long-horizon-agents

Presents a verification instrument for long-horizon agents that structurally separates commitment drift from binding drift, using a deterministic executive and pre-registered predictions. Reports ablation results showing commitment mechanism removal flips goal abandonment from 0 to 1 while binding error stays flat, though task efficacy is null on ARC-AGI-3.

A new survey from Renmin University reviews nearly 1,000 studies on long-horizon AI agents, arguing that reliable long-horizon intelligence depends on the whole model-harness system, not just larger context windows or stronger models.

This paper introduces JarvisBench, a benchmark for evaluating a continuous, real-time spoken mediation layer in long-horizon AI agent workflows, and presents a modular Jarvis prototype tested on WildClaw tasks with various LLM-based worker agents.

TRACE introduces a turn-level reward assignment method using frozen reference model log-probabilities and temporal-difference learning to address credit assignment in long-horizon agentic tasks, achieving significant improvements in search benchmarks without critic or process labels.

TRACE is a dense credit assignment method for multi-turn agentic reinforcement learning that uses a frozen reference model to compute per-action rewards from log-probability changes at tool boundaries, eliminating the need for a critic or process reward model. It significantly improves long-horizon tool-use performance on benchmarks like BrowseComp-Plus.

This paper introduces a proactive memory agent that operates alongside a standard action agent to selectively inject memory-grounded reminders during long-horizon tasks, mitigating behavioral state decay. Experiments on Terminal-Bench and τ²-Bench show significant improvements in pass@1, and the approach is demonstrated with both weak and strong action agents.

Introduces STRACE, a framework that performs structural trajectory analysis and causal extraction to construct high signal-to-noise optimization contexts for improving long-horizon agents, outperforming baselines on a formal verification task.

The paper introduces EnvProbe, a budgeted environment probing operator that allows long-horizon language agents to selectively query the environment for specific belief fields before acting, reducing world-model error by efficiently calibrating beliefs with limited interactions.

Introduces AgentOdyssey, a procedural text game generation framework designed to evaluate agents on test-time continual learning abilities including exploration, episodic memory, world knowledge acquisition, skill learning, and long-horizon planning. The framework highlights significant gaps between current agents and human performance.

Introduces Context Window Lifecycle (CWL), a structured context eviction scheme for long-horizon LLM agents that maintains an effectively unbounded working horizon by evicting content based on a dependency graph, avoiding the limitations of summarization-based compaction and recency truncation.

AutoLab is a new benchmark evaluating 17 frontier models on 36 expert-curated long-horizon tasks (system optimization, model development, CUDA kernels, puzzles), finding that persistence—not initial attempt quality—is the dominant predictor of success. Claude-opus-4.6 led all categories, while most other models terminated prematurely or exhausted budgets with minimal progress.

Researchers from University of Toronto and Vector Institute propose Segment Tree Memory (SegTreeMem), a memory architecture for long-horizon conversational agents that preserves temporal order using a hierarchical segment tree structure for both online construction and retrieval. Experiments across three datasets show nearly 20% improvement in LLM-judge accuracy over non-temporal tree baselines.

S3Mem proposes a structured spatiotemporal scene-event memory framework for long-horizon interactive question answering, using anchor-sensitive retrieval and token-budget-aware evidence interface to outperform standard RAG in multiple environments.

LongMINT is a benchmark for evaluating memory under multi-target interference in long-horizon agent systems.

The post explains why Reinforcement Learning struggles with long-horizon tasks due to sparse rewards and highlights GEPA, a method that uses trajectory-level textual reflection to preserve richer feedback signals for optimization.

Memanto introduces a typed semantic memory system using a schema, conflict resolution, and Moorcheh's information-theoretic retrieval engine, achieving state-of-the-art results on LongMemEval and LoCoMo benchmarks with zero ingestion cost and sub-90ms latency.

TACO is a self-evolving framework that automatically discovers and refines context compression rules for long-horizon terminal agents.