Embodied Conversational Agents (ECAs) aim to emulate human face-to-face interaction through speech, gestures, and facial expressions. Current large language model (LLM)-based conversational agents lack embodiment and the expressive gestures essential for natural interaction. Existing solutions for ECAs often produce rigid, low-diversity motions, that are unsuitable for human-like interaction. Alternatively, generative methods for co-speech gesture synthesis yield natural body gestures but depend on future speech context and require long run-times. To bridge this gap, we present MIBURI, the first online, causal framework for generating expressive full-body gestures and facial expressions synchronized with real-time spoken dialogue. We employ body-part aware gesture codecs that encode hierarchical motion details into multi-level discrete tokens. These tokens are then autoregressively generated by a two-dimensional causal framework conditioned on LLM-based speech-text embeddings, modeling both temporal dynamics and part-level motion hierarchy in real time. Further, we introduce auxiliary objectives to encourage expressive and diverse gestures while preventing convergence to static poses. Comparative evaluations demonstrate that our causal and real-time approach produces natural and contextually aligned gestures against recent baselines. We urge the reader to explore demo videos on https://vcai.mpi-inf.mpg.de/projects/MIBURI/.
Large Language Models (LLMs) demonstrate potentials for automating scientific code generation but face challenges in reliability, error propagation in multi-agent workflows, and evaluation in domains with ill-defined success metrics. We present a Bayesian adversarial multi-agent framework specifically designed for AI for Science (AI4S) tasks in the form of a Low-code Platform (LCP). Three LLM-based agents are coordinated under the Bayesian framework: a Task Manager that structures user inputs into actionable plans and adaptive test cases, a Code Generator that produces candidate solutions, and an Evaluator providing comprehensive feedback. The framework employs an adversarial loop where the Task Manager iteratively refines test cases to challenge the Code Generator, while prompt distributions are dynamically updated using Bayesian principles by integrating code quality metrics: functional correctness, structural alignment, and static analysis. This co-optimization of tests and code reduces dependence on LLM reliability and addresses evaluation uncertainty inherent to scientific tasks. LCP also streamlines human-AI collaboration by translating non-expert prompts into domain-specific requirements, bypassing the need for manual prompt engineering by practitioners without coding backgrounds. Benchmark evaluations demonstrate LCP's effectiveness in generating robust code while minimizing error propagation. The proposed platform is also tested on an Earth Science cross-disciplinary task and demonstrates strong reliability, outperforming competing models.
As large language models (LLMs) advance their mathematical capabilities toward the IMO level, the scarcity of challenging, high-quality problems for training and evaluation has become a significant bottleneck. Simultaneously, recent code agents have demonstrated sophisticated skills in agentic coding and reasoning, suggesting that code execution can serve as a scalable environment for mathematical experimentation. In this paper, we investigate the potential of code agents to autonomously evolve existing math problems into more complex variations. We introduce a multi-agent framework designed to perform problem evolution while validating the solvability and increased difficulty of the generated problems. Our experiments demonstrate that, given sufficient test-time exploration, code agents can synthesize new, solvable problems that are structurally distinct from and more challenging than the originals. This work provides empirical evidence that code-driven agents can serve as a viable mechanism for synthesizing high-difficulty mathematical reasoning problems within scalable computational environments. Our data is available at https://github.com/TarferSoul/Code2Math.
Saarthi is an agentic AI framework that uses multi-agent collaboration to perform end-to-end formal verification. Even though the framework provides a complete flow from specification to coverage closure, with around 40% efficacy, there are several challenges that need to be addressed to make it more robust and reliable. Artificial General Intelligence (AGI) is still a distant goal, and current Large Language Model (LLM)-based agents are prone to hallucinations and making mistakes, especially when dealing with complex tasks such as formal verification. However, with the right enhancements and improvements, we believe that Saarthi can be a significant step towards achieving domain-specific general intelligence for formal verification. Especially for problems that require Short Term, Short Context (STSC) capabilities, such as formal verification, Saarthi can be a powerful tool to assist verification engineers in their work. In this paper, we present two key enhancements to the Saarthi framework: (1) a structured rulebook and specification grammar to improve the accuracy and controllability of SystemVerilog Assertion (SVA) generation, and (2) integration of advanced Retrieval Augmented Generation (RAG) techniques, such as GraphRAG, to provide agents with access to technical knowledge and best practices for iterative refinement and improvement of outputs. We also benchmark these enhancements for the overall Saarthi framework using challenging test cases from NVIDIA's CVDP benchmark targeting formal verification. Our benchmark results stand out with a 70% improvement in the accuracy of generated assertions, and a 50% reduction in the number of iterations required to achieve coverage closure.
In order to flexibly act in an everyday environment, a robotic agent needs a variety of cognitive capabilities that enable it to reason about plans and perform execution recovery. Large language models (LLMs) have been shown to demonstrate emergent cognitive aspects, such as reasoning and language understanding; however, the ability to control embodied robotic agents requires reliably bridging high-level language to low-level functionalities for perception and control. In this paper, we investigate the extent to which an LLM can serve as a core component for planning and execution reasoning in a cognitive robot architecture. For this purpose, we propose a cognitive architecture in which an agentic LLM serves as the core component for planning and reasoning, while components for working and episodic memories support learning from experience and adaptation. An instance of the architecture is then used to control a mobile manipulator in a simulated household environment, where environment interaction is done through a set of high-level tools for perception, reasoning, navigation, grasping, and placement, all of which are made available to the LLM-based agent. We evaluate our proposed system on two household tasks (object placement and object swapping), which evaluate the agent's reasoning, planning, and memory utilisation. The results demonstrate that the LLM-driven agent can complete structured tasks and exhibits emergent adaptation and memory-guided planning, but also reveal significant limitations, such as hallucinations about the task success and poor instruction following by refusing to acknowledge and complete sequential tasks. These findings highlight both the potential and challenges of employing LLMs as embodied cognitive controllers for autonomous robots.
Coverage closure is a critical requirement in Integrated Chip (IC) development process and key metric for verification sign-off. However, traditional exhaustive approaches often fail to achieve full coverage within project timelines. This study presents an agentic AI-driven workflow that utilizes Large Language Model (LLM)-enabled Generative AI (GenAI) to automate coverage analysis for formal verification, identify coverage gaps, and generate the required formal properties. The framework accelerates verification efficiency by systematically addressing coverage holes. Benchmarking open-source and internal designs reveals a measurable increase in coverage metrics, with improvements correlated to the complexity of the design. Comparative analysis validates the effectiveness of this approach. These results highlight the potential of agentic AI-based techniques to improve formal verification productivity and support comprehensive coverage closure.
Scientific discoveries must be communicated clearly to realize their full potential. Without effective communication, even the most groundbreaking findings risk being overlooked or misunderstood. The primary way scientists communicate their work and receive feedback from the community is through peer review. However, the current system often provides inconsistent feedback between reviewers, ultimately hindering the improvement of a manuscript and limiting its potential impact. In this paper, we introduce a novel method APRES powered by Large Language Models (LLMs) to update a scientific papers text based on an evaluation rubric. Our automated method discovers a rubric that is highly predictive of future citation counts, and integrate it with APRES in an automated system that revises papers to enhance their quality and impact. Crucially, this objective should be met without altering the core scientific content. We demonstrate the success of APRES, which improves future citation prediction by 19.6% in mean averaged error over the next best baseline, and show that our paper revision process yields papers that are preferred over the originals by human expert evaluators 79% of the time. Our findings provide strong empirical support for using LLMs as a tool to help authors stress-test their manuscripts before submission. Ultimately, our work seeks to augment, not replace, the essential role of human expert reviewers, for it should be humans who discern which discoveries truly matter, guiding science toward advancing knowledge and enriching lives.
Scholarly data are largely fragmented across siloed databases with divergent metadata and missing linkages among them. We present the Science Data Lake, a locally-deployable infrastructure built on DuckDB and simple Parquet files that unifies eight open sources - Semantic Scholar, OpenAlex, SciSciNet, Papers with Code, Retraction Watch, Reliance on Science, a preprint-to-published mapping, and Crossref - via DOI normalization while preserving source-level schemas. The resource comprises approximately 960GB of Parquet files spanning ~293 million uniquely identifiable papers across ~22 schemas and ~153 SQL views. An embedding-based ontology alignment using BGE-large sentence embeddings maps 4,516 OpenAlex topics to 13 scientific ontologies (~1.3 million terms), yielding 16,150 mappings covering 99.8% of topics ($\geq 0.65$ threshold) with $F1 = 0.77$ at the recommended $\geq 0.85$ operating point, outperforming TF-IDF, BM25, and Jaro-Winkler baselines on a 300-pair gold-standard evaluation. We validate through 10 automated checks, cross-source citation agreement analysis (pairwise Pearson $r = 0.76$ - $0.87$), and stratified manual annotation. Four vignettes demonstrate cross-source analyses infeasible with any single database. The resource is open source, deployable on a single drive or queryable remotely via HuggingFace, and includes structured documentation suitable for large language model (LLM) based research agents.
Large Language Model (LLM)-based agents are increasingly adopted in high-stakes settings, but current benchmarks evaluate mainly whether a task was completed, not how. We introduce Procedure-Aware Evaluation (PAE), a framework that formalizes agent procedures as structured observations and exposes consistency relationships between what agents observe, communicate, and execute. PAE evaluates agents along complementary axes (Utility, Efficiency, Interaction Quality, Procedural Integrity) and applies multi-dimensional gating that categorically disqualifies corrupt outcomes. Evaluating state-of-the-art LLM agents on tau-bench yields findings at the axis, compliance, and benchmark levels. At the axis level, the dimensions capture non-redundant failure modes: utility masks reliability gaps, speed does not imply precision, and conciseness does not predict intent adherence. At the procedural compliance level, 27-78% of benchmark reported successes are corrupt successes concealing violations across interaction and integrity. Furthermore, gating substantially collapses Pass^4 rate and affects model rankings. The analysis of corrupt success cases reveals distinctive per-model failure signatures: GPT-5 spreads errors across policy, execution, and intent dimensions; Kimi-K2-Thinking concentrates 78% of violations in policy faithfulness and compliance; and Mistral-Large-3 is dominated by faithfulness failures. At the benchmark level, our analysis exposes structural flaws in the benchmark design, including task scope gaps, contradictory reward signals, and simulator artifacts that produce accidental successes.
Agentic Reinforcement Learning (Agentic RL) has shown remarkable potential in large language model-based (LLM) agents. These works can empower LLM agents to tackle complex tasks via multi-step, tool-integrated reasoning. However, an inherent limitation of existing Agentic RL methods is their reliance on a pure on-policy paradigm for exploration, restricting exploration to the agent's self-generated outputs and preventing the discovery of new reasoning perspectives for further improvement. While recent efforts incorporate auxiliary off-policy signals to enhance exploration, they typically utilize full off-policy trajectories for trajectory-level policy estimation, overlooking the necessity for the fine-grained, step-level exploratory dynamics within agentic rollout. In this paper, we revisit exploration in Agentic RL and propose Retrieval-Augmented Policy Optimization (RAPO), a novel RL framework that introduces retrieval to explicitly expand exploration during training. To achieve this, we decompose the Agentic RL training process into two phases: (i) Hybrid-policy Agentic Rollout, and (ii) Retrieval-aware Policy Optimization. Specifically, we propose a Hybrid-policy Agentic Rollout strategy, which allows the agents to continuously reason over the retrieved off-policy step-level traces. It dynamically extends the reasoning receptive field of agents, enabling broader exploration conditioned on external behaviors. Subsequently, we introduce the Retrieval-aware Policy Optimization mechanism, which calibrates the policy gradient estimation with retrieval reward and importance shaping, stabilizing training and prioritizing retrieval-illuminating exploration. Extensive experiments show that RAPO achieves an +5.0% average gain on fourteen datasets across three agentic reasoning tasks, while delivering 1.2x faster training efficiency.
Enterprise engineering organizations produce high-volume, heterogeneous telemetry from version control systems, CI/CD pipelines, issue trackers, and observability platforms. Large Language Models (LLMs) enable new forms of agentic automation, but grounding such agents on private telemetry raises three practical challenges: limited model context, locally defined semantic concepts, and evolving metric interfaces. We present REGAL, a registry-driven architecture for deterministic grounding of agentic AI systems in enterprise telemetry. REGAL adopts an explicitly architectural approach: deterministic telemetry computation is treated as a first-class primitive, and LLMs operate over a bounded, version-controlled action space rather than raw event streams. The architecture combines (1) a Medallion ELT pipeline that produces replayable, semantically compressed Gold artifacts, and (2) a registry-driven compilation layer that synthesizes Model Context Protocol (MCP) tools from declarative metric definitions. The registry functions as an "interface-as-code" layer, ensuring alignment between tool specification and execution, mitigating tool drift, and embedding governance policies directly at the semantic boundary. A prototype implementation and case study validate the feasibility of deterministic grounding and illustrate its implications for latency, token efficiency, and operational governance. This work systematizes an architectural pattern for enterprise LLM grounding; it does not propose new learning algorithms, but rather elevates deterministic computation and semantic compilation to first-class design primitives for agentic systems.
Multi-agent large language model frameworks are promising for complex multi step reasoning, yet existing systems remain weak for scientific and knowledge intensive domains due to static prompts and agent roles, rigid workflows, and homogeneous model reliance, leading to poor domain adaptation, limited reasoning flexibility, and high latency on heterogeneous or long-horizon scientific tasks. They also struggle to revise earlier decisions when intermediate reasoning diverges, reducing reliability in structured and calculation heavy settings. To address these limitations, we propose a scientific domain oriented interactive two tier multi model orchestration framework. A dedicated orchestration model analyzes each task, dynamically constructs a domain aware reasoning pipeline, and instantiates specialized expert agents with tailored prompts, while an execution model performs each step under generated role and instruction specifications. The orchestrator iteratively updates the pipeline based on intermediate feedback, enabling dynamic replanning, role reallocation, and prompt refinement across multi turn interactions, strengthening robustness and specialization for scientific reasoning through structured heterogeneous model collaboration. The framework is model agnostic and supports heterogeneous LLM integration with different capacities or costs, enabling flexible performance efficiency trade offs in practical scientific deployments. Experiments show consistent improvements over existing multi agent systems and strong baselines across diverse reasoning and scientific style benchmarks.
LLM agents increasingly act on users' personal information, yet existing privacy defenses remain limited in both design and adaptability. Most prior approaches rely on static or passive defenses, such as prompting and guarding. These paradigms are insufficient for supporting contextual, proactive privacy decisions in multi-step agent execution. We propose Contextualized Defense Instructing (CDI), a new privacy defense paradigm in which an instructor model generates step-specific, context-aware privacy guidance during execution, proactively shaping actions rather than merely constraining or vetoing them. Crucially, CDI is paired with an experience-driven optimization framework that trains the instructor via reinforcement learning (RL), where we convert failure trajectories with privacy violations into learning environments. We formalize baseline defenses and CDI as distinct intervention points in a canonical agent loop, and compare their privacy-helpfulness trade-offs within a unified simulation framework. Results show that our CDI consistently achieves a better balance between privacy preservation (94.2%) and helpfulness (80.6%) than baselines, with superior robustness to adversarial conditions and generalization.
Document-level event argument extraction (DEAE) is essential for knowledge acquisition, aiming to extract participants of events from documents.In the zero-shot setting, existing methods employ LLMs to generate synthetic data to address the challenge posed by the scarcity of annotated data. However, relying solely on Event-type-only prompts makes it difficult for the generated content to accurately capture the contextual and structural relationships of unseen events. Moreover, ensuring the reliability and usability of synthetic data remains a significant challenge due to the absence of quality evaluation mechanisms. To this end, we introduce a multi-agent collaboration framework for zero-shot document-level event argument extraction (ZS-DEAE), which simulates the human collaborative cognitive process of "Propose-Evaluate-Revise." Specifically, the framework comprises a generation agent and an evaluation agent. The generation agent synthesizes data for unseen events by leveraging knowledge from seen events, while the evaluation agent extracts arguments from the synthetic data and assesses their semantic consistency with the context. The evaluation results are subsequently converted into reward signals, with event structure constraints incorporated into the reward design to enable iterative optimization of both agents via reinforcement learning.In three zero-shot scenarios constructed from the RAMS and WikiEvents datasets, our method achieves improvements both in data generation quality and argument extraction performance, while the generated data also effectively enhances the zero-shot performance of other DEAE models.
RTL implementations frequently lack up-to-date or consistent specifications, making comprehension, maintenance, and verification costly and error-prone. While prior work has explored generating specifications from RTL using large language models (LLMs), ensuring that the generated documents faithfully capture design intent remains a major challenge. We present SpecLoop, an agentic framework for RTL-to-specification generation with a formal-verification-driven iterative feedback loop. SpecLoop first generates candidate specifications and then reconstructs RTL from these specifications; it uses formal equivalence checking tools between the reconstructed RTL and the original design to validate functional consistency. When mismatches are detected, counterexamples are fed back to iteratively refine the specifications until equivalence is proven or no further progress can be made. Experiments across multiple LLMs and RTL benchmarks show that incorporating formal verification feedback substantially improves specification correctness and robustness over LLM-only baselines, demonstrating the effectiveness of verification-guided specification generation.
The increasing diversity and scale of video data demand retrieval systems capable of multimodal understanding, adaptive reasoning, and domain-specific knowledge integration. This paper presents LLandMark, a modular multi-agent framework for landmark-aware multimodal video retrieval to handle real-world complex queries. The framework features specialized agents that collaborate across four stages: query parsing and planning, landmark reasoning, multimodal retrieval, and reranked answer synthesis. A key component, the Landmark Knowledge Agent, detects cultural or spatial landmarks and reformulates them into descriptive visual prompts, enhancing CLIP-based semantic matching for Vietnamese scenes. To expand capabilities, we introduce an LLM-assisted image-to-image pipeline, where a large language model (Gemini 2.5 Flash) autonomously detects landmarks, generates image search queries, retrieves representative images, and performs CLIP-based visual similarity matching, removing the need for manual image input. In addition, an OCR refinement module leveraging Gemini and LlamaIndex improves Vietnamese text recognition. Experimental results show that LLandMark achieves adaptive, culturally grounded, and explainable retrieval performance.
As LLM-powered agents have been used for high-stakes decision-making, such as clinical diagnosis, it becomes critical to develop reliable verification of their decisions to facilitate trustworthy deployment. Yet, existing verifiers usually underperform owing to a lack of domain knowledge and limited calibration. To address this, we establish GLEAN, an agent verification framework with Guideline-grounded Evidence Accumulation that compiles expert-curated protocols into trajectory-informed, well-calibrated correctness signals. GLEAN evaluates the step-wise alignment with domain guidelines and aggregates multi-guideline ratings into surrogate features, which are accumulated along the trajectory and calibrated into correctness probabilities using Bayesian logistic regression. Moreover, the estimated uncertainty triggers active verification, which selectively collects additional evidence for uncertain cases via expanding guideline coverage and performing differential checks. We empirically validate GLEAN with agentic clinical diagnosis across three diseases from the MIMIC-IV dataset, surpassing the best baseline by 12% in AUROC and 50% in Brier score reduction, which confirms the effectiveness in both discrimination and calibration. In addition, the expert study with clinicians recognizes GLEAN's utility in practice.
Failure is inevitable for embodied navigation in complex environments. To enhance the resilience, replanning (RP) is a viable option, where the robot is allowed to fail, but is capable of adjusting plan until success. However, existing RP approaches freeze the ego action model and miss the opportunities to explore better plans by upgrading the robot itself. To address this limitation, we propose Self-Evolutionary RePlanning, or SERP for short, which leads to a paradigm shift from frozen models towards evolving models by run-time learning from recent experiences. In contrast to existing model evolution approaches that often get stuck at predefined static parameters, we introduce agentic self-evolving action model that uses in-context learning with auto-differentiation (ILAD) for adaptive function adjustment and global parameter reset. To achieve token-efficient replanning for SERP, we also propose graph chain-of-thought (GCOT) replanning with large language model (LLM) inference over distilled graphs. Extensive simulation and real-world experiments demonstrate that SERP achieves higher success rate with lower token expenditure over various benchmarks, validating its superior robustness and efficiency across diverse environments.
Affective polarization has been central to political and social studies, with growing focus on social media, where partisan divisions are often exacerbated. Real-world studies tend to have limited scope, while simulated studies suffer from insufficient high-quality training data, as manually labeling posts is labor-intensive and prone to subjective biases. The lack of adequate tools to formalize different definitions of affective polarization across studies complicates result comparison and hinders interoperable frameworks. We present a multi-agent model providing a comprehensive approach to studying affective polarization in social media. To operationalize our framework, we develop a platform leveraging large language models (LLMs) to construct virtual communities where agents engage in discussions. We showcase the potential of our platform by (1) analyzing questions related to affective polarization, as explored in social science literature, providing a fresh perspective on this phenomenon, and (2) introducing scenarios that allow observation and measurement of polarization at different levels of granularity and abstraction. Experiments show that our platform is a flexible tool for computational studies of complex social dynamics such as affective polarization. It leverages advanced agent models to simulate rich, context-sensitive interactions and systematically explore research questions traditionally addressed through human-subject studies.
Optimizing communication topology is fundamental to the efficiency and effectiveness of Large Language Model (LLM)-based Multi-Agent Systems (MAS). While recent approaches utilize reinforcement learning to dynamically construct task-specific graphs, they typically rely on single-sample policy gradients with absolute rewards (e.g., binary correctness). This paradigm suffers from severe gradient variance and the credit assignment problem: simple queries yield non-informative positive rewards for suboptimal structures, while difficult queries often result in failures that provide no learning signal. To address these challenges, we propose Graph-GRPO, a novel topology optimization framework that integrates Group Relative Policy Optimization. Instead of evaluating a single topology in isolation, Graph-GRPO samples a group of diverse communication graphs for each query and computes the advantage of specific edges based on their relative performance within the group. By normalizing rewards across the sampled group, our method effectively mitigates the noise derived from task difficulty variance and enables fine-grained credit assignment. Extensive experiments on reasoning and code generation benchmarks demonstrate that Graph-GRPO significantly outperforms state-of-the-art baselines, achieving superior training stability and identifying critical communication pathways previously obscured by reward noise.
While Large Language Models (LLMs) form the cornerstone of sequential decision-making agent development, they have inherent limitations in high-frequency decision tasks. Existing research mainly focuses on discrete embodied decision scenarios with low-frequency and significant semantic differences in state space (e.g., household planning). These methods suffer from limited performance in high-frequency decision-making tasks, since high-precision numerical state information in such tasks undergoes frequent updates with minimal fluctuations, and exhibiting policy misalignment between the learned sub-tasks and composite tasks. To address these issues, this paper proposes Normalized Action Reward guided Consistency Policy Optimization (NAR-CP). 1) Our method first acquires predefined dense rewards from environmental feedback of candidate actions via reward functions, then completes reward shaping through normalization, and theoretically verifies action reward normalization does not impair optimal policy. 2) To reduce policy misalignment in composite tasks, we use LLMs to infer sub-observation candidate actions and generate joint policies, with consistency loss ensuring precise alignment between global semantic policies and sub-semantic policies. Experiments on UAV pursuit, a typical high-frequency task, show our method delivers superior performance on independent and composite tasks with excellent generalization to unseen tasks.
Learning causal structures typically represented by directed acyclic graphs (DAGs) from observational data is notoriously challenging due to the combinatorial explosion of possible graphs and inherent ambiguities in observations. This paper argues that causal learning is now ready for the emergence of a new paradigm supported by rapidly advancing technologies, fulfilling the long-standing vision of leveraging human causal knowledge. This paradigm integrates scalable crowdsourcing platforms for data collection, interactive knowledge elicitation for expert opinion modeling, robust aggregation techniques for expert reconciliation, and large language model (LLM)-based simulation for augmenting AI-driven information acquisition. In this paper, we focus on DAG learning for causal discovery and frame the problem as a distributed decision-making task, recognizing that each participant (human expert or LLM agent) possesses fragmented and imperfect knowledge about different subsets of the variables of interest in the causal graph. By proposing a systematic framework to synthesize these insights, we aim to enable the recovery of a global causal structure unachievable by any individual agent alone.We advocate for a new research frontier and outline a comprehensive framework for new research thrusts that range from eliciting, modeling, aggregating, and optimizing human causal knowledge contributions.
Modern machine learning (ML) workloads increasingly rely on GPUs, yet achieving high end-to-end performance remains challenging due to dependencies on both GPU kernel efficiency and host-side settings. Although LLM-based methods show promise on automated GPU kernel generation, prior works mainly focus on single-kernel optimization and do not extend to end-to-end programs, hindering practical deployment. To address the challenge, in this work, we propose StitchCUDA, a multi-agent framework for end-to-end GPU program generation, with three specialized agents: a Planner to orchestrate whole system design, a Coder dedicated to implementing it step-by-step, and a Verifier for correctness check and performance profiling using Nsys/NCU. To fundamentally improve the Coder's ability in end-to-end GPU programming, StitchCUDA integrates rubric-based agentic reinforcement learning over two atomic skills, task-to-code generation and feedback-driven code optimization, with combined rubric reward and rule-based reward from real executions. Therefore, the Coder learns how to implement advanced CUDA programming techniques (e.g., custom kernel fusion, cublas epilogue), and we also effectively prevent Coder's reward hacking (e.g., just copy PyTorch code or hardcoding output) during benchmarking. Experiments on KernelBench show that StitchCUDA achieves nearly 100% success rate on end-to-end GPU programming tasks, with 1.72x better speedup over the multi-agent baseline and 2.73x than the RL model baselines.
Prompt length is a major bottleneck in agentic large language model (LLM) workloads, where repeated inference steps and multi-call loops incur substantial prefill cost. Recent work on speculative prefill demonstrates that attention-based token importance estimation can enable training-free prompt compression, but this assumes the existence of a draft model that shares the same tokenizer as the target model. In practice, however, agentic pipelines frequently employ models without any smaller in-family draft model. In this work, we study cross-family speculative prefill, where a lightweight draft model from one model family is used to perform prompt compression for a target model from a different family. Using the same speculative prefill mechanism as prior work, we evaluate a range of cross-family draft-target combinations, including Qwen, LLaMA, and DeepSeek models. Across a broad diversity of tasks, we find that attention-based token importance estimation transfers reliably across different model families despite differences in model architectures and tokenizers between draft and target models. Cross-model prompt compression largely retains 90~100% of full-prompt baseline performance and, in some cases, slightly improves accuracy due to denoising effects, while delivering substantial reductions in time to first token (TTFT). These results suggest that speculative prefill depends mainly on task priors and semantic structure, thus serving as a generalizable prompt compression primitive. We discuss the implications of our findings for agentic systems, where repeated long-context inference and heterogeneous model stacks make cross-model prompt compression both necessary and practical.
Large Language Models (LLMs) have achieved great success in many real-world applications, especially the one serving as the cognitive backbone of Multi-Agent Systems (MAS) to orchestrate complex workflows in practice. Since many deployment scenarios preclude MAS workflow modifications and its performance is highly sensitive to the input prompts, prompt optimization emerges as a more natural approach to improve its performance. However, real-world prompt optimization for MAS is impeded by three key challenges: (1) the need of sample efficiency due to prohibitive evaluation costs, (2) topology-induced coupling among prompts, and (3) the combinatorial explosion of the search space. To address these challenges, we introduce MASPOB (Multi-Agent System Prompt Optimization via Bandits), a novel sample-efficient framework based on bandits. By leveraging Upper Confidence Bound (UCB) to quantify uncertainty, the bandit framework balances exploration and exploitation, maximizing gains within a strictly limited budget. To handle topology-induced coupling, MASPOB integrates Graph Neural Networks (GNNs) to capture structural priors, learning topology-aware representations of prompt semantics. Furthermore, it employs coordinate ascent to decompose the optimization into univariate sub-problems, reducing search complexity from exponential to linear. Extensive experiments across diverse benchmarks demonstrate that MASPOB achieves state-of-the-art performance, consistently outperforming existing baselines.
Autonomous web navigation requires agents to perceive complex visual environments and maintain long-term context, yet current Large Language Model (LLM) based agents often struggle with spatial disorientation and navigation loops. In this paper, we propose generally applicable V-GEMS(Visual Grounding and Explicit Memory System), a robust multimodal agent architecture designed for precise and resilient web traversal. Our agent integrates visual grounding to resolve ambiguous interactive elements and introduces an explicit memory stack with state tracking. This dual mechanism allows the agent to maintain a structured map of its traversal path, enabling valid backtracking and preventing cyclical failures in deep navigation tasks. We also introduce an updatable dynamic benchmark to rigorously evaluate adaptability. Experiments show V-GEMS significantly dominates the WebWalker baseline, achieving a substantial 28.7% performance gain. Code is available at https://github.com/Vaultttttttttttt/V-GEMS.
This project reproduces and extends the recently proposed ``Recursive Language Models'' (RLMs) framework by Zhang et al. (2026). This framework enables Large Language Models (LLMs) to process near-infinite contexts by offloading the prompt into an external REPL environment. While the original paper relies on a default recursion depth of 1 and suggests deeper recursion as a future direction, this study specifically investigates the impact of scaling the recursion depth. Using state-of-the-art open-source agentic models (DeepSeek v3.2 and Kimi K2), I evaluated pure LLM, RLM (depth=1), and RLM (depth=2) on the S-NIAH and OOLONG benchmarks. The findings reveal a compelling phenomenon: Deeper recursion causes models to ``overthink''. While depth-1 RLMs effectively boost accuracy on complex reasoning tasks, applying deeper recursion (depth=2) or using RLMs on simple retrieval tasks paradoxically degrades performance and exponentially inflates execution time (e.g., from 3.6s to 344.5s) and token costs. Code and data are available at: https://github.com/drbillwang/rlm-reproduction
We introduce Heterogeneous Agent Collaborative Reinforcement Learning (HACRL), a new learning paradigm that addresses the inefficiencies of isolated on-policy optimization. HACRL enables collaborative optimization with independent execution: heterogeneous agents share verified rollouts during training to mutually improve, while operating independently at inference time. Unlike LLM-based multi-agent reinforcement learning (MARL), HACRL does not require coordinated deployment, and unlike on-/off-policy distillation, it enables bidirectional mutual learning among heterogeneous agents rather than one-directional teacher-to-student transfer. Building on this paradigm, we propose HACPO, a collaborative RL algorithm that enables principled rollout sharing to maximize sample utilization and cross-agent knowledge transfer. To mitigate capability discrepancies and policy distribution shifts, HACPO introduces four tailored mechanisms with theoretical guarantees on unbiased advantage estimation and optimization correctness. Extensive experiments across diverse heterogeneous model combinations and reasoning benchmarks show that HACPO consistently improves all participating agents, outperforming GSPO by an average of 3.3\% while using only half the rollout cost.
Autonomous driving systems require comprehensive evaluation in safety-critical scenarios to ensure safety and robustness. However, such scenarios are rare and difficult to collect from real-world driving data, necessitating simulation-based synthesis. Yet, existing methods often exhibit limitations in both controllability and realism. From a capability perspective, LLMs excel at controllable generation guided by natural language instructions, while diffusion models are better suited for producing trajectories consistent with realistic driving distributions. Leveraging their complementary strengths, we propose AnchorDrive, a two-stage safety-critical scenario generation framework. In the first stage, we deploy an LLM as a driver agent within a closed-loop simulation, which reasons and iteratively outputs control commands under natural language constraints; a plan assessor reviews these commands and provides corrective feedback, enabling semantically controllable scenario generation. In the second stage, the LLM extracts key anchor points from the first-stage trajectories as guidance objectives, which jointly with other guidance terms steer the diffusion model to regenerate complete trajectories with improved realism while preserving user-specified intent. Experiments on the highD dataset demonstrate that AnchorDrive achieves superior overall performance in criticality, realism, and controllability, validating its effectiveness for generating controllable and realistic safety-critical scenarios.
The transition from sequential to parallel computing is essential for modern high-performance applications but is hindered by the steep learning curve of concurrent programming. This challenge is magnified for irregular data structures (such as sparse graphs, unbalanced trees, and non-uniform meshes) where static scheduling fails and data dependencies are unpredictable. Current Large Language Models (LLMs) often fail catastrophically on these tasks, generating code plagued by subtle race conditions, deadlocks, and sub-optimal scaling. We bridge this gap with ParEVO, a framework designed to synthesize high-performance parallel algorithms for irregular data. Our contributions include: (1) The Parlay-Instruct Corpus, a curated dataset of 13,820 tasks synthesized via a "Critic-Refine" pipeline that explicitly filters for empirically performant algorithms that effectively utilize Work-Span parallel primitives; (2) specialized DeepSeek, Qwen, and Gemini models fine-tuned to align probabilistic generation with the rigorous semantics of the ParlayLib library; and (3) an Evolutionary Coding Agent (ECA) that improves the "last mile" of correctness by iteratively repairing code using feedback from compilers, dynamic race detectors, and performance profilers. On the ParEval benchmark, ParEVO achieves an average 106x speedup (with a maximum of 1103x) across the suite, and a robust 13.6x speedup specifically on complex irregular graph problems, outperforming state-of-the-art commercial models. Furthermore, our evolutionary approach matches state-of-the-art expert human baselines, achieving up to a 4.1x speedup on specific highly-irregular kernels. Source code and datasets are available at https://github.com/WildAlg/ParEVO.