Tag
This paper argues that long-horizon AI agent reliability is lacking despite better models, as seen on WeaveBench with only 41.2% pass rate, and proposes the LongHorizon-Harness to manage task state for improved performance.
ProgramBench is a benchmark for agents to reproduce real software behavior from scratch. Factory is announced as the most advanced reverse-engineering system with breakthroughs in model-agnostic long-horizon agents.
The author discusses insights from rebuilding self-service onboarding as agentic flows, emphasizing how long-horizon agents that combine deterministic task management with non-deterministic LLM actions improve reliability and reduce hallucinations in enterprise AI deployments.
ECHO is a cognitively inspired, auditable memory architecture for long-horizon agents, evaluated on benchmarks like LoCoMo and LongMemEval-S with high retrieval performance.
The paper introduces Recuris, a recursive memory architecture that improves long-horizon agent success by tracking progress and guiding skill selection through localized, validation-gated updates.
This paper presents a holistic evaluation of memory substrates for memory-augmented LLM agents, finding that no single substrate dominates across all regimes and advocating for adaptive substrate routing to optimize performance in different operating conditions.
LongRCA Bench introduces a benchmark for diagnosing failures in long-horizon agent trajectories, and the RCTA method improves responsible role and root-cause step attribution.
Introduces Governed Persistent Memory (GPM), a bitemporal state-transition model for auditable long-horizon agent memory with source-bound semantics and fail-closed release, validated on benchmarks and sealed evaluations.
This paper introduces MESA, a framework that dynamically selects and fuses a query-adaptive subset of multiple structural memory views for long-horizon agents, achieving 8.5% accuracy improvement over the strongest baseline while using 41% fewer evidence tokens.
This arXiv survey (1,547 papers, 2024-2026) systematically maps the field of long-horizon LLM agents, disambiguating long-horizon, long-context, and long-term memory, and organizing research into six lifecycle categories while identifying the core 'horizon gap' and open measurement problems.
This empirical study investigates how recurrent context compression affects long-horizon agent behavior, showing that compression can weaken recent interaction influence and cause instability. The authors introduce TRACE, a verifier-guided framework that improves compression reliability and performance on AppWorld.
MemPrism proposes a task-conditioned relational memory framework that separates persistent experience storage from decision-time working memory, enabling long-horizon agents to dynamically construct relational views for improved performance and reduced token usage.
StepReflect reformulates per-step GUI reflection for mobile agents as supervised structured prediction, achieving higher transition accuracy than GPT-5.2 on AndroidWorld while reducing API costs.
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.