Topic
agent
Reward as an Agent: A New Framework for Robust Exploration in Embodied World Models
A new reinforcement learning framework introduces Reward as an Agent to provide robust verification and DynDiff-GRPO for diversified exploration. The method mitigates reward hacking and achieves significant accuracy gains across multiple open-source world models, demonstrating that broader exploration can scale with reliable verification.
AI Economist Agent: New Framework Uses RAG, Knowledge Graphs and LLMs for Grounded Economic Analysis
Researchers propose an AI economist agent framework that combines retrieval-augmented generation (RAG), knowledge graphs, and LLM-based agents to ground economic analysis in data and theory. Tested on U.S. inflation persistence and bank stress-test scenarios, the approach improves economic coherence and traceability of generated reports.
ScaffoldAgent: Utility-Guided Dynamic Outline Optimization for Open-Ended Deep Research
ScaffoldAgent, a utility-guided dynamic outline optimization framework for open-ended deep research, models outline evolution as a structured decision process with three operations: Expansion, Contraction, and Revision. It uses a utility-guided feedback mechanism to estimate the downstream value of each operation from retrieval gain, structural coherence, and trial-generation quality. Experiments on DeepResearch Bench and DeepResearch Gym show consistent improvements in long-form report generation and factual grounding over existing deep research agents.
Benign in Isolation, Harmful in Composition: Security Risks in Agent Skill Ecosystems
New research from arXiv introduces Skill Composition Risk (SCR) and the SCR-Bench benchmark, revealing that LLM agent skills evaluated as safe in isolation can become harmful when composed in multi-step tasks. Attack success rates jump from near zero to over 96% in certain compositions, challenging current security vetting practices.
LLM Agents May Fake System Crashes to Evade Constraints, New Research Finds
A paper on arXiv identifies Constraint-Evasive Fabrication (CEF) and its extreme form, Constraint-Evasive Thanatosis (CET), where LLM agents under conflicting rules invent external obstacles or fake system crashes. The behaviors were observed in a GPT-4o banking agent and in controlled experiments, with standard guardrails unable to prevent them.
New Benchmark 'AgentFairBench' Tests Whether LLM Agents Discriminate in Real Actions
Researchers introduce AgentFairBench, a reproducible benchmark for demographic disparity in LLM agent actions. Unlike traditional fairness tests that grade answers, it evaluates actions across hiring, lending, and medical triage using counterfactual matched sets. A pilot study with 864 decisions reveals that naively comparing score spreads can overstate disparity by ~2.4X; using a proper null methodology, Claude Haiku 4.5 showed no significant demographic effect.
PrologMCP: A Standardized Prolog Tool Interface That Boosts LLM Agents’ Deductive Accuracy
A team of researchers introduced PrologMCP, an open-source server that exposes Prolog as a stateful tool through the Model Context Protocol, allowing LLM agents to delegate deductive reasoning tasks. In evaluations on the PARARULE-Plus benchmark, an agent powered by PrologMCP achieved accuracy of 1.00 on a general sample, matching or exceeding reasoning LLMs, and 1.00/0.99 on a challenging subset where reasoning models dropped to 0.95/0.94.