Tag: #long-horizon8 results
Researchers propose a subgoal-driven framework and MiRA RL method to tackle long-horizon planning in LLM agents for web navigation. MiRA elevates Gemma3-12B's WebArena-Lite success rate from 6.4% to 43%, outperforming GPT-4-Turbo (17.6%) and GPT-4o (13.9%). This combo of planning and milestone rewards boosts autonomous AI capabilities.
ResearchGym introduces a benchmark with five containerized environments from ICML, ICLR, and ACL papers, totaling 39 sub-tasks, where agents propose hypotheses and run experiments to beat baselines. GPT-5-powered agents show a capability-reliability gap, succeeding in only 6.7% of evaluations and completing 26.5% of sub-tasks. It identifies key failure modes like impatience and poor resource management, while occasionally achieving SOTA results unreliably.
COSPLAY is a co-evolution framework where an LLM decision agent retrieves skills from a learnable skill bank to guide actions, while a skill pipeline discovers and refines reusable skills from rollouts. It enhances long-horizon decision-making in games under partial observability and delayed rewards. With an 8B model, it achieves 25.1% average reward improvement over baselines on single-player games.
Introduces HORIZON benchmark to systematically diagnose long-horizon failures in LLM agents across domains. Evaluates SOTA models like GPT-5 variants and Claude on 3100+ trajectories, revealing degradation patterns. Releases leaderboard and LLM-as-a-Judge pipeline validated with human annotations (κ=0.84).
Gemma 4 31B secures 3rd place on FoodTruck Bench, outperforming GLM 5, Qwen 3.5 397B, and all Claude Sonnets. It excels in long-horizon tasks by following its own planning advice. Poster excited but clarifies not the benchmark author.
LifeBench introduces a new benchmark for AI agents' long-term memory, integrating declarative and non-declarative types from diverse digital traces. It ensures data quality with real-world priors like social surveys and map APIs, and scales via cognitive-inspired event hierarchies. Top memory systems score only 55.2%, revealing challenges in long-horizon retrieval.
PseudoAct is a new framework that uses LLM-generated pseudocode to create structured plans for agents, encoding control flows like loops, conditionals, and parallelism. It overcomes ReAct's limitations in long-horizon tasks by reducing redundant actions and ensuring coherent execution. Experiments show 20.93% success gain on FEVER and new SOTA on HotpotQA.
Memory Bear AI introduces a memory-centered framework for multimodal affective intelligence, treating emotions as evolving variables in a structured memory system. Multimodal signals are converted into Emotion Memory Units (EMUs) for persistent storage, retrieval, and updating across interactions. It outperforms baselines in accuracy and robustness, especially under noisy or missing data.