Search: “llm”

Dimension-Free Saddle-Point Escape in Muon

arXiv:2605.09331v1 Announce Type: new Abstract: Modern Large Language Model (LLM) training is fundamentally bottlenecked by pathologically flat saddle points in extreme high-dimensional landscapes. Mo

researcharxiv-cs-lg

Do Agents Need to Plan Step-by-Step? Rethinking Planning Horizon in Data-Centric Tool Calling

Content type

All Blog X Post Paper YouTube Reddit GitHub

Clear filters

DGX agent

arXiv:2605.08477v1 Announce Type: new Abstract: Explicit planning is a critical capability for LLM-based agents solving complex data-centric tasks, which require precise tool calling over external dat

tutorialsarxiv-cs-cl

12 May 2026

EvoMAS: Learning Execution-Time Workflows for Multi-Agent Systems

arXiv:2605.08769v1 Announce Type: new Abstract: Large language model (LLM)-based multi-agent systems have shown strong potential on complex tasks through agent specialization, tool use, and collaborat

expo: Exploration-prioritized policy optimization via adaptive kl regulation and gaussian curriculum sampling

arXiv:2605.09923v1 Announce Type: new Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has become the standard paradigm for LLM mathematical reasoning, where Group Relative Policy Optim

Hardware

FlashSVD v1.5: Making Low-Rank Transformers Inference Actually Fast

arXiv:2605.08314v1 Announce Type: cross Abstract: SVD-based Low-rank compression reduces transformer parameters and nominal FLOPs, but these savings often translate poorly into real LLM serving speedu

hardwarearxiv-cs-ai

How You Begin is How You Reason: Driving Exploration in RLVR via Prefix-Tuned Priors

arXiv:2605.08817v1 Announce Type: new Abstract: Reinforcement learning with verifiable rewards (RLVR) recently thrives in large language model (LLM) reasoning tasks. However, the reward sparsity and t

Identified-Set Geometry of Distributional Model Extraction under Top-K Censored API Access

arXiv:2605.10407v1 Announce Type: new Abstract: Modern LLM APIs often reveal only top-K logit scores and censor the remaining vocabulary. We study the per-position distribution-recovery limits of this

researcharxiv-cs-lg

IntroLM: Introspective Language Models via Prefilling-Time Self-Evaluation

arXiv:2601.03511v2 Announce Type: replace-cross Abstract: A major challenge for the operation of large language models (LLMs) is how to predict whether a specific LLM will produce sufficiently high-qu

tutorialsarxiv-cs-ai

Local Ai

Large Language Models over Networks: Collaborative Intelligence under Resource Constraints

arXiv:2605.08626v1 Announce Type: cross Abstract: Large language models (LLMs) are transforming society, powering applications from smartphone assistants to autonomous driving. Yet cloud-based LLM ser

local-aiarxiv-cs-lg

LASSA Architecture-Based Autonomous Fault-Tolerant Control of Unmanned Underwater Vehicles

arXiv:2605.09494v1 Announce Type: cross Abstract: Unmanned underwater vehicles (UUVs) operate persistently in communication-constrained environments, thus requiring high-level autonomous fault-toleran

Less Diverse, Less Safe: The Indirect But Pervasive Risk of Test-Time Scaling in Large Language Models

arXiv:2510.08592v3 Announce Type: replace-cross Abstract: Test-Time Scaling (TTS) improves LLM reasoning by exploring multiple candidate responses and then operating over this set to find the best out

Machine Learning Research Has Outpaced Its Communication Norms and NeurIPS Should Act

arXiv:2605.08889v1 Announce Type: cross Abstract: Machine learning research has grown exponentially while its communication norms have not. We argue NeurIPS should adopt explicit, measurable writing s

researcharxiv-cs-cl

MemPrivacy: Privacy-Preserving Personalized Memory Management for Edge-Cloud Agents

arXiv:2605.09530v1 Announce Type: cross Abstract: As LLM-powered agents are increasingly deployed in edge-cloud environments, personalized memory has become a key enabler of long-term adaptation and u

MemQ: Integrating Q-Learning into Self-Evolving Memory Agents over Provenance DAGs

arXiv:2605.08374v1 Announce Type: new Abstract: Episodic memory allows LLM agents to accumulate and retrieve experience, but current methods treat each memory independently, i.e., evaluating retrieval

MIND-Skill: Quality-Guaranteed Skill Generation via Multi-Agent Induction and Deduction

arXiv:2605.08670v1 Announce Type: new Abstract: Large language model (LLM) powered AI agents have emerged as a promising paradigm for autonomous problem-solving, yet they continue to struggle with com

NanoResearch: Co-Evolving Skills, Memory, and Policy for Personalized Research Automation

arXiv:2605.10813v1 Announce Type: new Abstract: LLM-powered multi-agent systems can now automate the full research pipeline from ideation to paper writing, but a fundamental question remains: automati

Applications

On the Overscaling Curse of Parallel Thinking: System Efficacy Contradicts Sample Efficiency

arXiv:2601.21619v2 Announce Type: replace-cross Abstract: Parallel thinking improves LLM reasoning through multi-path sampling and aggregation. In standard evaluations, due to a lack of sample-specifi

applicationsarxiv-cs-ai

PATRA: Pattern-Aware Alignment and Balanced Reasoning for Time Series Question Answering

arXiv:2602.23161v2 Announce Type: replace Abstract: Time series reasoning demands both the perception of complex dynamics and logical depth. However, existing LLM-based approaches exhibit two limitati

PiCA: Pivot-Based Credit Assignment for Search Agentic Reinforcement Learning

arXiv:2605.09287v1 Announce Type: new Abstract: Large Language Model (LLM)-based search agents trained with reinforcement learning (RL) have significantly improved the performance of knowledge-intensi

Position: AI Security Policy Should Target Systems, Not Models

arXiv:2605.09504v1 Announce Type: cross Abstract: We present swarm-attack, an open-source adversarial testing framework in which multiple lightweight LLM agents coordinate through shared memory, paral

PruneTIR: Inference-Time Tool Call Pruning for Effective yet Efficient Tool-Integrated Reasoning

arXiv:2605.09931v1 Announce Type: cross Abstract: Tool-integrated reasoning (TIR) enables large language models (LLMs) to enhance their capabilities by interacting with external tools, such as code in

tutorialsarxiv-cs-ai

Reasoning-Aware Training for Time Series Forecasting

arXiv:2605.08625v1 Announce Type: cross Abstract: Time Series Foundation Models (TSFMs) excel at numerical forecasting but operate as black boxes lacking qualitative reasoning. Conversely, applying LL

Reasoning Trajectories for Socratic Debugging of Student Code: From Misconceptions to Contradictions and Updated Beliefs

arXiv:2511.00371v2 Announce Type: replace Abstract: In Socratic debugging, instructors guide students towards identifying and fixing a bug on their own, instead of providing the bug fix directly. Most

tutorialsarxiv-cs-cl

ShadowMerge: A Novel Poisoning Attack on Graph-Based Agent Memory via Relation-Channel Conflicts

arXiv:2605.09033v1 Announce Type: cross Abstract: Graph-based agent memory is increasingly used in LLM agents to support structured long-term recall and multi-hop reasoning, but it also creates a new

Skill-CMIB: Multimodal Agent Skill for Consistent Action via Conditional Multimodal Information Bottleneck

arXiv:2605.08526v1 Announce Type: new Abstract: While LLM-based agents excel at planning and executing long action sequences, their execution often remains inconsistent across trials, limiting reliabi

agentsarxiv-cs-lg

Skill-R1: Agent Skill Evolution via Reinforcement Learning

arXiv:2605.09359v1 Announce Type: cross Abstract: Agentic large language models often rely on skills, reusable natural language procedures that guide planning, action, and tool use. In practice, skill

SkillRAE: Agent Skill-Based Context Compilation for Retrieval-Augmented Execution

arXiv:2605.10114v1 Announce Type: new Abstract: Large Language Model (LLM)-based agents (e.g., OpenClaw) increasingly rely on reusable skill libraries to solve artifact-rich tasks such as document-cen

agentsarxiv-cs-cl

SLAM: Structural Linguistic Activation Marking for Language Models

arXiv:2605.05443v2 Announce Type: replace-cross Abstract: LLM watermarks must be detectable without compromising text quality, yet most existing schemes bias the next-token distribution and pay for de

Test-Time Speculation

arXiv:2605.09329v1 Announce Type: new Abstract: Speculative decoding accelerates LLM inference by using a fast draft model to generate tokens and a more accurate target model to verify them. Its perfo

The Agent Use of Agent Beings: Agent Cybernetics Is the Missing Science of Foundation Agents

arXiv:2605.10754v1 Announce Type: new Abstract: LLM-based foundation agents that perceive, reason, and act across thousands of reasoning steps are rapidly becoming the dominant paradigm for deploying

The Bystander Effect in Multi-Agent Reasoning: Quantifying Cognitive Loafing in Collaborative Interactions

arXiv:2605.10698v1 Announce Type: cross Abstract: Multi-agent systems (MAS) assume that collaborating inherently improves Large Language Model (LLM) reasoning. We challenge this by demonstrating that

The Differences Between Direct Alignment Algorithms are a Blur

model-releasesarxiv-cs-lg

arXiv:2502.01237v3 Announce Type: replace Abstract: Direct Alignment Algorithms (DAAs) simplify LLM alignment by directly optimizing policies, bypassing reward modeling and RL. While DAAs differ in th

The Extrapolation Cliff in On-Policy Distillation of Near-Deterministic Structured Outputs

arXiv:2605.08737v1 Announce Type: cross Abstract: On-policy distillation (OPD) is widely used for LLM post-training. When pushed with a reward-extrapolation coefficient lambda > 1, the student can lif

When Does Value-Aware KV Eviction Help? A Fixed-Contract Diagnostic for Non-Monotone Cache Compression

arXiv:2605.08234v1 Announce Type: cross Abstract: Long-context LLM inference is bottlenecked by the memory and bandwidth cost of reading large KV caches during decoding. KV compression reduces this co

When Prompts Become Payloads: A Framework for Mitigating SQL Injection Attacks in Large Language Model-Driven Applications

arXiv:2605.10176v1 Announce Type: cross Abstract: Natural language interfaces to structured databases are becoming increasingly common, largely due to advances in large language models (LLMs) that ena

Ah yes, chatGPT, tell me more about the 'colony mind', 'foraging patrol' and 'nost building' behaviors of the human liver. For anyone predic…

Ah yes, chatGPT, tell me more about the 'colony mind', 'foraging patrol' and 'nost building' behaviors of the human liver. For anyone predicting near-term physician replacement by LLM-based AI, please

safetygary-marcus--x

An Interpretable and Scalable Framework for Evaluating Large Language Models

arXiv:2605.07046v1 Announce Type: cross Abstract: Evaluation of large language models (LLMs) is increasingly critical, yet standard benchmarking methods rely on average accuracy, overlooking both the

Belief Memory: Agent Memory Under Partial Observability

arXiv:2605.05583v2 Announce Type: replace Abstract: LLM agents that operate over long context depend on external memory to accumulate knowledge over time. However, existing methods typically store eac

CASCADE: Case-Based Continual Adaptation for Large Language Models During Deployment

arXiv:2605.06702v1 Announce Type: new Abstract: Large language models (LLMs) have become a central foundation of modern artificial intelligence, yet their lifecycle remains constrained by a rigid sepa

Efficient Data Selection for Multimodal Models via Incremental Optimization Utility

arXiv:2605.07488v1 Announce Type: new Abstract: The scaling of Large Multimodal Models (LMMs) is constrained by the quality-quantity trade-off inherent in synthetic data. Previous approaches, such as

Experience Sharing in Mutual Reinforcement Learning for Heterogeneous Language Models

arXiv:2605.07244v1 Announce Type: cross Abstract: We introduce Mutual Reinforcement Learning, a framework for concurrent RL post-training in which heterogeneous LLM policies exchange typed experience

GASim: A Graph-Accelerated Hybrid Framework for Social Simulation

arXiv:2605.07692v1 Announce Type: new Abstract: Large-scale social simulators are essential for studying complex social patterns. Prior work explores hybrid methods to scale up simulations, combining

Implicit Compression Regularization: Concise Reasoning via Internal Shorter Distributions in RL Post-Training

arXiv:2605.07316v1 Announce Type: new Abstract: Reinforcement learning with verifiable rewards improves LLM reasoning but often induces overthinking, where models generate unnecessarily long reasoning

IntentGrasp: A Comprehensive Benchmark for Intent Understanding

arXiv:2605.06832v1 Announce Type: cross Abstract: Accurately understanding the intent behind speech, conversation, and writing is crucial to the development of helpful Large Language Model (LLM) assis

Memory-Efficient Looped Transformer: Decoupling Compute from Memory in Looped Language Models

arXiv:2605.07721v1 Announce Type: cross Abstract: Recurrent LLM architectures have emerged as a promising approach for improving reasoning, as they enable multi-step computation in the embedding space

MIPIAD: Multilingual Indirect Prompt Injection Attack Defense with Qwen -- TF-IDF Hybrid and Meta-Ensemble Learning

arXiv:2605.07269v1 Announce Type: new Abstract: Indirect prompt injection remains a persistent weakness in retrieval-augmented and tool-using LLM systems, and the problem becomes harder to characteris

Our internal AI agent was a brute-force nightmare. 🛑 It burned 40,000 tokens, took 2 minutes, and only hit 68% accuracy just to answer a si…

Our internal AI agent was a brute-force nightmare. 🛑 It burned 40,000 tokens, took 2 minutes, and only hit 68% accuracy just to answer a single question. Why? Traditional databases are built for human

agentspinecone--x

Semantic Integrity Matters: Benchmarking and Preserving High-Density Reasoning in KV Cache Compression

arXiv:2502.01941v3 Announce Type: replace-cross Abstract: While Key-Value (KV) cache compression is essential for efficient LLM inference, current evaluations disproportionately focus on sparse retrie