Live intelligence

Latest AI updates

1,729 updates

Research Editor selected

When Safety Checks Become Privacy Risks: A Privacy-Preserving Paradigm for LLM Agent Safety Auditing

The paper proposes a privacy-preserving paradigm for safety auditing of LLM agents, separating audit computation from plaintext access and execution authority. It introduces SealCall, a framework combining learned audit representations with CKKS-based encrypted similarity computation, achieving a mean F1 score of 0.840 on 1,252 public binary candidates. Encrypted scoring completes in 0.176 seconds per operation, preserving plaintext audit decisions in all measured cases. This addresses the risk of exposing sensitive user data during external safety checks.

Why it matters: Offers a concrete, evaluated solution to a critical privacy issue in LLM agent auditing, with measurable performance metrics.

arxiv5d agoScore 85

LLM safetyprivacyhomomorphic encryption

Research Editor selected

CryoAtlas: A Large Curated Dataset and Unified Benchmark for Cryo-EM Atomic Model Building

CryoAtlas is a rigorously curated dataset and unified benchmark for density-guided atomic model building of protein chains from cryo-EM maps. It provides 17,809 map-model-sequence entries, including a curated training split, and a 618-target benchmark with a shared evaluation protocol assessing structural accuracy, model-to-map fit, and stereochemical validity. The benchmark is used to evaluate state-of-the-art learning-based methods against Phenix and AlphaFold3, and retraining the ModelAngelo CNN component on CryoAtlas improves performance. This resource offers an open, reproducible foundation for future cryo-EM atomic model building research.

Why it matters: Provides a large, high-quality dataset and standardized benchmark that addresses a critical gap in cryo-EM atomic model building, enabling reproducible evaluation and advancement of deep learning methods.

arxiv5d agoScore 90

cryo-electron microscopyprotein structure predictiondeep learning

Research Editor selected

SparseGDN: Chunk Sparse in Linear Attention

SparseGDN introduces a hardware-aware framework for Gated Delta Networks that reduces computation and memory traffic by combining structured intra-chunk approximation with input-dependent recurrent-event selection. Evaluated on Qwen3.5 models across context lengths up to 128K, it retains about 29% of recurrent events while achieving 1.50-1.94x prefill speedups over native FLA kernels.

Why it matters: Provides a practical method to accelerate long-context sequence modeling with linear attention, offering significant speedups without sacrificing output quality.

arxiv5d agoScore 85

efficient transformerssequence modelinghardware-aware algorithms

Research Editor selected

BlindToChange: AI Agents Fail to Proactively Search for Recent Events

A new benchmark, BlindToChange, reveals that AI agents often fail to proactively search for recent real-world events that affect their responses, even though they can find changes when explicitly asked. The benchmark uses requests where an obvious candidate is invalidated by a recent event, and shows a performance gap of up to 27.5 points when the event is not mentioned. Over half of failures stem from agents not directly verifying relevant facts. This work highlights a new axis for progress in agentic systems.

Why it matters: Provides a concrete, measurable benchmark exposing a critical blind spot in AI agent behavior, directly relevant to improving real-world reliability.

arxiv5d agoScore 85

AI agentsbenchmarkproactive information seeking

Research

From Representation to Readout: Mechanistic Analysis of Membership Signals in Language Models

This arXiv paper presents a mechanistic study of membership inference signals in large language models, introducing Decodable Information Score, Used Information Score, and Readout Accessibility Gap (DUG) to separate membership information in internal representations from information exposed through readout. Across supervised fine-tuning and reinforcement-learning-based membership inference methods, it reports membership-specific effects in both decodable information and readout under SFT, but primarily in readout under RL methods, with activation patching localizing effects to intermediate layers. The paper also examines reinforcement learning with verifiable rewards contamination and finds behavioral changes can emerge without a corresponding increase in likelihood-based membership inference signals.

arxiv5d agoScore 72

membership inferenceLLM interpretabilityreinforcement learning

Research Editor selected

Evidence-Guided Recovery for AI Agents with Ambiguous Tool Outcomes and Partial Execution

A new arXiv paper introduces an evidence-guided framework for AI agents to recover from ambiguous tool outcomes and partial execution. The approach uses approved obligations, provider capabilities, and public evidence to determine safe completions, with a shared admissibility check and terminal certificate. In a study of 37 assignments, the method certified 28 completions, and a preregistered comparison across 58 paired units showed a 36.2-percentage-point advantage over evidence prompting.

Why it matters: Provides a concrete, formal method for improving AI agent reliability in real-world tool-use scenarios, with experimental validation.

arxiv5d agoScore 82

AI agent reliabilitytool outcome ambiguitytask recovery

Research Editor selected

EvoMCP: Co-evolving MCP agents and training data

EvoMCP is a framework that enables Model Context Protocol (MCP) agents to improve through a closed loop of task evolution, trajectory evolution, and supervised fine-tuning. It assesses the student's capabilities from rollout feedback, translates learning targets into executable tasks by adapting seed tasks, and uses reusable skills to recover failed teacher rollouts. Experiments on Toolathlon, MCPMark, and MCP-Atlas show absolute gains of 16.5%, 10.2%, and 12.6% respectively over the initial student.

Why it matters: Introduces a novel agent-data mutual evolution method that adapts training data to the student's evolving capabilities, addressing a key limitation of static supervised fine-tuning for MCP agents.

arxiv5d agoScore 88

Model Context Protocolagent trainingdata evolution

Research Editor selected

KEEPER: Learning Control-Sufficient Textual Memory for Long-Horizon LLM Agents

The paper introduces KEEPER, a framework that analyzes auxiliary objectives for recursive textual memory in LLM agents, identifying self-consistency bias where incorrect claims persist in memory. It compares four supervision targets and finds reward-related prediction achieves the highest task performance before and after offline reinforcement learning. A claim-level analysis reveals error persistence despite corrective observations, motivating evaluation of error persistence and valid-information retention alongside task outcomes.

Why it matters: Provides a concrete, recent research contribution addressing a critical failure mode in long-horizon LLM agent memory, with practical implications for agent reliability.

arxiv5d agoScore 85

LLM agentsmemory managementreinforcement learning

Research

Multi-Test Nation Bias Benchmarking of LLMs with Explicit Unbiased Baselines

Researchers introduce a benchmark using 722 official UN Security Council records and five complementary tests to measure nation-level bias in LLMs, including opinion ranking, association sorting, vote simulation, opposition attribution, and framing. They define explicit unbiased reference statuses for each test, including action-based pseudo-baselines derived from observable voting behavior for tests where ground truth is difficult to specify. Benchmarking GPT-4o-mini, Llama 3.3-70B, Mistral-22B-Small, and Qwen3-30B across the five permanent UNSC members reveals a test-agnostic macro-pattern: outputs lean toward the opposition side for China and Russia and toward the supportive side for the U.K., the U.S., and France. The authors argue that international-relations-facing LLM deployments need auditing with grounded, multi-test benchmarks tied to explicit reference points.

arxiv5d agoScore 76

LLM evaluationnation-level biasUN Security Council

Research Editor selected

Sparse Text Diffusion under Dual Budgets: Context Connectivity and Expert Activation

The paper introduces DualBudget, a sparse text diffusion method that separately budgets attention graph edges and expert activations. It combines local, dilated, and global attention with conditional expert computation. Experiments show gains over baselines up to 16K context length with a 20% cost reduction, though performance degrades at longer lengths.

Why it matters: Provides a principled approach to balancing efficiency and quality in long-context diffusion models.

arxiv5d agoScore 85

sparse text diffusiondual budgetslong-context modeling

Research Editor selected

Specification-Guided Generation for Large Language Models

A new neuro-symbolic decoding framework enforces formal specifications during LLM generation, not just syntax. It offers a forward-only variant guaranteeing specification satisfaction and a backtracking variant that revises earlier choices to avoid forcing unsuitable values. The method is model-agnostic, requires no retraining, and reduces repeated inference compared to rejection sampling on configuration and constraint-solving benchmarks.

Why it matters: Introduces a practical, model-agnostic approach to enforce formal specifications in LLM outputs, improving reliability for structured generation tasks.

arxiv5d agoScore 88

large language modelsconstrained generationneuro-symbolic AI

Research Editor selected

ORION: An Outcome-Driven LLM Advisor for RL-Based Multi-Domain 6G Slice Placement

ORION introduces a hierarchical orchestrator that uses two LLM agents to translate service intents and generate abstract placement plans for 6G network slices across administrative domains. A cooperative multi-agent RL coordinator selects cross-domain partitions while per-domain agents handle function placement. The key innovation is an outcome-driven episodic memory that records network conditions, proposed plans, and verified feasibility outcomes, feeding successes and failures back to the planner. In controlled experiments, this memory improves admission rates by 4.6–11.5 percentage points over the same planner without memory.

Why it matters: Provides a concrete, measurable improvement in 6G network orchestration by combining LLM planning with RL and a novel outcome-driven memory mechanism.

arxiv5d agoScore 88

ORION6G slice placementoutcome-driven memory

Research Editor selected

Verify–Backtrack Loop: Structured Precondition Recovery for Long-Horizon Language Agents

The paper introduces the Verify–Backtrack Loop (VBL), a training-free extension of ReAct that uses a bounded goal stack and dual-horizon verification to diagnose and recover from precondition failures in long-horizon tasks. VBL jointly checks whether an active goal contributes to the terminal goal and whether it is feasible in the current state, then triggers explicit operations like Push, Pop, retry, or BacktrackTo. Evaluated on AppWorld, ALFWorld, and ScienceWorld with Qwen3.5-9B, GPT-4o, and DeepSeek-V3.2, VBL improves task success by up to 22.07 percentage points over ReAct, and ablations confirm that both verification and diagnosis are complementary.

Why it matters: Provides a structured, training-free method for improving failure recovery in language agents, with strong empirical gains across multiple benchmarks and backbones.

arxiv5d agoScore 88

language agentstask decompositionfailure recovery

Research Editor selected

Patch-Based Deep Kernel Learning for Data Assimilation and Bias Correction with Extremely Sparse Observations

The paper proposes a hybrid framework that couples a CNN with a learnable-kernel Gaussian process to jointly perform data assimilation and bias correction from extremely sparse observations. A patch-based strategy reduces exact-GP training cost from O(N^3) to O(N) for a fixed patch size, while inference cost depends on the number of observations rather than the full grid. Experiments on sea surface temperature and salinity with synthetic and real sparse in-situ observations show improvements over 3D-Var, a classical GP, and a competing deep-learning model, with inference time reduced by more than 98% relative to 3D-Var.

Why it matters: It offers a concrete AI method for uncertainty-aware data assimilation and bias correction, with efficiency and accuracy claims relevant to weather and ocean modeling.

arxiv5d agoScore 83

data assimilationGaussian processesdeep learning

Research Editor selected

Pre-Registered Audit of 25 Cognitive-Architecture Modules in an LLM Agent System

The paper introduces a four-question static audit protocol with line-number evidence and call-chain tracing to test whether named cognitive mechanisms in LLM-based agent systems are actually implemented and used. In the audited system claiming 25 cognitive mechanisms, the computed value never reached any production output in 18 modules; the authors identify three idle-wiring patterns: connected-but-never-called, called-but-inert, and called-but-content-blind. Across 61 frozen criteria in 12 interventions, 53 were confirmed, five falsified, two partially adjudicated, and one adjudication-ambiguous. Two repair experiments raised planner task success from 0.028 to 1.000 and cut state-controller error by 52%.

Why it matters: It gives agent developers a reusable audit method and concrete evidence that module names can substantially overstate implemented computation.

arxiv5d agoScore 82

LLM agentscognitive architectureAI evaluation

Research Editor selected

VLA Policies Do Not Know When Not to Act: Unintended Task Execution and Start Suppression

A new arXiv study examines whether vision-language-action policies refrain from acting when instructions are missing, meaningless, infeasible, or refer to the wrong task. Across six checkpoints in LIBERO-Goal simulation, an empty instruction still caused movement in every episode, and five checkpoints often completed a task the instruction had not specified. Three suppression approaches were tested: an instruction prefix failed across all six checkpoints, an action-chunk gate held the robot on non-task inputs in five of six, and fine-tuning prevented task completion on non-task inputs but did not reliably keep the robot still or handle infeasible instructions.

Why it matters: It provides concrete evidence that current VLA policies can start and execute tasks without valid instructions, an important safety issue for embodied AI deployment.

arxiv5d agoScore 82

vision-language-action modelsrobot safetyinstruction following

Research

KARU: Reinforcement Unlearning of LLM Knowledge Associations

Researchers propose Knowledge Association (KA) as the target of LLM unlearning, defined as the conditional predictive distribution linking a query context to its target knowledge. They introduce Knowledge Association Reinforcement Unlearning (KARU), a reinforcement learning framework that combines token-level reward construction, distribution-level regularization, and proximity-aware retain sampling. Experiments across multiple LLMs and benchmarks report improved forgetting effectiveness while maintaining comparable retention and general utility.

arxiv5d agoScore 72

LLM unlearningKnowledge associationsReinforcement learning

Research

DA-CoT Uses Attention-Map Entropy to Decide When Reasoning Segmentation Needs Chain-of-Thought

Researchers propose Difficulty-Aware Chain-of-Thought (DA-CoT), a training-free framework for reasoning segmentation with multimodal large language models. The method uses the entropy of attention maps as a difficulty signal, triggering CoT only when grounding is uncertain to avoid “overthinking.” A Bidirectional Rethinking mechanism lets the model reverse its reasoning decision after an initial failure. The authors report state-of-the-art results on Reasoning Segmentation benchmarks with improved inference efficiency.

arxiv5d agoScore 73

Reasoning SegmentationChain-of-ThoughtMultimodal Large Language Models