Articles & Guides

OR-License evaluates whether LLM agents can turn original client materials into deployable operations research solutions. Its results show that strong conditional optimization performance does not yet translate into expert-level end-to-end reliability.

CyberArena moves cybersecurity evaluation beyond fixed targets by placing AI agents in repeated, symmetric attack-and-defense matches. Across 466 games and 40 challenges, live defenders exposed capability gaps hidden by static tests.

A new study treats scientific ideation as a set-level optimization problem rather than a search for one strong answer. Its 8B model improves the balance between idea quality and diversity without an apparent loss in executability.

A new study introduces EC2-Bench to evaluate edge-cloud collaboration for long-horizon LLM agents. Its results suggest that stronger edge models favor a shift from cloud execution toward edge-led decisions with the cloud acting as an advisor.

HERA co-evolves agent harnesses and execution environments so agents can distinguish solvable tasks from tasks with no valid solution. The approach improves abstention without sacrificing feasible-task completion and transfers across different language models.

Joint-Embedding Predictive Architectures are evolving from representation learners into latent world models for prediction, planning, and control. A new perspective argues that interpretability must examine how latent variables interact, change over time, and causally influence model behavior.

HandsOn is a composable simulation platform and benchmark for evaluating dexterous manipulation across robot embodiments, objects, relations, and task demands. Its results show that reliable contact and task transfer remain difficult even for modern multi-task policies.

CoPE adds multi-step action forecasting to agentic reinforcement learning, allowing successful trajectories to provide supervision before the terminal reward. Evaluations across five benchmarks show stronger aggregate results than GRPO and related predictive-auxiliary baselines, with sample efficiency reaching up to roughly 2× that of some baselines.

StopLoss is a small external monitor trained on completed coding-agent trajectories and verifier outcomes. It predicts likely failures from recent tool steps and ends expensive runs before they waste more resources.

A study of guardrail classifiers finds that safety scores can depend heavily on how training data defines an unsafe request. Small parameter changes, benchmark contamination, and teacher-model soft labels may all shift results without a comparable change in detection ability.

Long-horizon agents leave evidence across both environment state and execution traces, making simple rubric grading unreliable. Sauron investigates those sources first and independently verifies claims before issuing a verdict.

RoboValue is a fine-grained benchmark for evaluating robotic value models across simulation and real-world manipulation. It examines whether model feedback remains reliable throughout execution, not just at the final outcome.

Long-horizon agents can perform individual steps correctly while still producing incomplete outcomes. Interlock addresses this gap by representing task dependencies explicitly and using verified execution trajectories for training.

MemT-1.0 explores a foundation model architecture that treats memory as part of computation rather than relying solely on long prompts or external retrieval. The reported results point to better long-history efficiency while preserving broad model capabilities.

LIBERO-Industry extends the LIBERO benchmark with production-line conditions such as changing product geometry, moving conveyors, limited interaction windows, and tight placement tolerances. Tests show that generalist robot policies degrade steadily as these constraints are introduced and combined.

Stanford researchers introduce Halcyon, an early-fusion multimodal autoregressive model trained from scratch on 100 million image-text pairs. Their results suggest that internet-scale language pretraining is not essential for strong text-to-image generation.

A Tsinghua University team introduces Stateful Harness, a framework that separates hidden-state evolution from video rendering in interactive world models. On WorldState-Bench, it raises paired-outcome accuracy to 92% with the same frozen video model.

Streaming world models often drift when their own predictions are fed back over long rollouts. Frechet Forcing uses self-generated histories and statistics from real clips instead of a teacher model or long videos to preserve realism, scene identity, and action following.

AffordanceWAM introduces an affordance-aware joint world-action model that learns objects, interaction regions, and scene evolution from human videos before generating robot actions from limited robot trajectories. It avoids human action labels and motion retargeting.

Self-HPO is a recursive framework that alternates harness search with reinforcement-learning updates to a coding policy. On three coding benchmarks, it raises Qwen3.5-35B-A3B’s average resolve rate from 57.8% to 64.8%.

OpenVLHarness is a training-free harness that connects multimodal models to perception tools, web search, and Python computation. Its persistent multimodal memory keeps tool-derived evidence available throughout a reasoning process.