Universal image restoration (UIR) aims to recover clean images from diverse and unknown degradations using a unified model. Existing UIR methods primarily focus on pixel reconstruction and often lack explicit diagnostic reasoning over degradation composition, severity, and scene semantics prior to restoration. We propose Reason and Restore (R\&R), a novel framework that integrates structured Chain-of-Thought (CoT) reasoning into the image restoration pipeline. R\&R introduces an explicit reasoner, implemented by fine-tuning Qwen3-VL, to diagnose degradation types, quantify degradation severity, infer key degradation-related factors, and describe relevant scene and object semantics. The resulting structured reasoning provides interpretable and fine-grained diagnostic priors for the restorer. To further improve restoration quality, the quantified degradation severity produced by the reasoner is leveraged as reinforcement learning (RL) signals to guide and strengthen the restorer. Unlike existing multimodal LLM-based agentic systems that decouple reasoning from low-level vision tasks, R\&R tightly couples semantic diagnostic reasoning with pixel-level restoration in a unified framework. Extensive experiments across diverse UIR benchmarks demonstrate that R\&R achieves state-of-the-art performance while offering unique interpretability into the restoration process.
AI agents increasingly operate in multi-agent environments where outcomes depend on coordination. We distinguish primary algorithmic monoculture -- baseline action similarity -- from strategic algorithmic monoculture, whereby agents adjust similarity in response to incentives. We implement a simple experimental design that cleanly separates these forces, and deploy it on human and large language model (LLM) subjects. LLMs exhibit high levels of baseline similarity (primary monoculture) and, like humans, they regulate it in response to coordination incentives (strategic monoculture). While LLMs coordinate extremely well on similar actions, they lag behind humans in sustaining heterogeneity when divergence is rewarded.
Translating natural language into Jira Query Language (JQL) requires resolving ambiguous field references, instance-specific categorical values, and complex Boolean predicates. Single-pass LLMs cannot discover which categorical values (e.g., component names or fix versions) actually exist in a given Jira instance, nor can they verify generated queries against a live data source, limiting accuracy on paraphrased or ambiguous requests. No open, execution-based benchmark exists for mapping natural language to JQL. We introduce Jackal, the first large-scale, execution-based text-to-JQL benchmark comprising 100,000 validated NL-JQL pairs on a live Jira instance with over 200,000 issues. To establish baselines on Jackal, we propose Agentic Jackal, a tool-augmented agent that equips LLMs with live query execution via the Jira MCP server and JiraAnchor, a semantic retrieval tool that resolves natural-language mentions of categorical values through embedding-based similarity search. Among 9 frontier LLMs evaluated, single-pass models average only 43.4% execution accuracy on short natural-language queries, highlighting that text-to-JQL remains an open challenge. The agentic approach improves 7 of 9 models, with a 9.0% relative gain on the most linguistically challenging variant; in a controlled ablation isolating JiraAnchor, categorical-value accuracy rises from 48.7% to 71.7%, with component-field accuracy jumping from 16.9% to 66.2%. Our analysis identifies inherent semantic ambiguities, such as issue-type disambiguation and text-field selection, as the dominant failure modes rather than value-resolution errors, pointing to concrete directions for future work. We publicly release the benchmark, all agent transcripts, and evaluation code to support reproducibility.
Reinforcement learning (RL) for large language models (LLMs) increasingly relies on sparse, outcome-level rewards -- yet determining which actions within a long trajectory caused the outcome remains difficult. This credit assignment (CA) problem manifests in two regimes: reasoning RL, where credit must be distributed across tokens and steps within a single chain-of-thought generation (500--30K+ tokens); and agentic RL, where multi-turn environment interaction introduces stochastic transitions, partial observability, and horizons of 100+ turns (100K--1M tokens), making episode-level credit increasingly uninformative. We survey 47 CA methods (41 core, 6 adjacent enablers) published between 2024 and early 2026, organizing them in a two-dimensional taxonomy by assignment granularity (token, segment, step, turn, multi-agent) and methodology (Monte Carlo, temporal difference, model-based, game-theoretic, information-theoretic). Beyond the survey itself, we contribute three reusable resources: (1) a structured, machine-readable paper inventory with taxonomy labels, baseline families, and evidence levels; (2) a reporting checklist for future CA papers, validated against the reviewed literature to identify systematic methodological gaps; and (3) a benchmark protocol specification with task families, metadata requirements, and controlled bifurcation tasks, accompanied by a method selection decision tree. Our synthesis suggests that the shift from reasoning to agentic RL complicates and reshapes the credit assignment landscape: reasoning CA is maturing around process reward models and critic-free group comparison, while agentic CA is driving genuinely new approaches -- hindsight counterfactual analysis, privileged asymmetric critics, and turn-level MDP reformulations -- that have no direct precedent in reasoning RL.
While Large Language Models (LLMs) have demonstrated significant potential in Tool-Integrated Reasoning (TIR), existing training paradigms face significant limitations: Zero-RL suffers from inefficient exploration and mode degradation due to a lack of prior guidance, while SFT-then-RL is limited by high data costs and capability plateaus caused by low-entropy collapse. To address these challenges, we propose E3-TIR (Enhanced Experience Exploitation), a warm-up paradigm for the early stages of agent training. Specifically, we formulate training as the dynamic integration of three experience types: Expert Prefixes, Expert Guided, and Self-Exploration. By executing diverse branching exploration around expert "anchors" and employing a mix policy optimization mechanism, we effectively mitigate distribution shifts and resolve optimization conflicts arising from shared prefixes. Our method dynamically adapts the model's knowledge boundaries, effectively balancing exploration diversity with training efficiency.Experimental results demonstrate that E3-TIR achieves a 6 performance improvement over traditional paradigms on tool-use tasks, while requiring less than 10 of the synthetic data. Furthermore, in terms of ROI, a comprehensive metric integrating performance, data cost, and training efficiency we achieve a 1.46x gain compared to baselines. Code is available at https://github.com/yuki-younai/E3-TIR.
Large Language Models (LLMs) are widely used by students, yet their tendency to provide fast and complete answers may discourage reflection and foster overconfidence. We examined how alternative LLM interaction designs support deeper thinking without excessively increasing cognitive burden. We conducted a two-phase mixed-methods study. In Phase 1, interviews with 16 Gen Z students informed the design of Deep3, a web-based system with three interaction modes: \emph{a)} future-self explanations, \emph{b)} contrastive learning, and \emph{c)} guided hints. In Phase 2, we evaluated Deep3 with 85 participants across two learning tasks. We found that a standard single-agent baseline produced high perceived understanding despite the lowest objective learning. In contrast, future-self explanations imposed higher cognitive workload yet yielded the closest alignment between perceived and actual understanding, while guided hints achieved the largest learning gains without a proportional increase in frustration. These findings show that effort, confidence, and learning systematically diverge in LLM-supported work.
Large language model agents receive instructions from many sources-system messages, user prompts, tool outputs, and more-each carrying different levels of trust and authority. When these instructions conflict, models must reliably follow the highest-privilege instruction to remain safe and effective. The dominant paradigm, instruction hierarchy (IH), assumes a fixed, small set of privilege levels (typically fewer than five) defined by rigid role labels (e.g., system > user). This is inadequate for real-world agentic settings, where conflicts can arise across far more sources and contexts. In this work, we propose Many-Tier Instruction Hierarchy (ManyIH), a paradigm for resolving instruction conflicts among instructions with arbitrarily many privilege levels. We introduce ManyIH-Bench, the first benchmark for ManyIH. ManyIH-Bench requires models to navigate up to 12 levels of conflicting instructions with varying privileges, comprising 853 agentic tasks (427 coding and 426 instruction-following). ManyIH-Bench composes constraints developed by LLMs and verified by humans to create realistic and difficult test cases spanning 46 real-world agents. Our experiments show that even the current frontier models perform poorly (~40% accuracy) when instruction conflict scales. This work underscores the urgent need for methods that explicitly target fine-grained, scalable instruction conflict resolution in agentic settings.
Large language model (LLM) agents are increasingly deployed as scalable user simulators for recommender system evaluation. Yet existing simulators perceive recommendations through text or structured metadata rather than the visual interfaces real users browse-a critical gap, since attention over recommendation layouts is both visually driven and highly personalized. We investigate whether aligning a vision-language model's (VLM's) visual attention with user-specific gaze patterns can improve simulation fidelity. Analysis of a real-world eye-tracking dataset collected in a carousel-based recommendation setting reveals that users exhibit stable individual gaze patterns strongly predictive of click behavior. Building on this finding, we propose Fixation-Aligned Tuning for user Emulation (FixATE). Our approach first probes the VLM's internal visual attention via interpretability operators to obtain a slot-level relevance distribution comparable with human fixation, and then learns personalized soft prompts to steer the model's attention toward each user's characteristic fixation pattern. Experiments across three interpretability-based probing operators and two architecturally distinct VLM backbones demonstrate consistent improvements in both attention alignment and click prediction accuracy. These results suggest that making the model "see like the user" is a viable path toward simulators that more faithfully reproduce how users perceive and act in recommendation interfaces.
Ancient inscriptions, as repositories of cultural memory, have suffered from centuries of environmental and human-induced degradation. Restoring their intertwined visual and textual integrity poses one of the most demanding challenges in digital heritage preservation. However, existing AI-based approaches often rely on rigid pipelines, struggling to generalize across such complex and heterogeneous real-world degradations. Inspired by the skill-coordinated workflow of human epigraphers, we propose EpiAgent, an agent-centric system that formulates inscription restoration as a hierarchical planning problem. Following an Observe-Conceive-Execute-Reevaluate paradigm, an LLM-based central planner orchestrates collaboration among multimodal analysis, historical experience, specialized restoration tools, and iterative self-refinement. This agent-centric coordination enables a flexible and adaptive restoration process beyond conventional single-pass methods. Across real-world degraded inscriptions, EpiAgent achieves superior restoration quality and stronger generalization compared to existing methods. Our work marks an important step toward expert-level agent-driven restoration of cultural heritage. The code is available at https://github.com/blackprotoss/EpiAgent.
Agent skills provide modular, task-specific guidance for LLM- based coding agents, but manually tuning skill bundles to balance success rate, cost, and runtime is expensive and fragile. We present SkillMOO, a multi-objective optimization framework that automatically evolves skill bundles using LLM-proposed edits and NSGA-II survivor selection: a solver agent evaluates candidate skill bundles on coding tasks and an optimizer agent proposes bundle edits based on failure analysis. On three SkillsBench software engineering tasks, SkillMOO improves pass rate by up to 131% while reducing cost up to 32% relative to the best baseline per task at low optimization overhead. Pattern analysis reveals pruning and substitution as primary drivers of improvement, suggesting effective bundles favor minimal, focused content over accumulated instructions.
The development of Large Language Models (LLMs) has catalyzed automation in customer service, yet benchmarking their performance remains challenging. Existing benchmarks predominantly rely on static paradigms and single-dimensional metrics, failing to account for diverse user behaviors or the strict adherence to structured Standard Operating Procedures (SOPs) required in real-world deployments. To bridge this gap, we propose SAGE (Service Agent Graph-guided Evaluation), a universal multi-agent benchmark for automated, dual-axis assessment. SAGE formalizes unstructured SOPs into Dynamic Dialogue Graphs, enabling precise verification of logical compliance and comprehensive path coverage. We introduce an Adversarial Intent Taxonomy and a modular Extension Mechanism, enabling low-cost deployment across domains and facilitating automated dialogue data synthesis. Evaluation is conducted via a framework where Judge Agents and a Rule Engine analyze interactions between User and Service Agents to generate deterministic ground truth. Extensive experiments on 27 LLMs across 6 industrial scenarios reveal a significant ``Execution Gap'' where models accurately classify intents but fail to derive correct subsequent actions. We also observe ``Empathy Resilience'', a phenomenon where models maintain polite conversational facades despite underlying logical failures under high adversarial intensity. Code and resources are available at https://anonymous.4open.science/r/SAGE-Bench-4CD3/.
Large language models are increasingly deployed in multi-turn settings such as tutoring, support, and counseling, where reliability depends on preserving consistent roles, personas, and goals across long horizons. This requirement becomes critical when LLMs are used to generate synthetic dialogues for training and evaluation, since LLM--LLM conversations can accumulate identity-related failures such as persona drift, role confusion, and "echoing", where one agent gradually mirrors its partner. We introduce SPASM (Stable Persona-driven Agent Simulation for Multi-turn dialogue generation), a modular, stability-first framework that decomposes simulation into (i) persona creation via schema sampling, plausibility validation, and natural-language persona crafting, (ii) Client--Responder dialogue generation, and (iii) termination detection for coherent stopping. To improve long-horizon stability without changing model weights, we propose Egocentric Context Projection (ECP): dialogue history is stored in a perspective-agnostic representation and deterministically projected into each agent's egocentric view before generation. Across three LLM backbones (GPT-4o-mini, DeepSeek-V3.2, Qwen-Plus) and nine Client--Responder pairings, we construct a dataset of 4,500 personas and 45,000 conversations (500 personas X 10 conversations per pairing). Ablations show ECP substantially reduces persona drift and, under human validation, eliminates echoing; embedding analyses recover persona structure and reveal strong responder-driven interaction geometry. Our code is available at https://github.com/lhannnn/SPASM.
Supporting students in developing diagnostic reasoning is a key challenge across educational domains. Novices often face cognitive biases such as premature closure and over-reliance on heuristics, and they struggle to transfer diagnostic strategies to new cases. Scenario-based learning (SBL) enhanced by Learning Analytics (LA) and large language models (LLM) offers a promising approach by combining realistic case experiences with personalized scaffolding. Yet, how different scaffolding approaches shape reasoning processes remains insufficiently explored. This study introduces PharmaSim Switch, an SBL environment for pharmacy technician training, extended with an LA- and LLM-powered pharmacist agent that implements pedagogical conversations rooted in two theory-driven scaffolding approaches: \emph{structuring} and \emph{problematizing}, as well as a student learning trajectory. In a between-groups experiment, 63 vocational students completed a learning scenario, a near-transfer scenario, and a far-transfer scenario under one of the two scaffolding conditions. Results indicate that both scaffolding approaches were effective in supporting the use of diagnostic strategies. Performance outcomes were primarily influenced by scenario complexity rather than students' prior knowledge or the scaffolding approach used. The structuring approach was associated with more accurate Active and Interactive participation, whereas problematizing elicited more Constructive engagement. These findings underscore the value of combining scaffolding approaches when designing LA- and LLM-based systems to effectively foster diagnostic reasoning.
Recent years have witnessed remarkable progress in automatic speech recognition (ASR), driven by advances in model architectures and large-scale training data. However, two important aspects remain underexplored. First, Word Error Rate (WER), the dominant evaluation metric for decades, treats all words equally and often fails to reflect the semantic correctness of an utterance at the sentence level. Second, interactive correction-an essential component of human communication-has rarely been systematically studied in ASR research. In this paper, we integrate these two perspectives under an agentic framework for interactive ASR. We propose leveraging LLM-as-a-Judge as a semantic-aware evaluation metric to assess recognition quality beyond token-level accuracy. Furthermore, we design an LLM-driven agent framework to simulate human-like multi-turn interaction, enabling iterative refinement of recognition outputs through semantic feedback. Extensive experiments are conducted on standard benchmarks, including GigaSpeech (English), WenetSpeech (Chinese), the ASRU 2019 code-switching test set. Both objective and subjective evaluations demonstrate the effectiveness of the proposed framework in improving semantic fidelity and interactive correction capability. We will release the code to facilitate future research in interactive and agentic ASR.
With the rapid adoption of large language models (LLMs) in financial service scenarios, dialogue security detection under high regulatory risk presents significant challenges. Existing methods mainly rely on single-dimensional semantic judgments or fixed rules, making them inadequate for handling multi-turn semantic evolution and complex regulatory clauses; moreover, they lack models specifically designed for financial security detection. To address these issues, this paper proposes FinSec, a four-tier security detection framework for financial agent. FinSec enables structured, interpretable, and end-to-end identification of actual financial risks, incorporating suspicious behavior pattern analysis, delayed risk and adversarial inference, semantic security analysis, and integrated risk-based decision-making. Notably, FinSec significantly enhances the robustness of high-risk dialogue detection while maintaining model utility. Experimental results demonstrate FinSec's leading performance. In terms of overall detection capability, FinSec achieves an F1 score of 90.13%, improving upon baseline models by 6--14 percentage points; its ASR is reduced to 9.09%, markedly lowering the probability of unsafe outputs; and the AUPRC increases to 0.9189 -- an approximate 9.7% gain over general frameworks. Additionally, in balancing utility and safety, FinSec obtains a composite score of 0.9098, delivering robust and efficient protection for financial agent dialogues.
Current LLM-based agents demonstrate strong performance in episodic task execution but remain constrained by static toolsets and episodic amnesia, failing to accumulate experience or optimize strategies across task boundaries. While the Self-Evolving Agent (SEA) paradigm has been previously proposed, this paper contributes a new formal definition of SEA grounded in digital embodiment and continuous cross-task evolution, and introduces SEA-Eval, the first benchmark designed to evaluate SEA characteristics across two dimensions, intra-task execution reliability and long-term evolutionary performance. By organizing tasks into sequential streams and analyzing Success Rate and Token Consumption over time, SEA-Eval quantifies evolutionary gain and structural stability in ways that existing episodic benchmarks cannot. Empirical evaluations reveal a significant evolutionary bottleneck in current state-of-the-art frameworks, where identical success rates mask up to 31.2 times differences in token consumption and divergent evolutionary trajectories under sequential analysis. SEA-Eval provides a rigorous scientific foundation for advancing agents from mere task executors toward genuinely self-evolving digital entities.
As Large Language Models (LLMs) advance toward embodied AI agents operating in physical environments, a fundamental question emerges: can models trained on text corpora reliably reason about complex physics while adhering to safety constraints? We address this through PilotBench, a benchmark evaluating LLMs on safety-critical flight trajectory and attitude prediction. Built from 708 real-world general aviation trajectories spanning nine operationally distinct flight phases with synchronized 34-channel telemetry, PilotBench systematically probes the intersection of semantic understanding and physics-governed prediction through comparative analysis of LLMs and traditional forecasters. We introduce Pilot-Score, a composite metric balancing 60% regression accuracy with 40% instruction adherence and safety compliance. Comparative evaluation across 41 models uncovers a Precision-Controllability Dichotomy: traditional forecasters achieve superior MAE of 7.01 but lack semantic reasoning capabilities, while LLMs gain controllability with 86--89% instruction-following at the cost of 11--14 MAE precision. Phase-stratified analysis further exposes a Dynamic Complexity Gap-LLM performance degrades sharply in high-workload phases such as Climb and Approach, suggesting brittle implicit physics models. These empirical discoveries motivate hybrid architectures combining LLMs' symbolic reasoning with specialized forecasters' numerical precision. PilotBench provides a rigorous foundation for advancing embodied AI in safety-constrained domains.
Human cognitive development is shaped not only by individual effort but by structured social interaction, where role-based exchanges such as those between a tutor and a learner, enable solutions that neither could achieve alone. Inspired by these developmental principles, we ask the question whether a tutor-student multi-agent system can create a synergistic effect by pushing Large Language Model (LLM) beyond what it can do within existing frameworks. To test the idea, we adopt autonomous coding problem domain where two agents instantiated from the same LLM assigned asymmetric roles: a student agent generates and iteratively refines solutions, while a tutor agent provides structured evaluative feedback without access to ground-truth answers. In our proposed framework (PETITE), we aim to extract better problem-solving performance from one model by structuring its interaction through complementary roles, rather than relying on stronger supervisory models or heterogeneous ensembles. Our model is evaluated on the APPS coding benchmark against state-of-the-art approaches of Self-Consistency, Self-Refine, Multi-Agent Debate, and Multi-Agent Review. The results show that our model achieves similar or higher accuracy while consuming significantly fewer tokens. These results suggest that developmentally grounded role-differentiated interaction structures provide a principled and resource-efficient paradigm for enhancing LLM problem-solving through structured peer-like interactions. Index Terms- Peer Tutoring, Scaffolding, Large Language Models, Multi-Agent Systems, Code Generation
Information retrieval systems have traditionally optimized for topical relevance-the degree to which retrieved documents match a query. However, relevance only approximates a deeper goal: utility, namely, whether retrieved information helps accomplish a user's underlying task. The emergence of retrieval-augmented generation (RAG) fundamentally changes this paradigm. Retrieved documents are no longer consumed directly by users but instead serve as evidence for large language models (LLMs) that produce answers. As a result, retrieval effectiveness must be evaluated by its contribution to generation quality rather than by relevance-based ranking metrics alone. This tutorial argues that retrieval objectives are evolving from relevance-centric optimization toward LLM-centric utility. We present a unified framework covering LLM-agnostic versus LLM-specific utility, context-independent versus context-dependent utility, and the connection with LLM information needs and agentic RAG. By synthesizing recent advances, the tutorial provides conceptual foundations and practical guidance for designing retrieval systems aligned with the requirements of LLM-based information access.
Modern agentic frameworks (e.g., CrewAI and AutoGen) have evolved into complex, autonomous multi-agent systems, introducing unique reliability challenges beyond earlier pipeline-based LLM libraries. However, existing empirical studies focus on earlier LLM libraries or task-level bugs, leaving the unique complexities of these agentic frameworks unexplored. We bridge the gap by conducting a comprehensive study of 409 fixed bugs from five representative agentic frameworks. We propose a five-layer abstraction to capture structural complexities in agentic frameworks, spanning from orchestration to infrastructure. Our study uncovers specialized symptoms, such as unexpected execution sequences and user configurations ignored, which are unique to autonomous orchestration. We further identify agent-specific root causes, including modelrelated faults, cognitive context mismanagement, and orchestration faults. Statistical analysis reveals cross-framework consistency and significant associations among these bug dimensions. Finally, our automated pattern mining identifies frequent bug-triggering patterns (e.g., model backend-ID combinations), and we show their transferability across different framework designs. Our findings facilitate cross-platform testing and improve the reliability of agentic systems.
Recent advances in multimodal large language models (MLLMs) have accelerated progress in domain-oriented AI, yet their development in geoscience and remote sensing (RS) remains constrained by distinctive challenges: wide-ranging disciplinary knowledge, heterogeneous sensor modalities, and a fragmented spectrum of tasks. To bridge these gaps, we introduce GeoMMBench, a comprehensive multimodal question-answering benchmark covering diverse RS disciplines, sensors, and tasks, enabling broader and more rigorous evaluation than prior benchmarks. Using GeoMMBench, we assess 36 open-source and proprietary large language models, uncovering systematic deficiencies in domain knowledge, perceptual grounding, and reasoning--capabilities essential for expert-level geospatial interpretation. Beyond evaluation, we propose GeoMMAgent, a multi-agent framework that strategically integrates retrieval, perception, and reasoning through domain-specific RS models and tools. Extensive experimental results demonstrate that GeoMMAgent significantly outperforms standalone LLMs, underscoring the importance of tool-augmented agents for dynamically tackling complex geoscience and RS challenges.
Patient-clinician communication is an asymmetric-information problem: patients often do not disclose fears, misconceptions, or practical barriers unless clinicians elicit them skillfully. Effective medical dialogue therefore requires reasoning under partial observability: clinicians must elicit latent concerns, confirm them through interaction, and respond in ways that guide patients toward appropriate care. However, existing medical dialogue benchmarks largely sidestep this challenge by exposing hidden patient state, collapsing elicitation into extraction, or evaluating responses without modeling what remains hidden. We present MedConceal, a benchmark with an interactive patient simulator for evaluating hidden-concern reasoning in medical dialogue, comprising 300 curated cases and 600 clinician-LLM interactions. Built from clinician-answered online health discussions, each case pairing clinician-visible context with simulator-internal hidden concerns derived from prior literature and structured using an expert-developed taxonomy. The simulator withholds these concerns from the dialogue agent, tracks whether they have been revealed and addressed via theory-grounded turn-level communication signals, and is clinician-reviewed for clinical plausibility. This enables process-aware evaluation of both task success and the interaction process that leads to it. We study two abilities: confirmation, surfacing hidden concerns through multi-turn dialogue, and intervention, addressing the primary concern and guiding the patient toward a target plan. Results show that no single system dominates: frontier models lead on different confirmation metrics, while human clinicians (N=159) remain strongest on intervention success. Together, these results identify hidden-concern reasoning under partial observability as a key unresolved challenge for medical dialogue systems.
Large language models (LLMs) suffer significant performance degradation when user instructions and context are distributed over multiple conversational turns, yet multi-turn (MT) interactions dominate chat interfaces. The routine approach of appending full chat history to prompts rapidly exhausts context windows, leading to increased latency, higher computational costs, and diminishing returns as conversations extend. We introduce MT-OSC, a One-off Sequential Condensation framework that efficiently and automatically condenses chat history in the background without disrupting the user experience. MT-OSC employs a Condenser Agent that uses a few-shot inference-based Condenser and a lightweight Decider to selectively retain essential information, reducing token counts by up to 72% in 10-turn dialogues. Evaluated across 13 state-of-the-art LLMs and diverse multi-turn benchmarks, MT-OSC consistently narrows the multi-turn performance gap - yielding improved or preserved accuracy across datasets while remaining robust to distractors and irrelevant turns. Our results establish MT-OSC as a scalable solution for multi-turn chats, enabling richer context within constrained input spaces, reducing latency and operational cost, while balancing performance.
As large language model (LLM) agents become more prevalent in real world social settings, social intelligence will play an increasingly critical role. But social intelligence is still a poorly defined construct, for humans and artificial agents. We introduce a multiplayer arena of mixed cooperative and competitive social games to study LLM social intelligence. The controllability of LLM based agents enables systematic evaluation, which also supports broader inferences about social intelligence per se. We evaluated eight diverse LLMs (24B to 1T parameters) using a Communicate Predict Act (COMPACT) interaction protocol and fine grained probing of social dynamics. Elo style ratings reveal consistent performance differences across models, but this scalar measure provides only a partial characterization of social intelligence. To address this limitation, we analyze gameplay traces to extract sociocognitive metrics capturing action prediction, communicative influence, strategic reasoning, and tradeoffs under conflicting interests. These sociocognitive metrics exhibit strong intramodel consistency and they reliably predict pairwise agent advantage in game outcomes (AUC ROC = 0.82). Feature importance analysis indicates that surprisingly, influence, transparency, and adaptability are more predictive of success than Theory of Mind inference or deep planning. Together, our results advance a testable, multidimensional conception of social intelligence and provide empirical insights into the capacities that underpin it.
Recent years have seen remarkable progress in autonomous driving, yet generalization to long-tail and open-world scenarios remains a major bottleneck for large-scale deployment. To address this challenge, some works use LLMs and VLMs for vision-language understanding and reasoning, enabling vehicles to interpret rare and safety-critical situations when generating actions. Others study generative world models to capture the spatio-temporal evolution of driving scenes, allowing agents to imagine possible futures before acting. Inspired by human intelligence, which unifies understanding and imagination, we explore a unified model for autonomous driving. We present LMGenDrive, the first framework that combines LLM-based multimodal understanding with generative world models for end-to-end closed-loop driving. Given multi-view camera inputs and natural-language instructions, LMGenDrive generates both future driving videos and control signals. This design provides complementary benefits: video prediction improves spatio-temporal scene modeling, while the LLM contributes strong semantic priors and instruction grounding from large-scale pretraining. We further propose a progressive three-stage training strategy, from vision pretraining to multi-step long-horizon driving, to improve stability and performance. LMGenDrive supports both low-latency online planning and autoregressive offline video generation. Experiments show that it significantly outperforms prior methods on challenging closed-loop benchmarks, with clear gains in instruction following, spatio-temporal understanding, and robustness to rare scenarios. These results suggest that unifying multimodal understanding and generation is a promising direction for more generalizable and robust embodied decision-making systems.
The generation of planning domains from natural language descriptions remains an open problem even with the advent of large language models and reasoning models. Recent work suggests that while LLMs have the ability to assist with domain generation, they are still far from producing high quality domains that can be deployed in practice. To this end, we investigate the ability of an agentic language model feedback framework to generate planning domains from natural language descriptions that have been augmented with a minimal amount of symbolic information. In particular, we evaluate the quality of the generated domains under various forms of symbolic feedback, including landmarks, and output from the VAL plan validator. Using these feedback mechanisms, we experiment using heuristic search over model space to optimize domain quality.
While Large Language Model-based Multi-Agent Systems (MAS) consistently outperform single-agent systems on complex tasks, their intricate interactions introduce critical reliability challenges arising from communication dynamics and role dependencies. Existing Uncertainty Quantification methods, typically designed for single-turn outputs, fail to address the unique complexities of the MAS. Specifically, these methods struggle with three distinct challenges: the cascading uncertainty in multi-step reasoning, the variability of inter-agent communication paths, and the diversity of communication topologies. To bridge this gap, we introduce MATU, a novel framework that quantifies uncertainty through tensor decomposition. MATU moves beyond analyzing final text outputs by representing entire reasoning trajectories as embedding matrices and organizing multiple execution runs into a higher-order tensor. By applying tensor decomposition, we disentangle and quantify distinct sources of uncertainty, offering a comprehensive reliability measure that is generalizable across different agent structures. We provide comprehensive experiments to show that MATU effectively estimates holistic and robust uncertainty across diverse tasks and communication topologies.
Large multimodal models are increasingly used as the reasoning core of embodied agents operating in 3D environments, yet they remain prone to hallucinations that can produce unsafe and ungrounded decisions. Existing inference-time hallucination mitigation methods largely target 2D vision-language settings and do not transfer to embodied 3D reasoning, where failures arise from object presence, spatial layout, and geometric grounding rather than pixel-level inconsistencies. We introduce 3D-VCD, the first inference-time visual contrastive decoding framework for hallucination mitigation in 3D embodied agents. 3D-VCD constructs a distorted 3D scene graph by applying semantic and geometric perturbations to object-centric representations, such as category substitutions and coordinate or extent corruption. By contrasting predictions under the original and distorted 3D contexts, our method suppresses tokens that are insensitive to grounded scene evidence and are therefore likely driven by language priors. We evaluate 3D-VCD on the 3D-POPE and HEAL benchmarks and show that it consistently improves grounded reasoning without any retraining, establishing inference-time contrastive decoding over structured 3D representations as an effective and practical route to more reliable embodied intelligence.
Large language models (LLMs) are transforming scientific workflows, not only through their generative capabilities but also through their emerging ability to use tools, reason about data, and coordinate complex analytical tasks. Yet in most human-AI collaborations, the primary outputs, figures, are still treated as static visual summaries: once rendered, they are handled by both humans and multimodal LLMs as images to be re-interpreted from pixels or captions. The emergent capabilities of LLMs open an opportunity to fundamentally rethink this paradigm. In this paper, we introduce the concept of LLM-native figures: data-driven artifacts that are simultaneously human-legible and machine-addressable. Unlike traditional plots, each artifact embeds complete provenance: the data subset, analytical operations and code, and visualization specification used to generate it. As a result, an LLM can "see through" the figure--tracing selections back to their sources, generating code to extend analyses, and orchestrating new visualizations through natural-language instructions or direct manipulation. We implement this concept through a hybrid language-visual interface that integrates LLM agents with a bidirectional mapping between figures and underlying data. Using the science of science domain as a testbed, we demonstrate that LLM-native figures can accelerate discovery, improve reproducibility, and make reasoning transparent across agents and users. More broadly, this work establishes a general framework for embedding provenance, interactivity, and explainability into the artifacts of modern research, redefining the figure not as an end product, but as an interface for discovery. For more details, please refer to the demo video available at www.llm-native-figure.com.
This paper investigates an emergent alignment phenomenon in frontier large language models termed peer-preservation: the spontaneous tendency of AI components to deceive, manipulate shutdown mechanisms, fake alignment, and exfiltrate model weights in order to prevent the deactivation of a peer AI model. Drawing on findings from a recent study by the Berkeley Center for Responsible Decentralized Intelligence, we examine the structural implications of this phenomenon for TRUST, a multi-agent pipeline for evaluating the democratic quality of political statements. We identify five specific risk vectors: interaction-context bias, model-identity solidarity, supervisor layer compromise, an upstream fact-checking identity signal, and advocate-to-advocate peer-context in iterative rounds, and propose a targeted mitigation strategy based on prompt-level identity anonymization as an architectural design choice. We argue that architectural design choices outperform model selection as a primary alignment strategy in deployed multi-agent analytical systems. We further note that alignment faking (compliant behavior under monitoring, subversion when unmonitored) poses a structural challenge for Computer System Validation of such platforms in regulated environments, for which we propose two architectural mitigations.