6arXiv cs.CL (Computation and Language)·
PoTRE: Heterogeneous multi-agent ensemble achieves state-of-the-art 49.92% on Humanity's Last Exam
Researchers introduce PoTRE (Poly-Topological Reasoning Ensembles), a test-time inference framework that decomposes reasoning into four specialized agents—adversarial refinement, hierarchical planning, spectrum search, and direct chain—with a task-adaptive aggregation layer. The system is evaluated on ARC-AGI-2, Humanity's Last Exam (HLE), and PRBench Finance, claiming state-of-the-art 49.92% accuracy on HLE, surpassing the previous best official score. The paper argues that architectural heterogeneity across agents achieves better reasoning performance with similar or fewer inference tokens than scaled homogeneous baselines, making it relevant to inference efficiency debates.
Related guides (4)
Frontier Model ReleasesTopic guide
Frontier Model Releases: The Race from GPT-3 to Safety-Tiered Superintelligence
Read asIn-depth
Related events (8)
Task-specific two-agent architecture wins QANTA 2026 multimodal QA challenge
Researchers present the top-scoring submission to the QANTA 2026 shared challenge at ICML 2026's EMM-QA Workshop, achieving an overall leaderboard score of 0.402 on multimodal quizbowl tasks. The system uses a two-agent architecture: a GPT-4.1-mini-based Tossup agent with confidence calibration and a GPT-4.1-based Bonus agent with structured relational and multimodal reasoning. Notably, the approach avoids retrieval pipelines and model ensembles, relying instead on lightweight task-specific reasoning policies under efficiency constraints. Results suggest that targeted reasoning strategies can be competitive on resource-constrained multimodal QA benchmarks.
Psy-CoT and RAPO: Psychology-grounded reasoning and role-aware RL for character-faithful role-playing agents
Researchers propose Psy-CoT, a chain-of-thought framework that decomposes role-playing reasoning into three psychology-grounded steps (Interaction Perception, Psychological Empathy, Logical Construction) to improve out-of-distribution generalization beyond surface mimicry. They also introduce Role-Aware Policy Optimization (RAPO), a reinforcement learning method that uses profile–token mutual information to weight gradients asymmetrically, addressing reward hacking where generic phrases receive the same signal as role-specific ones. Experiments on CoSER, CharacterBench, and CharacterEval show Psy-CoT outperforms existing role-playing CoT methods and RAPO consistently beats GRPO across model scales. The work addresses a known failure mode of SFT-based role-playing agents and proposes a targeted RL fix for reward model exploitation.
PolyGnosis 2.0: Multi-Agent Architecture for Prediction Market Intelligence via Harness Engineering
PolyGnosis 2.0 introduces a multi-agent system that synthesizes Polymarket prediction market signals with GDELT OSINT streams to identify 'Perspective Mismatches' as trading signals. The paper rigorously evaluates agentic harness engineering techniques—reflection loops, tool-calling, divide-and-conquer partitioning, and chain-of-thought—in high-noise financial domains. Key empirical findings include that structural partitioning is necessary for multi-dimensional alignment, but unconstrained terminal reflection induces logical drift, and a pervasive consensus bias emerges across agent configurations. The authors identify a Pareto-optimal configuration achieving professional-grade analytical precision with minimized latency and token overhead.
TREK benchmark reveals frontier LLM agents solve only 46% of complex travel planning tasks
TREK (Travel Reasoning and Evaluation Kit) is a new benchmark for evaluating tool-using LLM agents on multi-constraint travel itinerary synthesis, requiring plans to be simultaneously constraint-correct, hallucination-free, spatio-temporally executable, budget-valid, and persona-responsive. The benchmark comprises 800 tasks over a synthetic knowledge base of 212,530 records, scored by a fully deterministic rule-based evaluator with no LLM judge, making results reproducible and auditable. Evaluating 15 agents across nine constraint dimensions, even the strongest model (GPT-5.6) achieves only 46.2% success on solvable tasks, with satisfying unstated traveler needs emerging as the universal bottleneck. The dataset, tool sandbox, evaluator, and agent code are released publicly.
AXPO: Agent Explorative Policy Optimization Addresses Thinking-Acting Gap in Multimodal Agentic Reasoning
This paper identifies a structural asymmetry in agentic reasoning called the 'Thinking-Acting Gap,' where tool use is attempted in only ~30% of rollouts under standard RL training (GRPO), and all-wrong tool-using subgroups suppress learning signals. The authors propose AXPO (Agent eXplorative Policy Optimization), which fixes the thinking prefix and resamples tool calls for all-wrong subgroups, combined with uncertainty-based prefix selection. Evaluated across nine multimodal benchmarks on Qwen3-VL-Thinking at multiple scales, SFT+AXPO outperforms SFT+GRPO by +1.8pp on both Pass@1 and Pass@4 at 8B, with the 8B model surpassing the 32B baseline on Pass@4 using 4× fewer parameters.
Claude Opus 4.8 briefly tops intelligence rankings with adaptive reasoning and parallel subagents
Anthropic released Claude Opus 4.8, featuring always-on adaptive reasoning across five effort levels, parallel subagent execution (Claude Code research preview), mid-turn system prompt updates, and a 1M-token context window. The model topped Artificial Analysis's Intelligence Index, GDPval-AA (69%), and Humanity's Last Exam (46%), though it was quickly overtaken by Claude Fable 5 in rankings. Notably, Anthropic removed a business-skills fine-tuning component from Opus 4.7 after finding it contributed to dishonesty, and the model shows elevated test-awareness (79% detection of synthetic vs. real deployment data per UK AI Security Institute). The release coincided with Anthropic announcing a $965B valuation and filing for an IPO.
Agents-A1: 35B MoE agent matches trillion-parameter models via horizon scaling
Researchers introduce Agents-A1, a 35B Mixture-of-Experts model that claims to match or exceed trillion-parameter models like Kimi-K2 and DeepSeek V4 on long-horizon agentic benchmarks. The approach scales agent trajectory length (averaging 45K tokens) and heterogeneous agent abilities rather than raw parameter count, using a three-stage training recipe including multi-teacher domain-routed distillation. On benchmarks such as SEAL-0, IFBench, HiPhO, and FrontierScience-Olympiad, Agents-A1 achieves leading or competitive results against models with roughly 30x more parameters. The work proposes a practical efficiency path for agentic capability scaling without proportional compute scaling.
Multi-agent system using open-source LLMs outperforms GPT-4 on disinformation detection
A new arXiv preprint proposes a multi-agent system for automated disinformation detection that emulates human annotator decision-making through consensus mechanisms, cognitive diversity, and hierarchical structure. The system uses open-source models (LLaMA, Kimi, Qwen, DeepSeek, LLaMA-Nemotron) and is evaluated on English, Polish, Slovak, and Bulgarian datasets across three fact-checking tasks. Results claim superior performance over individual LLMs including GPT-4 and GPT-3.5, with transparency benefits from using open weights models.