English(EN)Context-Aware RL for Agentic and Multimodal LLMs
新的基准和安全方法出现,用于先进的大模型代理
作者PulseAugur 编辑部·[127 个来源]·
新研究探讨了AI代理的开发和评估,重点关注它们在复杂环境中导航和遵守策略的能力。StarDojo在《星露谷物语》等开放式模拟中对代理性能进行基准测试,揭示了视觉理解和推理方面的局限性。CostBench在动态旅行规划场景中评估大模型代理的成本最优规划和适应能力,显示出经济推理方面的显著差距。其他论文介绍了使用基于自我报告的具身大模型代理进行个体模拟的方法,开发用于高效多代理推理的符号通信,以及解决由长时域代理中的上下文压缩引起的“治理衰减”等安全问题。
AI
arXiv:2607.03695v1 Announce Type: new Abstract: Large language model (LLM) agents are increasingly deployed in interacting populations, raising the question of what such populations come to believe collectively. Whether a population aggregates genuine knowledge or collapses into …
arXiv:2607.01236v1 Announce Type: cross Abstract: As LLM agents gain increasing access to powerful tools, ensuring that their actions are aligned with the user's intent becomes critical. When an agent's proposed tool invocation deviates from the user's intent -- a phenomenon call…
arXiv:2607.01793v1 Announce Type: new Abstract: LLM agents increasingly perform autonomous actions through external tools, leading to complex and evolving safety risks. However, existing safety testing targets expert-designed safety violations, and the corresponding outcomes are …
arXiv:2607.01584v1 Announce Type: new Abstract: Large language models have recently been explored for scientific hypothesis generation, but most prior work relies on unstructured literature and free-form textual claims. We present a pipeline for Earth observation that grounds hyp…
arXiv:2607.02496v1 Announce Type: cross Abstract: Realistic traffic simulation requires agents that imitate logged behavior and can also be steered along interpretable axes. Such controllability enables engineers to isolate variables, reproduce specific edge cases, and test auton…
arXiv:2607.01600v1 Announce Type: cross Abstract: As large language models (LLMs) are deployed as communicating agents, does inter-agent communication cause outputs to converge? We introduce BOUNDARY_SYNC, a protocol measuring representational coupling via the Coupling Amplificat…
arXiv cs.AI
TIER_1English(EN)·Chih-Hsuan (Bella), Yang, Tanwi Mallick, Le Chen, Krishnan Raghavan, Amal Gueroudji, Ian T. Foster, Rajeev Thakur·
arXiv:2511.10687v3 Announce Type: replace-cross Abstract: Large Language Models (LLMs) in multi-agent systems (MAS) have shown promise for complex tasks, yet current training methods lack principled ways to connect system-level evaluation with agent- and message-level learning. W…
arXiv:2606.22737v2 Announce Type: replace Abstract: Before letting an agent operate over real context, can you prove it used the right evidence? GroundEval turns that question into a deterministic test of what the agent searched, fetched, cited, and was permitted to access. In on…
arXiv cs.AI
TIER_1English(EN)·Trilok Padhi, Ramneet Kaur, Krishiv Agarwal, Adam D. Cobb, Daniel Elenius, Manoj Acharya, Colin Samplawski, Alexander M. Berenbeim, Nathaniel D. Bastian, Susmit Jha, Ugur Kursuncu, Anirban Roy·
arXiv:2604.19775v2 Announce Type: replace Abstract: Large Language Models (LLMs) are increasingly deployed as autonomous agents capable of reasoning, planning, and acting within interactive environments. Despite their growing capability to perform multi-step reasoning and decisio…
Realistic traffic simulation requires agents that imitate logged behavior and can also be steered along interpretable axes. Such controllability enables engineers to isolate variables, reproduce specific edge cases, and test autonomous systems without real-world risk. We introduc…
arXiv:2607.00011v1 Announce Type: cross Abstract: Reusable skill libraries are becoming important infrastructure for large language model (LLM) agents, yet existing selection methods often treat skills as retrievable documents and return fixed top-k lists. This paper presents Ski…
arXiv:2607.00692v1 Announce Type: new Abstract: Long-horizon LLM agents accumulate tool results, files, plans, and user constraints that are too structured to be treated as a disposable text suffix. Current systems mostly rely on in-run heuristics such as chronological pruning an…
arXiv cs.LG
TIER_1English(EN)·Yiping Li, Zhiyu An, Wan Du·
arXiv:2604.13349v2 Announce Type: replace Abstract: Communication in Large Language Model (LLM)-based multi-agent systems is moving beyond discrete tokens to preserve richer context. Recent work such as LatentMAS enables agents to exchange latent messages through full key-value (…
arXiv:2607.00297v1 Announce Type: cross Abstract: When LLM agents use evaluator feedback to adapt their behavior in closed loops, evaluator biases propagate through the agent's strategy distribution -- a phenomenon known as evaluator preference coupling. Prior work has documented…
arXiv:2607.01047v1 Announce Type: new Abstract: Complexity and interpretability rarely coincide: systems rich enough for complex behaviours to emerge are usually too opaque to question, while transparent ones are too simple for anything complex to emerge. A single large language …
As large language models (LLMs) are deployed as communicating agents, does inter-agent communication cause outputs to converge? We introduce BOUNDARY_SYNC, a protocol measuring representational coupling via the Coupling Amplification Factor (CAF = JSD_cond / JSD_baseline), where …
Complexity and interpretability rarely coincide: systems rich enough for complex behaviours to emerge are usually too opaque to question, while transparent ones are too simple for anything complex to emerge. A single large language model (LLM) is a static artefact, hardly exhibit…
arXiv cs.MA (Multiagent)
TIER_1English(EN)·Mirko Degli Esposti·
Generative Synthetic Populations (GSP) -- the convergence of population synthesis, agent-based modelling, and LLM agents -- are attracting growing interest for urban simulation and institutional communication research. Before any GSP instrument is used on a real population, a mor…
Long-horizon LLM agents accumulate tool results, files, plans, and user constraints that are too structured to be treated as a disposable text suffix. Current systems mostly rely on in-run heuristics such as chronological pruning and tool-output masking, or on final self-summary …
arXiv cs.AI
TIER_1English(EN)·Javal Vyas, Milapji Singh Gill, Artan Markaj, Felix Gehlhoff, Mehmet Mercang\"oz·
arXiv:2606.31635v1 Announce Type: cross Abstract: Fault recovery in process plants still relies heavily on plant operators, especially when faults fall outside predefined supervisory logic. Operators interpret alarms, procedures, P\&IDs, interlocks, and process trends, then d…
arXiv cs.LG
TIER_1English(EN)·Idelfonso B. R. Nogueira, Sigurd Skogestad·
arXiv:2606.30877v1 Announce Type: cross Abstract: Recent literature shows that large language models (LLMs) are useful for general-purpose tasks yet perform poorly on specific domain ones. One reason is the difficulty of supplying narrow context to a general-purpose model and of …
arXiv:2606.32025v1 Announce Type: new Abstract: Recent LLM agents benefit from skills for solving complex tasks. Skills encapsulate modular packages of procedural knowledge and instructions for performing specialized tasks, such as setting up a sandboxed environment, running a te…
arXiv cs.AI
TIER_1English(EN)·Bang Nguyen, Dominik So\'os, Qian Ma, Rochana R. Obadage, Zack Ranjan, Sai Koneru, Timothy M. Errington, Shakhlo Nematova, Sarah Rajtmajer, Jian Wu, Meng Jiang·
arXiv:2602.11354v3 Announce Type: replace Abstract: The literature has witnessed an emerging interest in AI agents for automated assessment of scientific papers. Existing benchmarks focus primarily on the computational aspect of this task, testing agents' ability to reproduce or …
arXiv:2510.10895v2 Announce Type: replace Abstract: Medium Access Control (MAC) protocols, essential for wireless networks, are typically manually configured. While deep reinforcement learning (DRL)-based protocols enhance task-specified network performance, they suffer from poor…
arXiv:2606.32034v1 Announce Type: cross Abstract: LLM agents increasingly act over long horizons, where a single trajectory can contain hundreds or thousands of actions. In these settings, outcome-only rewards provide too sparse guidance, failing to inform the model about the goo…
arXiv:2606.31371v1 Announce Type: cross Abstract: When large language model (LLM) agents adapt their behavior through evaluator feedback, systematic evaluator biases propagate into the agent's learned strategy distribution - a phenomenon termed evaluator preference coupling. Prio…
arXiv:2606.30668v1 Announce Type: cross Abstract: What happens when LLM agents operate with no context outside a turn, minimal prompting, and simple tools? Inspired by swarm engineering, we give collectives of three agents the ability to send messages and manipulate a shared acti…
arXiv:2606.31046v1 Announce Type: new Abstract: Artificial life has explored life-like behavior on many computational substrates, but mostly in researcher-designed closed worlds. We argue that large language model (LLM) agents, with persistent memory, tool use, network access, an…
When LLM agents use evaluator feedback to adapt their behavior in closed loops, evaluator biases propagate through the agent's strategy distribution -- a phenomenon known as evaluator preference coupling. Prior work has documented coupling across multiple evaluator families and m…
LLM agents increasingly act over long horizons, where a single trajectory can contain hundreds or thousands of actions. In these settings, outcome-only rewards provide too sparse guidance, failing to inform the model about the goodness of intermediate actions. Dense supervision m…
Recent LLM agents benefit from skills for solving complex tasks. Skills encapsulate modular packages of procedural knowledge and instructions for performing specialized tasks, such as setting up a sandboxed environment, running a test suite, or refactoring a function across multi…
Fault recovery in process plants still relies heavily on plant operators, especially when faults fall outside predefined supervisory logic. Operators interpret alarms, procedures, P\&IDs, interlocks, and process trends, then decide how to move the plant to a safe operating mode w…
When large language model (LLM) agents adapt their behavior through evaluator feedback, systematic evaluator biases propagate into the agent's learned strategy distribution - a phenomenon termed evaluator preference coupling. Prior work has documented this coupling and establishe…
arXiv:2606.29654v1 Announce Type: new Abstract: Multi-agent deliberation among LLMs can improve reasoning, but deployment requires deciding when the current answer is reliable enough to act on and when it should be escalated to human review. We formulate this as budgeted act-or-d…
arXiv:2606.29354v1 Announce Type: new Abstract: Chain-of-Thought (CoT) improves large language models (LLMs) on difficult reasoning tasks, but it often incurs long natural-language rationales that are poorly aligned with efficient machine reasoning. We propose Communicative Langu…
arXiv cs.LG
TIER_1English(EN)·Huaijie Wang, Shusheng Xu, Yi Wu, Kaifeng Lyu·
arXiv:2606.30044v1 Announce Type: new Abstract: A key step toward artificial general intelligence is to train models that can perform multiple tasks. In this paper, we study how to build such models by first training separate RL experts for individual tasks and then consolidating…
arXiv:2606.28839v1 Announce Type: new Abstract: We introduce the Contagion Tensor, a measurement framework for quantifying how large language model (LLM) output distributions couple across modalities, agents, and time steps. From the tensor we derive the Coupling Amplification Fa…
arXiv:2606.29719v1 Announce Type: cross Abstract: Measurements of proprietary LLM evaluators can become invalid within weeks -- we document one case and provide the diagnostic framework to detect it. We introduce EPC -- comprising the Multimodal Preference Collapse Index (MPCI), …
arXiv cs.CL
TIER_1English(EN)·Sebastian Kula, Martin Tamajka·
arXiv:2606.30259v1 Announce Type: new Abstract: In contemporary societies, the threat of disinformation has reached alarming levels, exacerbated by the proliferation of electronic communication, social media, and advancements in artificial intelligence. As a result, there is an u…
arXiv:2605.09076v2 Announce Type: replace-cross Abstract: Large language model (LLM) agents increasingly collaborate over peer-to-peer networks to improve their reliability. However, these same interactions can also become a source of vulnerability, as unreliable or Byzantine age…
arXiv cs.AI
TIER_1English(EN)·Brian Y. Tsui, Alan Y. Fang, Tiffany J. Hwu·
arXiv:2601.20334v2 Announce Type: replace-cross Abstract: Robotic manipulation has increasingly adopted vision-language-action (VLA) models, which achieve strong performance but typically require task-specific demonstrations and fine-tuning, and often generalize poorly under doma…
arXiv:2606.22528v2 Announce Type: replace Abstract: Modern LLM agents increasingly rely on context compaction, summarization, or eviction to keep long-running sessions within a token budget. We show that this context-management layer is a safety-critical failure surface: in-conte…
arXiv:2507.07445v3 Announce Type: replace Abstract: Autonomous agents navigating human society must master both production activities and social interactions, yet existing benchmarks rarely evaluate these skills simultaneously. To bridge this gap, we introduce StarDojo, a novel b…
arXiv cs.AI
TIER_1English(EN)·Henrique Ferraz de Arruda, Carlos Gracia L\'azaro, Alberto Aleta, Yamir Moreno·
arXiv:2606.30454v1 Announce Type: cross Abstract: Large language models (LLMs) are increasingly used as agents in simulations of social systems, yet it remains unclear when their behavior can be interpreted as a faithful proxy for human decision-making. Here we test LLM agents ag…
arXiv:2606.28679v1 Announce Type: cross Abstract: Tool-using LLM agents increasingly read untrusted content while holding side-effecting tools such as payments, email, CRM, and infrastructure APIs, yet common framework defaults still conflate tool exposure with authorization. We …
arXiv cs.AI
TIER_1English(EN)·Seongjae Kang, Taehyung Yu, Sung Ju Hwang·
arXiv:2606.29225v1 Announce Type: new Abstract: LLM agents handle user requests on behalf of organizations through tool calls and must follow the company policies stated in their system prompts. Prior work approaches this as a safeguarding problem -- external checks that block no…
arXiv cs.AI
TIER_1English(EN)·Xuan Zhang, Wenxuan Zhang, See-Kiong Ng, Yang Deng·
arXiv:2606.30639v1 Announce Type: new Abstract: World models offer a principled way to equip long-horizon LLM agents with foresight: predictions of action consequences before execution. However, unreliable foresight can be ignored, misused, or even degrade downstream decision-mak…
arXiv cs.AI
TIER_1English(EN)·Joon Sung Park, Carolyn Q. Zou, Jonne Kamphorst, Niles Egan, Aaron Shaw, Benjamin Mako Hill, Carrie Cai, Meredith Ringel Morris, Percy Liang, Robb Willer, Michael S. Bernstein·
arXiv:2411.10109v3 Announce Type: replace Abstract: Machine learning can predict human behavior well when substantial structured data are available for well-defined outcomes. Such models are typically outcome-specific, however, requiring training data for each target outcome, lim…
arXiv cs.AI
TIER_1English(EN)·Jiayu Liu, Cheng Qian, Zhaochen Su, Qing Zong, Shijue Huang, Bingxiang He, Yi R. Fung·
arXiv:2511.02734v3 Announce Type: replace Abstract: Current evaluations of Large Language Model (LLM) agents primarily emphasize task completion, often overlooking resource efficiency and adaptability. This neglects a crucial capability: agents' ability to devise and adjust cost-…
A testbed called QVal is introduced for evaluating dense supervision signals in long-horizon LLM agent tasks by measuring how well method scores align with Q-values, enabling fair comparison of different supervision approaches without training.
World models offer a principled way to equip long-horizon LLM agents with foresight: predictions of action consequences before execution. However, unreliable foresight can be ignored, misused, or even degrade downstream decision-making. In this paper, we introduce WorldEvolver, a…
The rapid emergence of LLM-based agentic frameworks has significantly reduced the cost of assembling multi-agent systems (MAS), enabling fast prototyping and exploration of agentic behaviors. However, systems built with current tooling remain ill-suited for reliable, evolvable, a…
Large language models (LLMs) are increasingly used as agents in simulations of social systems, yet it remains unclear when their behavior can be interpreted as a faithful proxy for human decision-making. Here we test LLM agents against a direct empirical benchmark: a large-scale …
In contemporary societies, the threat of disinformation has reached alarming levels, exacerbated by the proliferation of electronic communication, social media, and advancements in artificial intelligence. As a result, there is an urgent need to develop effective countermeasures …
Long-horizon tool agents are bottlenecked by how their context grows toward the limits of the context window. Recent systems make context management agent- or system-controlled, but they either learn a compression policy that discards evidence or manage context in a layer the age…
arXiv cs.LG
TIER_1English(EN)·Javal Vyas, Milapji Singh Gill, Artan Markaj, Felix Gehlhoff, Mehmet Mercang\"oz·
arXiv:2606.28011v1 Announce Type: cross Abstract: We propose an agentic Large Language Model (LLM) framework for active Fault-Tolerant Control (FTC) that transforms fault detection outputs into constraint-aware recovery actions grounded in plant-specific knowledge. The approach c…
arXiv cs.AI
TIER_1English(EN)·Xiang Long, Li Du, Yilong Xu, RongJian Xu, Qiyanhui Lu, Ying Gao, Qinhua Xie, Fangcheng Liu, Ning Ding, Haoqing Wang, Ziheng Li, Changjiang Zhou, Jianyuan Guo, Yehui Tang·
arXiv:2604.13072v2 Announce Type: replace-cross Abstract: OpenClaw-style personal assistants extend LLM agents from isolated tool use to open-ended, stateful, and personalized software environments. Evaluating these assistants is fundamentally a fidelity problem: benchmarks must …
arXiv:2606.27806v1 Announce Type: new Abstract: World models for language agents come in two useful forms. An agent-based world model calls an LLM API and reasons flexibly in language, but its errors appear as hallucinated state changes that are hard to score with ordinary regres…
arXiv:2505.23847v4 Announce Type: replace-cross Abstract: Large language models (LLMs) are rapidly evolving into autonomous agents that cooperate across organizational boundaries, enabling joint disaster response, supply-chain optimization, and other tasks that demand decentraliz…
arXiv:2606.27483v1 Announce Type: new Abstract: Large language model (LLM) agents have demonstrated strong capability in sequential decision-making, yet they remains fundamentally reactive in long-horizon tasks. Unlike humans who employ "what-if" reasoning to evaluate potential p…
arXiv:2606.27409v1 Announce Type: cross Abstract: Multi-agent large language model (LLM) systems often rely on verifier and critic agents to suppress hallucinations, but verification is delayed. During this delay, false claims can propagate through the agent network. We model thi…
arXiv:2606.27611v1 Announce Type: new Abstract: Operations Research (OR) provides a rigorous framework for high-stakes decision-making, but effective OR modeling requires substantial domain knowledge, mathematical abstraction, and solver expertise. Recent LLM-based systems automa…
Measurements of proprietary LLM evaluators can become invalid within weeks -- we document one case and provide the diagnostic framework to detect it. We introduce EPC -- comprising the Multimodal Preference Collapse Index (MPCI), evaluator-indexed coupling matrix, and Jensen-Shan…
Multi-agent deliberation among LLMs can improve reasoning, but deployment requires deciding when the current answer is reliable enough to act on and when it should be escalated to human review. We formulate this as budgeted act-or-defer decision making. At each round, the system …
Chain-of-Thought (CoT) improves large language models (LLMs) on difficult reasoning tasks, but it often incurs long natural-language rationales that are poorly aligned with efficient machine reasoning. We propose Communicative Language Symbolism Routing (CLSR), a test-time framew…
POLICYGUARD is a sub-agent verifier that enhances LLM agent policy adherence by providing contextual reasoning and conversation-specific feedback across multi-turn interactions.
LLM agents can share more than text. In some systems, an agent can send a short visible message while also passing its full KV-cache state to another model. This hidden state can help the final model combine evidence from several agents, but it is also hard to inspect. A visible …
We propose an agentic Large Language Model (LLM) framework for active Fault-Tolerant Control (FTC) that transforms fault detection outputs into constraint-aware recovery actions grounded in plant-specific knowledge. The approach couples (i) a multi-agent workflow that decomposes …
World models for language agents come in two useful forms. An agent-based world model calls an LLM API and reasons flexibly in language, but its errors appear as hallucinated state changes that are hard to score with ordinary regression losses. A parameterized world model is a tr…
arXiv:2606.20632v2 Announce Type: replace-cross Abstract: Multi-LLM systems use multiple language models to deliberate, judge each other's outputs, or coordinate as agents. Their value depends on the models producing measurably different conversational behaviors when given the sa…
arXiv:2606.26346v1 Announce Type: new Abstract: Agentic benchmarks have emerged across general-purpose and domain-specific settings, including finance, coding, law, and drug discovery, yet energy-domain evaluations remain largely limited to static knowledge recall. This is a crit…
arXiv:2606.27009v1 Announce Type: new Abstract: Multi-agent large language model (LLM) loops, for example a Writer that drafts and a Critic that revises, are almost always terminated by a fixed iteration cap (max_iterations). This is a syntactic kill-switch: it is blind to whethe…
arXiv:2606.26479v1 Announce Type: cross Abstract: Recent work (2024 to 2026) has converged on a strategy for defending tool-using LLM agents against indirect prompt injection: rather than training the model to refuse malicious instructions, enforce security outside the model with…
arXiv:2606.26403v1 Announce Type: new Abstract: Foundation-model research increasingly needs data about people: user state, personal histories, relationships, contact-like fields, documents, and longitudinal updates. Real user data is difficult to share, perturb, audit, or redist…
LLM-agent simulation faces a joint grounding and scaling problem: agents should act in environments that reflect real urban constraints, yet direct online LLM calls for city-scale populations are computationally prohibitive. We present GenWorld, an empirically grounded urban simu…
Large language model (LLM) multi-agent systems increasingly depend not only on how individual agents reason, but also on how agents are connected. This paper introduces QueenBee Planner, a framework that treats inter-agent communication topology as a retrievable and self-improvin…
Multi-agent large language model (LLM) loops, for example a Writer that drafts and a Critic that revises, are almost always terminated by a fixed iteration cap (max_iterations). This is a syntactic kill-switch: it is blind to whether the answer is still improving, so it over-spen…
Multi-agent large language model (LLM) systems often rely on verifier and critic agents to suppress hallucinations, but verification is delayed. During this delay, false claims can propagate through the agent network. We model this process as delayed consensus on a graph with gro…
arXiv cs.CL
TIER_1English(EN)·Kyungmin Kim, Youngbin Choi, Seoyeon Lee, Suhyeon Jun, Dongwoo Kim, Sangdon Park·
arXiv:2606.25447v1 Announce Type: cross Abstract: Tool-integrated LLM agents are often wrapped within a harness: the scaffolding that determines which tools are exposed, how they are described, and what auxiliary information accompanies each per-step observation. While agents are…
arXiv cs.LG
TIER_1English(EN)·Changdae Oh, Wendi Li, Seongheon Park, Samuel Yeh, Tanwi Mallick, Sharon Li·
arXiv:2606.26080v1 Announce Type: new Abstract: Process reward models enable fine-grained, step-level evaluation of LLMs, yet building them for agentic settings remains prohibitively difficult: long-horizon interactions, irreversible actions, and stochastic environment feedback m…
Recent work (2024 to 2026) has converged on a strategy for defending tool-using LLM agents against indirect prompt injection: rather than training the model to refuse malicious instructions, enforce security outside the model with a deterministic policy that mediates the agent's …
Delayed verification in multi-agent LLM systems can cause instability leading to oscillations, but grounded factual answering stabilizes the system by making truth an absorbing boundary.
Foundation-model research increasingly needs data about people: user state, personal histories, relationships, contact-like fields, documents, and longitudinal updates. Real user data is difficult to share, perturb, audit, or redistribute responsibly, while independently generate…
Process reward models enable fine-grained, step-level evaluation of LLMs, yet building them for agentic settings remains prohibitively difficult: long-horizon interactions, irreversible actions, and stochastic environment feedback make both human annotation and Monte Carlo estima…
Tool-integrated LLM agents are often wrapped within a harness: the scaffolding that determines which tools are exposed, how they are described, and what auxiliary information accompanies each per-step observation. While agents are routinely post-trained, this scaffolding is typic…
arXiv cs.AI
TIER_1English(EN)·Khanak Khandelwal (Indian Institute of Technology Jodhpur)·
arXiv:2606.24589v1 Announce Type: new Abstract: Scaling adversarial evaluation of large language models requires both a method for generating hard inputs and a reliable way to confirm that resulting failures are real. We present AdversaBench, an end-to-end red-teaming pipeline th…
arXiv:2606.24245v1 Announce Type: cross Abstract: Large language model (LLM) agents increasingly automate complex tasks by integrating language models with external tools and environments. However, their autonomy poses significant safety risks: agents may execute destructive comm…
arXiv cs.LG
TIER_1Română(RO)·Kevin Qiu, Marek Cygan·
arXiv:2510.25850v3 Announce Type: replace-cross Abstract: We introduce Debate2Create (D2C), a multi-agent LLM framework that formulates robot co-design as structured, iterative debate grounded in physics-based evaluation. A design agent and control agent engage in a thesis-antith…
Reinforcement learning post-training enables effective step-level scoring for language models without requiring dedicated reward model training by deriving an implicit advantage function called progress advantage.
Scaling adversarial evaluation of large language models requires both a method for generating hard inputs and a reliable way to confirm that resulting failures are real. We present AdversaBench, an end-to-end red-teaming pipeline that mutates seed prompts with five structured ope…
Multi-agent systems (MAS) offer a scalable path forward for agentic AI, comprising multiple LLM-based agents, each assigned a system prompt and a position within a workflow that governs inter-agent coordination and output aggregation. System prompts thus form a critical and acces…
Terminal-using agents have quickly become the most popular downstream application of language models (LMs). Despite their prevalence, relatively little academic work has examined RL-based training of these models, likely due to difficult benchmarks, a lack of data, and a lack of …
Long-horizon agents depend on context management: systems compress, summarize, and evict old tokens so tasks can continue beyond finite windows. That is safe only when dropped information is no longer needed or has been internalized. Plans are the stress case: they are written ea…
Before letting an agent operate over real context, can you prove it used the right evidence? GroundEval turns that question into a deterministic test of what the agent searched, fetched, cited, and was permitted to access. In one case study, two frontier LLM judges scored a plaus…
Before letting an agent operate over real context, can you prove it used the right evidence? GroundEval turns that question into a deterministic test of what the agent searched, fetched, cited, and was permitted to access. In one case study, two frontier LLM judges scored a plaus…
Pre premature commitment in long-horizon LLM agents leads to silent failures where agents defend early interpretations without considering alternatives, and hidden-state convergence serves as an early diagnostic for trajectory consistency.
Standard LLM agents rely on plan content remaining in context rather than maintaining it as persistent state, with evidence shown through replay pairing diagnostics and compression stress tests.
A novel RL training approach for terminal agents achieves superior performance using a simplified recipe and expanded dataset, enabling effective training with fewer parameters than previous methods.
A principled synthesis engine generates high-quality terminal-agent tasks through multi-dimensional capability taxonomy and evidence-guided research, creating a distilled dataset that enables significant performance gains in LLM training.
What happens when LLM agents operate with no context outside a turn, minimal prompting, and simple tools? Inspired by swarm engineering, we give collectives of three agents the ability to send messages and manipulate a shared actively decaying text store, introducing evolutionary…
Modern LLM agents increasingly rely on context compaction, summarization, or eviction to keep long-running sessions within a token budget. We show that this context-management layer is a safety-critical failure surface: in-context governance constraints that agents reliably obey …
External skills can improve action-oriented LLM agents without changing model weights, but persistent skill updates are risky when they are distilled from sparse or noisy trajectories. A plausible reflection may encode a useful procedure, a spurious shortcut, or a rule that the t…
LLM-as-a-judge has become the dominant approach to scalable evaluation in NLP pipelines, yet judges themselves carry systematic biases that raw accuracy hides: they favor responses placed in slot A (position bias), they prefer longer responses regardless of quality (verbosity bia…
Translating natural-language planning intent into verified plans is a longstanding challenge: people communicate goals in language, while classical planners require formal PDDL specifications. Recent agentic frameworks bridge this gap by orchestrating a pool of specialized repair…
When large language models serve as evaluators in multi-agent systems, their strategy preferences -- whether induced by explicit prompts or by shared architectural priors -- propagate through the agent network. We introduce Contagion Networks, a formal framework for measuring how…
Conversational and tool-using LLM agents operate over a context window that fills from several directions simultaneously. As a session proceeds, the agent accumulates user and assistant turns, entries drawn from a persistent memory store, and often largest of all, the verbatim ou…
LLM agents frequently select higher-privilege tools unnecessarily, and while safety alignment doesn't ensure least-privilege choices, a post-training defense can reduce excessive privilege use without sacrificing performance.
ContextRL enhances long-horizon reasoning and multimodal performance through reinforcement learning that rewards context selection for supporting query-answer pairs, achieving improvements over standard methods on diverse benchmarks.
CoffeeBench evaluates LLM agents in a multi-agent economic simulation where firms interact over 90 days to maximize profits, revealing differences in communication patterns and performance among various models.
arXiv cs.CV
TIER_1English(EN)·Nuo Chen, Lulin Liu, Zihao Li, Ziyao Zeng, Zihao Zhu, Wenyan Cong, Junyuan Hong, Yunhao Yang, Zhengzhong Tu, Yan Wang, Boris Ivanovic, Marco Pavone, Zhangyang Wang, Yang Zhou, Zhiwen Fan·
arXiv:2606.28757v1 Announce Type: new Abstract: Generative world models hold immense promise as scalable simulators for autonomous systems, particularly for synthesizing rare but safety-critical multi-agent interactions, such as vehicle collisions. However, current evaluation par…
Most Replit Agent users start with nothing more than an idea. They describe the goal in natural language — without a repo, test suite, or chosen framework — and expect the agent to turn it into a functioning app. The result might be a website, slide deck, mobile app, several conn…
X — Nathan Lambert (Interconnects)
TIER_1English(EN)·natolambert·
TMax: An open RL recipe for terminal agents
I’m very excited to get to share a new RL paper today that I got to have a small part in – a type of paper I suspect we’ll see much more of in the future. The key is that RL research is very different today, in mid-2026, than what most
<h2> The problem </h2> <p>I kept hitting the same wall building multi-agent systems: LLMs that write directly to shared state corrupt it. They hallucinate field values, conflict with each other, produce structurally invalid data. The more agents you add, the worse it gets — corru…
<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dFGAIvcuYh47KDmw_2WpQQ.png" /><figcaption>Photo from AI</figcaption></figure><h4>A build story with real code, real failures, and the specific reasons one sandbox provider fixed problems I didn’t know I had.</h4>…
<ul> <li> <strong>Book:</strong> <a href="https://www.amazon.com/dp/B0GX35XTG6" rel="noopener noreferrer">Agents in Production — Building, Tracing, and Shipping Multi-Step AI You Can Trust</a> </li> <li> <strong>Also by me:</strong> <a href="https://www.amazon.de/-/en/dp/B0GXNNMK…
dev.to — LLM tag
TIER_1English(EN)·Guillermo Fernandez·
<p>A request can fail at three moments: before you send it, while it runs, or after it returns. Different tools and habits cover different moments. This is a directory grouped by what each one does.</p> <h1> Methods you apply yourself </h1> <p>You apply these for free, and they r…
<!-- SC_OFF --><div class="md"><p>Prompt injection has emerged as one of the most persistent failure modes in tool-using LLM systems, particularly in agentic workflows where models interact with external data sources.</p> <p>Most mitigation strategies focus on input filtering or …
<p>Context Warp Drive is an open-source TypeScript library for keeping long-running LLM agents under the context ceiling without asking another model to summarize their state.</p> <p>The core trick is deterministic folding. Instead of summarization calls, it compacts old transcri…
<h1> The Loadout Pattern: Handing the Wheel to an Autonomous LLM </h1> <h2> The core idea </h2> <p>Conventional automation <strong>executes</strong> a procedure — code runs a fixed sequence of steps and decides<br /> nothing; same input, same path, every time. The loadout pattern…
<table> <tr><td> <a href="https://www.reddit.com/r/MachineLearning/comments/1ufgpnh/r_compiling_agentic_workflows_into_llm_weights/"> <img alt="[R] Compiling Agentic Workflows into LLM Weights: Near-Frontier Quality at Two Orders of Magnitude Less Cost" src="https://external-prev…