CyberArena Benchmark Pits AI Agents in Live Attack-and-Defense Cyber Combat
A new benchmark moves beyond static cybersecurity tests by staging live, symmetric contests between AI agents.
Research Digest
Preprints and papers, rewritten in plain English.
MemComposer: Reinforcement Learning for LLM Agent Memory Assembly
A new paper argues that semantic similarity alone is not enough for LLM agent memory, proposing reinforcement learning to assemble memories for better downstream performance.
New benchmark tests AI agents on long-horizon bioinformatics tasks
OmicsBench evaluates AI agents across multi-turn, research-scale bioinformatics analyses, addressing a gap in existing benchmarks.
New Benchmark Probes Human-Agent Collaboration in Long-Horizon Tasks
A benchmark called Kolab aims to quantify how AI agents contribute when humans work with them on extended, multi-step workflows.
LLM Planners Differ in Commitment Length Under Partial Observability
A new arXiv study examines how LLM-based planning agents decide how much to commit before acting in partially observable environments.
Probe-and-Resume Improves Agent Decisions via Test-Time Search
A new technique lets terminal-based agents explore alternative actions and resume from promising branches, boosting performance.
AgentEval assesses stateful agents by evidence validity, not just task completion
A new arXiv paper argues that task completion and recorded judgments are insufficient for evaluating stateful agents, proposing evidence-bounded assessment instead.
Condition-Contrastive Extrapolation for On-Policy Distillation
A new arXiv paper proposes a method to address a teacher-model failure mode in on-policy distillation of language models.
Latent chain-of-thought: do hidden reasoning steps actually reason?
A new taxonomy aims to clarify what computation happens in latent chain-of-thought models.
MemScope Evolves Agent Memory Architectures via Localized Code Repair
A new arXiv paper proposes MemScope, a method for updating LLM agent memory systems through targeted code repair rather than leaving them fixed after deployment.
GRADS: Graph-Regulated Data Synthesis for Research Agents
A new framework uses graph structure to generate synthetic training data with controllable difficulty, targeting multi-hop reasoning in deep research agents.
Test-Time Method Aims to Fix Generative Models' Diversity Gap
A new approach called MINE-SAMPLE-RENDER tackles the tendency of generative models to produce repetitive outputs despite user instructions.
Who May Cancel an AI Agent's Promise? A New 'Standing' Principle
A new arXiv paper proposes that only the creditor or authorized parties should be able to release an AI assistant's commitments in multi-party chats.
Conformal Weighting Adds Uncertainty Estimates to Rubric-Based LLM Evaluation
A new approach accounts for varying confidence in LLM verifier judgments, making rubric-based evaluation more reliable.
AutoOnboarder Aims to Restore Reasoning in Harnessed LLM Agents
A new arXiv paper proposes AutoOnboarder to address the documented loss of reasoning capability in agentic LLMs when placed inside a software harness.
Diviner: Agentic root-cause analysis for large-scale systems
A new system applies agentic AI with human oversight to root-cause analysis in complex distributed infrastructure.
DocPilot: Training Agents to Navigate Long Visual Documents
A new arXiv paper introduces DocPilot, an interactive agent framework for multimodal question answering across long documents.
Detecting Money-Mule Accounts Demands Relational Analysis
A new study on an Indian financial crime dataset shows that identifying money-mule accounts depends on network context rather than individual transactions.