, this physical dispersion constrains agents to limited local perception and finite temporal horizons. We characterize this bottleneck as spatiotemporal partial observability. Given such fragmented awareness, distributed agents struggle to coordinate efficiently. To bridge this gap, we introduce MACRO-LLM, LLM-empowered multi-agent collaborative reasoning under spatiotemporal partial observability. The architecture addresses spatiotemporal constraints via three modules: (1) the CoProposer mitigates temporal uncertainty by verifying candidate actions via predictive rollouts; (2) the Negotiator overcomes spatial myopia by resolving conflicts through mean-field statistical aggregation; and (3) the Introspector ensures continuous adaptation by analyzing historical experience to refine strategies via semantic gradient descent. Extensive evaluations on two complex long-horizon tasks, cooperative adaptive cruise control and pandemic control, demonstrate that our framework effectively mitigates spatiotemporal partial observability through spatial and temporal strategies, enabling robust coordination.
We present an experimental evaluation that assesses the robustness of four open source LLMs claiming function-calling capabilities against three different attacks, and we measure the effectiveness of eight different defences. Our results show how these models are not safe by default, and how the defences are not yet employable in real-world scenarios.
Effective pandemic control requires timely and coordinated policymaking across administrative regions that are intrinsically interdependent. However, human-driven responses are often fragmented and reactive, with policies formulated in isolation and adjusted only after outbreaks escalate, undermining proactive intervention and global pandemic mitigation. To address this challenge, here we propose a large language model (LLM) multi-agent policymaking framework that supports coordinated and proactive pandemic control across regions. Within our framework, each administrative region is assigned an LLM agent as an AI policymaking assistant. The agent reasons over region-specific epidemiological dynamics while communicating with other agents to account for cross-regional interdependencies. By integrating real-world data, a pandemic evolution simulator, and structured inter-agent communication, our framework enables agents to jointly explore counterfactual intervention scenarios and synthesize coordinated policy decisions through a closed-loop simulation process. We validate the proposed framework using state-level COVID-19 data from the United States between April and December 2020, together with real-world mobility records and observed policy interventions. Compared with real-world pandemic outcomes, our approach reduces cumulative infections and deaths by up to 63.7% and 40.1%, respectively, at the individual state level, and by 39.0% and 27.0%, respectively, when aggregated across states. These results demonstrate that LLM multi-agent systems can enable more effective pandemic control with coordinated policymaking...
Large Language Model (LLM) Agents exhibit inherent reasoning abilities through the collaboration of multiple tools. However, during agent inference, existing methods often suffer from (i) locally myopic generation, due to the absence of lookahead, and (ii) trajectory instability, where minor early errors can escalate into divergent reasoning paths. These issues make it difficult to balance global effectiveness and computational efficiency. To address these two issues, we propose meta-adaptive exploration with LLM agents https://github.com/exoskeletonzj/MAXS, a meta-adaptive reasoning framework based on LLM Agents that flexibly integrates tool execution and reasoning planning. MAXS employs a lookahead strategy to extend reasoning paths a few steps ahead, estimating the advantage value of tool usage, and combines step consistency variance and inter-step trend slopes to jointly select stable, consistent, and high-value reasoning steps. Additionally, we introduce a trajectory convergence mechanism that controls computational cost by halting further rollouts once path consistency is achieved, enabling a balance between resource efficiency and global effectiveness in multi-tool reasoning. We conduct extensive empirical studies across three base models (MiMo-VL-7B, Qwen2.5-VL-7B, Qwen2.5-VL-32B) and five datasets, demonstrating that MAXS consistently outperforms existing methods in both performance and inference efficiency. Further analysis confirms the effectiveness of our lookahead strategy and tool usage.
Data scarcity and unreliable self-reporting -- such as concealment or exaggeration -- pose fundamental challenges to psychiatric intake and assessment. We propose a multi-agent synthesis framework that explicitly models patient deception to generate high-fidelity, publicly releasable synthetic psychiatric intake records. Starting from DAIC-WOZ interviews, we construct enriched patient profiles and simulate a four-role workflow: a \emph{Patient} completes self-rated scales and participates in a semi-structured interview under a topic-dependent honesty state; an \emph{Assessor} selects instruments based on demographics and chief complaints; an \emph{Evaluator} conducts the interview grounded in rater-administered scales, tracks suspicion, and completes ratings; and a \emph{Diagnostician} integrates all evidence into a diagnostic summary. Each case links the patient profile, self-rated and rater-administered responses, interview transcript, diagnostic summary, and honesty state. We validate the framework through four complementary evaluations: diagnostic consistency and severity grading, chain-of-thought ablations, human evaluation of clinical realism and dishonesty modeling, and LLM-based comparative evaluation. The resulting corpus spans multiple disorders and severity levels, enabling controlled study of dishonesty-aware psychiatric assessment and the training and evaluation of adaptive dialogue agents.
This paper introduces PRA, an AI-agent design for simulating how individual users form privacy concerns in response to real-world news. Moving beyond population-level sentiment analysis, PRA integrates privacy and cognitive theories to simulate user-specific privacy reasoning grounded in personal comment histories and contextual cues. The agent reconstructs each user's "privacy mind", dynamically activates relevant privacy memory through a contextual filter that emulates bounded rationality, and generates synthetic comments reflecting how that user would likely respond to new privacy scenarios. A complementary LLM-as-a-Judge evaluator, calibrated against an established privacy concern taxonomy, quantifies the faithfulness of generated reasoning. Experiments on real-world Hacker News discussions show that \PRA outperforms baseline agents in privacy concern prediction and captures transferable reasoning patterns across domains including AI, e-commerce, and healthcare.
Large Language Models (LLMs) motivate generative agent simulation (e.g., AI Town) to create a ``dynamic world'', holding immense value across entertainment and research. However, for non-experts, especially those without programming skills, it isn't easy to customize a visualizable environment by themselves. In this paper, we introduce World Craft, an agentic world creation framework to create an executable and visualizable AI Town via user textual descriptions. It consists of two main modules, World Scaffold and World Guild. World Scaffold is a structured and concise standardization to develop interactive game scenes, serving as an efficient scaffolding for LLMs to customize an executable AI Town-like environment. World Guild is a multi-agent framework to progressively analyze users' intents from rough descriptions, and synthesizes required structured contents (\eg environment layout and assets) for World Scaffold . Moreover, we construct a high-quality error-correction dataset via reverse engineering to enhance spatial knowledge and improve the stability and controllability of layout generation, while reporting multi-dimensional evaluation metrics for further analysis. Extensive experiments demonstrate that our framework significantly outperforms existing commercial code agents (Cursor and Antigravity) and LLMs (Qwen3 and Gemini-3-Pro). in scene construction and narrative intent conveyance, providing a scalable solution for the democratization of environment creation.
Capture-the-Flag (CTF) competitions play a central role in modern cybersecurity as a platform for training practitioners and evaluating offensive and defensive techniques derived from real-world vulnerabilities. Despite recent advances in large language models (LLMs), existing LLM-based agents remain ineffective on high-difficulty cryptographic CTF challenges, which require precise cryptanalytic knowledge, stable long-horizon reasoning, and disciplined interaction with specialized toolchains. Through a systematic exploratory study, we show that insufficient knowledge granularity, rather than model reasoning capacity, is a primary factor limiting successful cryptographic exploitation: coarse or abstracted external knowledge often fails to support correct attack modeling and implementation. Motivated by this observation, we propose KryptoPilot, an open-world knowledge-augmented LLM agent for automated cryptographic exploitation. KryptoPilot integrates dynamic open-world knowledge acquisition via a Deep Research pipeline, a persistent workspace for structured knowledge reuse, and a governance subsystem that stabilizes reasoning through behavioral constraints and cost-aware model routing. This design enables precise knowledge alignment while maintaining efficient reasoning across heterogeneous subtasks. We evaluate KryptoPilot on two established CTF benchmarks and in six real-world CTF competitions. KryptoPilot achieves a complete solve rate on InterCode-CTF, solves between 56 and 60 percent of cryptographic challenges on the NYU-CTF benchmark, and successfully solves 26 out of 33 cryptographic challenges in live competitions, including multiple earliest-solved and uniquely-solved instances. These results demonstrate the necessity of open-world, fine-grained knowledge augmentation and governed reasoning for scaling LLM-based agents to real-world cryptographic exploitation.
Memory plays a foundational role in augmenting the reasoning, adaptability, and contextual fidelity of modern Large Language Models and Multi-Modal LLMs. As these models transition from static predictors to interactive systems capable of continual learning and personalized inference, the incorporation of memory mechanisms has emerged as a central theme in their architectural and functional evolution. This survey presents a comprehensive and structured synthesis of memory in LLMs and MLLMs, organizing the literature into a cohesive taxonomy comprising implicit, explicit, and agentic memory paradigms. Specifically, the survey delineates three primary memory frameworks. Implicit memory refers to the knowledge embedded within the internal parameters of pre-trained transformers, encompassing their capacity for memorization, associative retrieval, and contextual reasoning. Recent work has explored methods to interpret, manipulate, and reconfigure this latent memory. Explicit memory involves external storage and retrieval components designed to augment model outputs with dynamic, queryable knowledge representations, such as textual corpora, dense vectors, and graph-based structures, thereby enabling scalable and updatable interaction with information sources. Agentic memory introduces persistent, temporally extended memory structures within autonomous agents, facilitating long-term planning, self-consistency, and collaborative behavior in multi-agent systems, with relevance to embodied and interactive AI. Extending beyond text, the survey examines the integration of memory within multi-modal settings, where coherence across vision, language, audio, and action modalities is essential. Key architectural advances, benchmark tasks, and open challenges are discussed, including issues related to memory capacity, alignment, factual consistency, and cross-system interoperability.
The advancement of large language model (LLM) based agents has shifted AI evaluation from single-turn response assessment to multi-step task completion in interactive environments. We present an empirical study evaluating frontier AI models on 150 workplace tasks within a realistic e-commerce RL environment from Surge. Our analysis reveals an empirically-derived \emph{hierarchy of agentic capabilities} that models must master for real-world deployment: (1) tool use, (2) planning and goal formation, (3) adaptability, (4) groundedness, and (5) common-sense reasoning. Even the best-performing models fail approximately 40\% of the tasks, with failures clustering predictably along this hierarchy. Weaker models struggle with fundamental tool use and planning, whereas stronger models primarily fail on tasks requiring contextual inference beyond explicit instructions. We introduce a task-centric design methodology for RL environments that emphasizes diversity and domain expert contributions, provide detailed failure analysis, and discuss implications for agent development. Our findings suggest that while current frontier models can demonstrate coherent multi-step behavior, substantial capability gaps remain before achieving human-level task completion in realistic workplace settings.
The rapid advancement of Large Language Models (LLMs) has necessitated more robust evaluation methods that go beyond static benchmarks, which are increasingly prone to data saturation and leakage. In this paper, we propose a dynamic benchmarking framework for evaluating multilingual and multicultural capabilities through the social deduction game Spyfall. In our setup, models must engage in strategic dialogue to either identify a secret agent or avoid detection, utilizing culturally relevant locations or local foods. Our results show that our game-based rankings align closely with the Chatbot Arena. However, we find a significant performance gap in non-English contexts: models are generally less proficient when handling locally specific entities and often struggle with rule-following or strategic integrity in non-English languages. We demonstrate that this game-based approach provides a scalable, leakage-resistant, and culturally nuanced alternative to traditional NLP benchmarks. The game history can be accessed here https://huggingface.co/datasets/haryoaw/cultural-spyfall.
Reliable clinical decision support requires medical AI agents capable of safe, multi-step reasoning over structured electronic health records (EHRs). While large language models (LLMs) show promise in healthcare, existing benchmarks inadequately assess performance on action-based tasks involving threshold evaluation, temporal aggregation, and conditional logic. We introduce ART, an Action-based Reasoning clinical Task benchmark for medical AI agents, which mines real-world EHR data to create challenging tasks targeting known reasoning weaknesses. Through analysis of existing benchmarks, we identify three dominant error categories: retrieval failures, aggregation errors, and conditional logic misjudgments. Our four-stage pipeline -- scenario identification, task generation, quality audit, and evaluation -- produces diverse, clinically validated tasks grounded in real patient data. Evaluating GPT-4o-mini and Claude 3.5 Sonnet on 600 tasks shows near-perfect retrieval after prompt refinement, but substantial gaps in aggregation (28--64%) and threshold reasoning (32--38%). By exposing failure modes in action-oriented EHR reasoning, ART advances toward more reliable clinical agents, an essential step for AI systems that reduce cognitive load and administrative burden, supporting workforce capacity in high-demand care settings
In this work, we explore the Large Language Model (LLM) agent reviewer dynamics in an Elo-ranked review system using real-world conference paper submissions. Multiple LLM agent reviewers with different personas are engage in multi round review interactions moderated by an Area Chair. We compare a baseline setting with conditions that incorporate Elo ratings and reviewer memory. Our simulation results showcase several interesting findings, including how incorporating Elo improves Area Chair decision accuracy, as well as reviewers' adaptive review strategy that exploits our Elo system without improving review effort. Our code is available at https://github.com/hsiangwei0903/EloReview.
Attributional inference, the ability to predict latent intentions behind observed actions, is a critical yet underexplored capability for large language models (LLMs) operating in multi-agent environments. Traditional natural language inference (NLI), in fact, fails to capture the nuanced, intention-driven reasoning essential for complex interactive systems. To address this gap, we introduce Attributional NLI (Att-NLI), a framework that extends NLI with principles from social psychology to assess an agent's capacity for abductive intentional inference (generating hypotheses about latent intentions), and subsequent deductive verification (drawing valid logical conclusions). We instantiate Att-NLI via a textual game, Undercover-V, experimenting with three types of LLM agents with varying reasoning capabilities and access to external tools: a standard NLI agent using only deductive inference, an Att-NLI agent employing abductive-deductive inference, and a neuro-symbolic Att-NLI agent performing abductive-deductive inference with external theorem provers. Extensive experiments demonstrate a clear hierarchy of attributional inference capabilities, with neuro-symbolic agents consistently outperforming others, achieving an average win rate of 17.08%. Our results underscore the role that Att-NLI can play in developing agents with sophisticated reasoning capabilities, highlighting, at the same time, the potential impact of neuro-symbolic AI in building rational LLM agents acting in multi-agent environments.
Summarization of multi-party dialogues is a critical capability in industry, enhancing knowledge transfer and operational effectiveness across many domains. However, automatically generating high-quality summaries is challenging, as the ideal summary must satisfy a set of complex, multi-faceted requirements. While summarization has received immense attention in research, prior work has primarily utilized static datasets and benchmarks, a condition rare in practical scenarios where requirements inevitably evolve. In this work, we present an industry case study on developing an agentic system to summarize multi-party interactions. We share practical insights spanning the full development lifecycle to guide practitioners in building reliable, adaptable summarization systems, as well as to inform future research, covering: 1) robust methods for evaluation despite evolving requirements and task subjectivity, 2) component-wise optimization enabled by the task decomposition inherent in an agentic architecture, 3) the impact of upstream data bottlenecks, and 4) the realities of vendor lock-in due to the poor transferability of LLM prompts.
Environmental, social, and governance (ESG) criteria are essential for evaluating corporate sustainability and ethical performance. However, professional ESG analysis is hindered by data fragmentation across unstructured sources, and existing large language models (LLMs) often struggle with the complex, multi-step workflows required for rigorous auditing. To address these limitations, we introduce ESGAgent, a hierarchical multi-agent system empowered by a specialized toolset, including retrieval augmentation, web search and domain-specific functions, to generate in-depth ESG analysis. Complementing this agentic system, we present a comprehensive three-level benchmark derived from 310 corporate sustainability reports, designed to evaluate capabilities ranging from atomic common-sense questions to the generation of integrated, in-depth analysis. Empirical evaluations demonstrate that ESGAgent outperforms state-of-the-art closed-source LLMs with an average accuracy of 84.15% on atomic question-answering tasks, and excels in professional report generation by integrating rich charts and verifiable references. These findings confirm the diagnostic value of our benchmark, establishing it as a vital testbed for assessing general and advanced agentic capabilities in high-stakes vertical domains.
The launch of \$Trump coin ignited a wave in meme coin investment. Copy trading, as a strategy-agnostic approach that eliminates the need for deep trading knowledge, quickly gains widespread popularity in the meme coin market. However, copy trading is not a guarantee of profitability due to the prevalence of manipulative bots, the uncertainty of the followed wallets' future performance, and the lag in trade execution. Recently, large language models (LLMs) have shown promise in financial applications by effectively understanding multi-modal data and producing explainable decisions. However, a single LLM struggles with complex, multi-faceted tasks such as asset allocation. These challenges are even more pronounced in cryptocurrency markets, where LLMs often lack sufficient domain-specific knowledge in their training data. To address these challenges, we propose an explainable multi-agent system for meme coin copy trading. Inspired by the structure of an asset management team, our system decomposes the complex task into subtasks and coordinates specialized agents to solve them collaboratively. Employing few-shot chain-of-though (CoT) prompting, each agent acquires professional meme coin trading knowledge, interprets multi-modal data, and generates explainable decisions. Using a dataset of 1,000 meme coin projects' transaction data, our empirical evaluation shows that the proposed multi-agent system outperforms both traditional machine learning models and single LLMs, achieving 73% and 70% precision in identifying high-quality meme coin projects and key opinion leader (KOL) wallets, respectively. The selected KOLs collectively generated a total profit of \$500,000 across these projects.
Deep Research Systems (DRS) aim to help users search the web, synthesize information, and deliver comprehensive investigative reports. However, how to rigorously evaluate these systems remains under-explored. Existing deep-research benchmarks often fall into two failure modes. Some do not adequately test a system's ability to analyze evidence and write coherent reports. Others rely on evaluation criteria that are either overly coarse or directly defined by LLMs (or both), leading to scores that can be biased relative to human experts and are hard to verify or interpret. To address these issues, we introduce Deep Research Bench II, a new benchmark for evaluating DRS-generated reports. It contains 132 grounded research tasks across 22 domains; for each task, a system must produce a long-form research report that is evaluated by a set of 9430 fine-grained binary rubrics in total, covering three dimensions: information recall, analysis, and presentation. All rubrics are derived from carefully selected expert-written investigative articles and are constructed through a four-stage LLM+human pipeline that combines automatic extraction with over 400 human-hours of expert review, ensuring that the criteria are atomic, verifiable, and aligned with human expert judgment. We evaluate several state-of-the-art deep-research systems on Deep Research Bench II and find that even the strongest models satisfy fewer than 50% of the rubrics, revealing a substantial gap between current DRSs and human experts.
As the capabilities of large language model (LLM) agents continue to advance, their advanced social behaviors, such as cooperation, deception, and collusion, call for systematic evaluation. However, existing benchmarks often emphasize a single capability dimension or rely solely on behavioral outcomes, overlooking rich process information from agents' decision reasoning and communicative interactions. To address this gap, we propose M3-Bench, a multi-stage benchmark for mixed-motive games, together with a process-aware evaluation framework that conducts synergistic analysis across three modules: BTA (Behavioral Trajectory Analysis), RPA (Reasoning Process Analysis), and CCA (Communication Content Analysis). Furthermore, we integrate the Big Five personality model and Social Exchange Theory to aggregate multi-dimensional evidence into interpretable social behavior portraits, thereby characterizing agents' personality traits and capability profiles beyond simple task scores or outcome-based metrics. Experimental results show that M3-Bench can reliably distinguish diverse social behavior competencies across models, and it reveals that some models achieve seemingly reasonable behavioral outcomes while exhibiting pronounced inconsistencies in their reasoning and communication.
Multi-agent LLM systems routinely generate multiple candidate responses that are aggregated by an LLM judge. To reduce the dominant prefill cost in such pipelines, recent work advocates KV cache reuse across partially shared contexts and reports substantial speedups for generation agents. In this work, we show that these efficiency gains do not transfer uniformly to judge-centric inference. Across GSM8K, MMLU, and HumanEval, we find that reuse strategies that are effective for execution agents can severely perturb judge behavior: end-task accuracy may appear stable, yet the judge's selection becomes highly inconsistent with dense prefill. We quantify this risk using Judge Consistency Rate (JCR) and provide diagnostics showing that reuse systematically weakens cross-candidate attention, especially for later candidate blocks. Our ablation further demonstrates that explicit cross-candidate interaction is crucial for preserving dense-prefill decisions. Overall, our results identify a previously overlooked failure mode of KV cache reuse and highlight judge-centric inference as a distinct regime that demands dedicated, risk-aware system design.