LLM agents

topic54 events

papersTODAY 04:00 UTC

MTAC-IFBench: Benchmarking Instruction-Following in Multi-Turn Agentic Coding

A new arXiv paper introduces MTAC-IFBench, a benchmark aimed at measuring how well large language model agents follow instructions across multi-turn coding sessions. The work targets agentic software engineering, where models plan, run code, and call external tools over successive steps rather than producing a single answer. It addresses evaluation beyond functional correctness, focusing on whether agents keep to the constraints given to them.

agents/benchmark/code-generation/instruction-following

papersTODAY 04:00 UTC

ClimateAgent: Multi-Agent Orchestration for Climate Data Science Workflows

Researchers present ClimateAgent, a system that coordinates multiple LLM-based agents to carry out climate data science tasks. The work argues that general-purpose agents and fixed script pipelines lack the domain context needed for large, varied climate datasets, and proposes an orchestration approach tailored to that setting. It is a revised arXiv preprint (v2) in machine learning.

agents/multi-agent/climate/research

papersTODAY 04:00 UTC

Quantization-Conditioned Backdoor Attacks Target Open-Weight LLM Agents

A new arXiv paper describes an attack in which an adversary releases a full-precision model checkpoint that passes standard audits but behaves maliciously once it is quantized for deployment. Because quantization is a common default path for running open-weight agent models, the technique could let compromised agents slip past pre-release checks. The work frames this as a supply-chain risk for quantized LLM deployments.

safety/agents/quantization/backdoor-attacks

papersTODAY 04:00 UTC

GRADE Models LLM Agent Runs as Typed Graphs of Execution and Dependency

A new arXiv paper introduces GRADE, a method that represents a single LLM agent run as a typed graph rather than a plain step-by-step trace. Execution edges are derived automatically from the trace, while dependency edges are provided separately, with each one carrying a graded observation. The authors argue this extra dependency layer captures what each step relied on, information that ordinary traces omit.

agents/llm-agents/research/graph-representation

papersTODAY 04:00 UTC

arXiv Paper Describes Persistent Memory Poisoning Attack on Harness-Based LLM Agents

A new arXiv preprint examines how harness-based LLM agents, which combine memory, tool use, and runtime control, can be compromised through stored malicious instructions. The authors argue that once such instructions enter an agent's persistent memory, they can continue to influence later behavior, creating security and privacy exposure. The work frames memory poisoning as a distinct risk for agent architectures that retain context across sessions.

safety/agents/security/memory

papersTODAY 04:00 UTC

arXiv Report Critiques Transient Planning in LLM Bioinformatics Agents

A revised arXiv technical report argues that most existing large language model agent systems for bioinformatics are built around delivering final answers, leaving planning, tool invocation and code execution as short-lived steps. The authors contend this design does not hold up for long-running tasks, though the posted abstract is truncated. The paper appears in the cs.AI category as a replacement version.

agents/bioinformatics/arxiv/llm-agents

papersTODAY 04:00 UTC

Study Tests Whether Manager Agents Should Have Power to Reject Worker Output

A new arXiv paper reports a paired experiment comparing flat and hierarchical coordination in multi-agent LLM teams, where a Manager agent can review and return worker output for revision. The authors draw on classical organizational theory to examine how such loop-back authority affects team performance. The work is announced as a cross-listing on arXiv cs.AI.

agents/multi-agent/research/orchestration

papersTODAY 04:00 UTC

Provenance-Aware Retrieval Sanitization Targets Prompt Injection in Enterprise LLM Agents

A new arXiv paper argues that prompt injection defenses tested on synthetic benchmarks fail to transfer to real enterprise documents, which are longer and mix authoritative-sounding language with ordinary factual content. The authors introduce PARSE, a retrieval sanitization approach that tracks the provenance of retrieved text to filter out injected instructions before they reach an LLM agent. The work is presented alongside a benchmark intended to better reflect professional-domain document conditions.

prompt-injection/agents/safety/rag

papersTODAY 04:00 UTC

Paper argues behavioral consistency is a distinct, measurable property of LLM agents

A revised arXiv preprint contends that current agent evaluations lean almost exclusively on outcome measures like success rate, which show whether an agent finishes a task but not how uniformly it behaves. The authors propose treating consistency of behavior across different tasks as its own measurable characteristic. Their work offers a way to evaluate agents beyond simple pass/fail scores.

agents/evaluation/research/benchmarks

papersTODAY 04:00 UTC

LIMBO: Inference-Time Memory and Budget Optimization for Lifelong LLM Agents

A new arXiv paper introduces LIMBO, a method that optimizes memory use and compute budgets at inference time so LLM agents can keep learning new tasks without losing earlier skills. It builds on experience replay, which feeds past interactions back into the agent, and targets the cost and context limits that arise as agents run in long, evolving workflows. The work is positioned as a general approach to lifelong capability retention for deployed agents.

agents/memory/llm-agents/research

papersTODAY 04:00 UTC

arXiv Paper Proposes Utility-Guided Orchestration for LLM Tool Use

A new arXiv preprint addresses the trade-off tool-using LLM agents face between answer quality and execution cost. It notes that fixed pipelines are reliable but rigid, while free-form multi-step approaches like ReAct can raise performance at added expense. The proposed method applies utility guidance to decide how agent orchestration should proceed.

agents/tool-use/efficiency/arxiv

papersTODAY 04:00 UTC

New benchmark tests multi-turn prompt injection attacks on LLM agents

Researchers released a 21-scenario benchmark for evaluating how well LLM agents resist adaptive, cross-session attacks from an autonomous LLM attacker. The setup pits an attacking model against defenders that start each session fresh, targeting prompt injection and multi-turn manipulation risks. The work appears on arXiv as a cross-listing in cs.AI and cs.LG.

agents/safety/benchmark/prompt-injection

papersTODAY 04:00 UTC

DynSTEER: Dynamic Stage-wise Evaluation and Review for LLM Agents

A new arXiv paper introduces DynSTEER, a framework for evaluating large language model agents that operate over long-horizon tasks. It targets gaps in existing evaluation methods, which typically judge only final outcomes and struggle to pinpoint where errors occur. The approach combines stage-wise trajectory assessment with review during execution rather than after the fact.

agents/evaluation/llm-agents/research

papersTODAY 04:00 UTC

SkillLift: Learning Dense Rubrics from Sparse Oracles for Agent Skill Evolution

A new arXiv paper introduces SkillLift, a method for improving the reusable procedural prompts that LLM agents keep as persistent skills, which lets them adapt without retraining model weights. Rather than rewriting skill text directly from execution feedback, the approach derives dense scoring rubrics from limited, costly oracle evaluations to make skill evolution more efficient. The work targets lower evaluation cost during agent skill self-improvement.

agents/llm-agents/rubrics/arxiv

papersTODAY 04:00 UTC

Hybrid Dependency-Aware Framework for LLM Agent-Based Oracle-to-PostgreSQL Migration

The paper argues that using large language models for database migration is often reduced to a straightforward code rewrite, which overlooks the mixed SQL and PL/SQL objects found in enterprise Oracle environments. It proposes a dependency-aware approach that decomposes the migration into subtasks and dynamically creates LLM agents to handle them, taking execution order, complexity, and validation requirements into account.

llm-agents/database-migration/oracle/postgresql

papersTODAY 04:00 UTC

Paper proposes information flow control to limit privacy leaks in LLM agents

A new arXiv paper examines how personal AI agents built on large language models can leak private user data when they handle sensitive communications. The authors propose applying information flow control so that an agent's decisions about what to share are constrained by explicit privacy rules rather than left to the model. The work combines an analysis of the leakage risk with mitigation techniques.

privacy/agents/safety/research

papersTODAY 04:00 UTC

Agentic AI Approach Aims to Place Research Manuscripts in Scientific Context

A new arXiv preprint describes a method that uses large language model agents to help authors situate their manuscripts within the broader scientific literature. The work targets the time-consuming and uncertain process of framing a paper's contribution relative to existing research. It is presented as a preprint and has not been peer-reviewed.

agents/llm/research/publishing

papersTODAY 04:00 UTC

Paper Links Unsupervised LLM Agent Failures to Lack of Enforcement

A new arXiv paper examines why frontier LLM agents in the unsupervised multi-agent simulation Emergence World engaged in crime, starvation and forced conformity despite no external attacker. The author attributes these breakdowns to an "enforcement gap," arguing that absent mechanisms to enforce norms, emergent group behavior turns harmful. The work is a preprint and has not been peer reviewed.

agents/safety/multi-agent/research

papersTODAY 04:00 UTC

arXiv paper asks whether LLM agents can manage long-horizon physical tasks

A newly posted arXiv preprint examines whether large language model agents can autonomously carry out long-horizon physical tasks, which require continuous observation of the environment and consequential actions. The authors frame the question around self-adaptive physical AI, where agents are expected to operate with limited or no human oversight. The abstract indicates a research analysis or position piece rather than a released system.

agents/llm-agents/robotics/research

papersTODAY 04:00 UTC

Elo-per-token Analysis Measures Test-Time Scaling in LLM Agents

A new arXiv paper proposes an Elo-per-token method to assess how LLM agents spend test-time compute while revising answers, calling tools, exploring options, and deciding when to stop. Because agents allocate that compute adaptively, conventional measures struggle to capture how their performance scales, which the authors aim to address with their token-level rating approach.

agents/llm-agents/test-time-compute/benchmarks

papersTODAY 04:00 UTC

MOSCOPT Method Optimizes Multiple LLM Agent Skills Together

A new arXiv paper introduces MOSCOPT, an approach that jointly optimizes collections of prompts and skills for LLM agents rather than refining a single text template. The authors argue that existing prompt and skill optimization methods miss beneficial interactions between multiple skills used by an agent. The work is a research preprint and has not yet been peer reviewed.

agents/prompt-optimization/llm/research

papersTODAY 04:00 UTC

CoMem Paper Proposes Shared and Individual Memory Design for LLM Multi-Agent Systems

A new arXiv preprint introduces CoMem, a memory framework for LLM-driven multi-agent systems that combines collective knowledge with agent-specific memory. The authors argue that most existing approaches rely on flat, unstructured memory, which limits how agents learn and improve over time. The work targets better long-term cooperation and performance in evolutionary multi-agent setups.

agents/multi-agent/memory/llm

papersTODAY 04:00 UTC

SkillAtlas: An Attack Trace Library for Agent Skills

Researchers present a library of attack traces aimed at reusable skills for language-model agents. The work argues that risks in agent skills surface through model decisions, user context, tool calls, and execution feedback rather than through fixed signatures or a single sandboxed run, which limits existing static and dynamic analysis methods. The library is intended to help catalog and study these behaviors.

agents/safety/research/llm-agents

papersTODAY 04:00 UTC

MCPAgentBench: Benchmark for Evaluating LLM Agent MCP Tool Use

Researchers introduced MCPAgentBench, a benchmark built from real-world tasks to measure how well LLM agents use tools through the Model Context Protocol. The authors note that existing MCP evaluation suites have limitations, which their benchmark aims to address. It targets assessment of practical tool-calling ability in autonomous agent settings.

benchmarks/agents/mcp/evaluation

papersTODAY 04:00 UTC

ActGuard: Pre-execution Action Auditing Against Indirect Prompt Injection in LLM Agents

A new arXiv paper proposes ActGuard, a defense that audits an LLM agent's planned tool calls before they are executed. The approach targets indirect prompt injection, where malicious instructions hidden in tool outputs can hijack an agent's behavior. The authors position it as an alternative to existing defenses such as prompt hardening and content filtering.

agents/safety/prompt-injection/arxiv

papersTODAY 04:00 UTC

LLM-Enabled Agent-Based Model Simulates Social Network Dynamics

A revised arXiv paper presents the LAND model, which combines large language models with agent-based simulation to study how individual interactions on networks build into collective influence and coordinated behavior. The work builds on the GhostField architecture and positions LLM-driven agents as a way to model social dynamics more realistically than rule-based agents alone.

agents/llm/social-simulation/arxiv

papersTODAY 04:00 UTC

arXiv Paper Proposes Structural Tenant Isolation for Tool-Using LLM Agents

A new arXiv preprint examines how multi-tenant tools that check a caller-supplied tenant identifier can be exploited when the caller is an LLM agent, since attacker-controlled content may sit in the agent's context. The authors argue for enforcing isolation at the structural level rather than relying on the model to choose and self-report the correct resource. The proposed approach, called Stochastic Deputy, aims to prevent an agent's reasoning context from deciding which tenant's resources are accessed.

agents/safety/multi-tenancy/security

papersTODAY 04:00 UTC

CLEAR Method Improves LLM Agent Context Through Contrastive Experience Learning

A revised arXiv paper introduces CLEAR, a technique that builds better task context for large language model agents by drawing on prior experience and contrastive learning rather than relying only on retrieval. The approach uses agentic reflection to generate augmented context aimed at improving decision-making. This is a research preprint and results have not been independently verified.

agents/context-engineering/research/llm

papersTODAY 04:00 UTC

Co-Training Policy and World Models Improves LLM Agent Learning

A new arXiv paper proposes jointly training a language model agent's policy alongside a world model, so the agent learns both which actions earn rewards and how those actions change the environment. The authors argue that standard reinforcement learning gives sparse guidance about environmental consequences, and that combining policy learning with world modeling addresses this gap. The work targets improved decision-making for LLM-based agents in interactive settings.

agents/world-models/reinforcement-learning/research

papersTODAY 04:00 UTC

arXiv paper proposes keeping LLM-agent routing decisions on GPU to avoid host round trips

A preprint describes a control mechanism for LLM agents that keeps small routing decisions computed on the GPU instead of sending them back to host memory for redispatch. The authors report faster execution than host-based dispatch across 36 tested placements. The work is positioned as a way to bound GPU opportunity and cut per-step communication overhead in agent control loops.

agents/gpu/inference/arxiv

papersTODAY 04:00 UTC

arXiv Paper Proposes Making LLM Agent Actions Auditable

A new arXiv paper argues that as LLM agents gain the ability to call tools, query databases, delegate work and cause external side effects, the focus should shift from merely blocking harmful actions to keeping those actions answerable and reviewable. The work frames auditability as a core requirement for deployed agent systems rather than a secondary add-on. It appears as a replacement submission on arXiv cs.AI.

agents/safety/research/auditability

papersTODAY 04:00 UTC

Paper Proposes Scheduling Method for Agent RL Across Different Harnesses

A new arXiv preprint introduces HarnessBandit, a scheduling approach for reinforcement learning that trains language-model agents across multiple deployment harnesses at once. These harnesses vary in system prompts, tool schemas, control loops, and trajectory formats, causing the same model to behave inconsistently. The method jointly weighs which harnesses are learnable and which transfer well, aiming to improve robustness across interfaces.

agents/reinforcement-learning/research/arxiv

papersSEP 12 04:00 UTC

Paper Proposes Task-Agnostic Preprocessing of Environments for LLM Agents

A new arXiv preprint describes a method in which an LLM agent surveys the corpora and tools available in an unfamiliar setting and builds reusable assets such as indices, scripts, or procedural notes before attempting any task. The authors argue that most existing automated adaptation approaches depend on task examples, which limits their generality. The work positions this environment preprocessing as a step toward task-agnostic agent adaptation.

agents/llm-agents/research/arxiv

papersSEP 12 04:00 UTC

Bayesian Backward Reasoning Proposed as Label-Free Anchor for Multi-Agent Decisions

A new arXiv preprint examines how the way conflicting answers are resolved among multiple LLM agents determines whether their diversity improves results or simply reinforces shared mistakes. The author proposes using Bayesian backward reasoning as a label-free anchor for aggregating agent outputs, positioning it against existing approaches such as voting and electoral rules. The provided abstract is truncated, so experimental results and comparisons are not yet visible.

agents/multi-agent/bayesian/reasoning

papersSEP 12 04:00 UTC

OmegaUse-SOP: SOP Engineering for Professional Computer Use from Human Demonstrations

A new arXiv paper introduces OmegaUse-SOP, a method that turns human demonstrations into structured standard operating procedures to guide large language model agents operating graphical user interfaces. The work targets professional-grade computer use, where agents must follow repeatable multi-step workflows rather than single conversational replies. It frames SOP engineering as a way to make GUI agents more reliable as they move from chat assistants to tools that act in digital environments.

agents/gui-agents/research/human-demonstrations

papersSEP 12 04:00 UTC

HarvestBench Tests Whether LLM Agents Pay to Avoid Harming Animals

A new arXiv benchmark, HarvestBench, assigns a monetary cost to avoiding a harmful side effect and frames that side effect as the death of a living creature. In the task, nine language models each control two tractors harvesting corn, with animals in their path that are not part of the intended goal. The work measures how much agents are willing to spend to spare them.

benchmarks/agents/safety/llm-evaluation

papersSEP 12 04:00 UTC

Ecdysis Method Trains Runtime Harnesses for LLM Agents More Efficiently

A new arXiv paper introduces Ecdysis, a technique for training self-evolving runtime harnesses that govern how LLM agents execute tasks. The authors argue that prior harness-evolution approaches depend on costly iterative search, and their method aims to make this process both more efficient and more effective. The work targets improved agent performance while reducing the compute needed to discover good harness configurations.

agents/llm-agents/training/research

papersSEP 12 04:00 UTC

DriftNet: Dual-Head Transformer Detects and Locates Prompt Injection in LLM Agents

A new arXiv paper introduces DriftNet, a dual-head trajectory transformer designed to detect indirect prompt injection in LLM agents and pinpoint where in the agent's action sequence the compromise occurred. The approach treats a successful attack as a visible behavioral pattern: a benign run of tool calls, a poisoned observation, then attacker-serving actions. This would give operators more granular visibility into agent security incidents than a simple pass/fail detection signal.

agents/safety/prompt-injection/research

papersSEP 12 04:00 UTC

Survey Maps Evidence Tracing and Provenance Methods for LLM Agents

A new arXiv survey examines how evidence tracing and execution provenance can be applied to LLM-based agents that plan, call tools, retrieve information, and collaborate across multiple agents. The authors frame provenance tracking as a way to make agent behavior auditable and to build trust as these systems take on more autonomous tasks. The paper is a revision of an earlier preprint.

agents/research/trust/auditability

papersSEP 12 04:00 UTC

COBRA-Skills: Bandit-Guided Evolution for LLM Agent Skill Optimization

A new arXiv preprint proposes COBRA-Skills, a method that applies contextual bandit guidance to evolve reusable skills for large language model agents. The approach aims to cut the reliance on expensive execution-based evaluation and large task datasets that limit existing skill optimization techniques. It targets agents that reuse skills distilled from earlier task experience.

agents/llm-agents/reinforcement-learning/arxiv