New Research Shows Claude Opus 4.6 Generates 30× More Errors Than GPT-5.4 Despite Similar Success Rates While RobustSGPO Improves Agent Harness Evolution Completion From 60% To 80%

Frontier models show divergent error rates despite similar success; OpenDiscoveryTrace reveals Claude Opus 4.6 generates 30× more errors than GPT-5.4. Decision-Focused Active Learning optimizes NdFeB magnet recovery in 16–24 experiments versus 48 for nonadaptive methods. State-Path Tool Menus boost ToolBench success from 0.737 to 0.898, while RobustSGPO raises held-out task completion from 60% to 80%. TrimSFT improves mathematical reasoning by +26.9 points on MATH500 via logit-gap reweighting. Autonomous GeoAI agents integrate ecological criteria, and Multi-Agent Agentic Graph Learning partitions graphs better than SOTA.

Safety and reliability research exposes critical gaps: a black-box audit found 85% behavior vulnerabilities in CrewAI and AutoGen. Trust evaluation frameworks show top models score 95.1 on task completion but only 22.6 on safety. LexAgentHallu profiles cascading legal hallucinations invisible to outcome metrics, while RESCUE-BENCH demonstrates LLM failures in relation-sensitive multi-party support. CareGuard detects cyberbullying efficiently, and OntologyAligner achieves 88.78% accuracy in biomedical normalization. Proof-carrying cognition proposals aim to close verification gaps, though unsound verifiers show soundness loss.

Long-horizon tasks benefit from subagents with clear contracts, outperforming context-loading skills despite overhead. Adaptive entangled game modules explain 82–94% of trading patterns, supporting nonlocal brain hypotheses. Cyber-financial contagion models reveal AI vendor compromises propagate through banking networks with AUROC 0.82 early warnings. Memory management improves via storage-usage separation, boosting personalization scores by 2.4–4.0 points and reducing latency by 15–61%. CityPlanner uses sandbox agents and atomic-task RL to solve urban planning, surpassing baselines.

New architectures address agent limitations: a four-layer Cognitive Digital Twin enables self-evolving operations, and a Belief-State Engine augments LLMs with Bayesian posteriors for partial observability. ConvMem uses hierarchical convolution for efficient long-context reasoning without RL overfitting. Fortunate Recall applies behavioral ontologies to manage memory lifecycles, reducing confabulation by half. SmartWeatherAgent fuses rules with LLMs for weather alerts, boosting warning quality by 112%. TRACE uses synthesized rewards to improve diagnostic reasoning, and Valerant generates navigable 3D game maps using action-conditioned world models.