long-horizon-papers

A curated collection of papers on long-horizon tasks, spanning LLM agents, vision-language models (VLM), and vision-language-action models (VLA). Focus is on 2026 work, plus key 2025 foundations on the VLA side. Used as a submodule of paper-shelf; follows the same conventions (one folder per paper, source.txt metadata, PDFs gitignored).

LLM — long-horizon agents

Date	Paper	arXiv
Jun 2026	Agent Memory: Characterization and System Implications of Stateful Long-Horizon Workloads — first systems-level characterization of agent memory, with a taxonomy along four axes	2606.06448
May 2026	Meta-Cognitive Memory Policy Optimization for Long-Horizon LLM Agents — "Belief Entropy" probes how uncertain the model is about latent task state given its compressed memory	2605.30159
Apr 2026	Rethinking Agentic Reinforcement Learning in Large Language Models — RL for agentic LLMs with focus on long-horizon planning and tool use	2604.27859
Apr 2026	YC-Bench: Benchmarking AI Agents for Long-Term Planning and Consistent Execution — agent runs a simulated startup over a one-year horizon, hundreds of turns	2604.01212
Mar 2026	Beyond pass@1: A Reliability Science Framework for Long-Horizon LLM Agents — reliability-focused evaluation methodology	2603.29231
Mar 2026	The Y-Combinator for LLMs: Solving Long-Context Rot with Lambda-Calculus — tackles context degradation over long executions	2603.20105
Mar 2026	A Subgoal-driven Framework for Improving Long-Horizon LLM Agents — online subgoal decomposition + MiRA, RL with milestone-based dense rewards	2603.19685
Mar 2026	Memex(RL): Scaling Long-Horizon LLM Agents via Indexed Experience Memory — indexed experience memory for workflows spanning hundreds of steps	2603.04257
Feb 2026	KLong: Training LLM Agent for Extremely Long-horizon Tasks — withdrawn by author (Apr 2026) due to data errors	2602.17547
Feb 2026	AgentLAB: Benchmarking LLM Agents against Long-Horizon Attacks — security angle: adaptive multi-turn attacks	2602.16901
Jan 2026	Continuum Memory Architectures for Long-Horizon LLM Agents — memory for long-running workflows	2601.09913
Jan 2026	InfiAgent: An Infinite-Horizon Framework for General-Purpose Autonomous Agents	2601.03204

VLM — long-horizon planning

Date	Paper	arXiv
Apr 2026	Goal2Skill: Long-Horizon Manipulation with Adaptive Planning and Reflection — VLM planner with structured task memory: post-condition verification + reflection to diagnose failures (32.4% vs 9.8% strongest baseline)	2604.13942
Mar 2026	AgentVLN: Towards Agentic Vision-and-Language Navigation — agentic VLN over extended horizons	2603.17670
Mar 2026	Recurrent Reasoning with Vision-Language Models for Estimating Long-Horizon Embodied Task Progress — progress estimation as the key signal for multi-step execution	2603.17312

VLA — long-horizon manipulation

Organized by the competing answers to "how does a VLA survive a long task?" — the first four folders are strategies for not failing; Recovery/ is the newer line on failing gracefully.

Hierarchical/ — VLM planner ⟷ VLA executor

Date	Paper	arXiv
Jul 2026	Harness VLA: Steering Frozen VLAs into Reliable Manipulation Primitives via Memory-Guided Agents — memory-guided agent composes frozen VLAs with analytic primitives	2607.08448
Jul 2026	Cortex: A Bidirectionally Aligned Embodied Agent Framework — 32 canonical skill primitives + tractability principles for subtask transitions; attacks the planner/executor semantic gap head-on	2607.05377
Jun 2026	What Matters in Orchestrating Robot Policies: A Systematic Study of Hierarchical VLA Agents — options-style framework benchmarking planner-controller configs; the closest thing to a survey of this folder	2606.10267
Jun 2026	VoLo: A Physical Orchestrator for Open-Vocabulary Long-Horizon Manipulation — VLA/WAM as an interruptible tool the VLM steers mid-execution (cf. Tool-Aligned VLA)	2606.07723
May 2026	Towards Long-horizon Embodied Agents with Tool-Aligned VLA Models — "VLAs-as-tools": planning VLM invokes bounded VLA executions; tool-family residual adapters	2605.13119
May 2026	Anticipation-VLA — unified multimodal model as both anticipation + value model to generate subgoals; refines subgoals when progress stalls	2605.01772
Apr 2026	LoHo-Manip: Long-Horizon Manipulation via Trace-Conditioned VLA Planning — task-management VLM above any VLA; trace-conditioned executor makes replanning implicit	2604.21924
Feb 2026	LiLo-VLA: Compositional Long-Horizon Manipulation via Linked Object-Centric Policies — modular linked policies, dynamic replanning + skill reuse	2602.21531

Memory/ — temporal context inside the policy

Date	Paper	arXiv
Jul 2026	NativeMEM: Native Memory Compression — compresses history into single tokens using the VLA's own vision encoder; no external memory module	2607.06678
Jul 2026	HiMe: Hierarchical Embodied Memory — high-frequency Executor / working-memory Sentry / Planner split; names the "frequency-competence paradox"	2607.03449
Jun 2026	Chronos: A Physics-Informed Full-History Framework for Non-Markovian Long-Horizon Manipulation — selective SSM over full history (cf. Keyframe Chaining, VQ-Memory)	2606.30318
Jun 2026	S²-VLA: State-Space Guided VLA — maintains a belief state that gates fusion of visual/language/action features by execution phase	2606.27872
Jun 2026	EventVLA: Event-Driven Visual Evidence Memory — learns which keyframes to keep by predicting future keyframe probability	2606.20092
Jun 2026	μVLA: On Recurrent Memory for Partially Observable Manipulation — the control experiment: what does recurrence alone buy, with no auxiliary losses?	2606.12497
Apr 2026	HELM: Harness-Enhanced Long-horizon Memory — names the three execution-loop gaps: memory (cross-phase context loss), verification, recovery	2604.18791
Mar 2026	VQ-Memory — discretizes past joint states via VQ-VAE; compact, model-agnostic temporal memory	2603.09513
Mar 2026	Non-Markovian Long-Horizon Robot Manipulation via Keyframe Chaining — sparse semantic keyframe history, reasons over events thousands of timesteps apart	2603.01465
Feb 2026	Recursive Belief Vision Language Action Models	2602.20659
Feb 2026	StemVLA — future 3D spatial geometry + 4D historical representation	2602.23721
Feb 2026	Efficient Long-Horizon VLA Models via Static-Dynamic Disentanglement — splits visual tokens into static/dynamic to cut context length	2602.03983
Nov 2025	EchoVLA: Synergistic Declarative Memory for VLA-Driven Mobile Manipulation	2511.18112
Aug 2025	MemoryVLA: Perceptual-Cognitive Memory in VLA Models — the foundational one; perceptual-cognitive memory bank, +26 pts on long-horizon real-world tasks	2508.19236

WorldModel/ — subgoal imagination & visual foresight

Date	Paper	arXiv
Jul 2026	Imagined Rollouts are Kinematic, Not Dynamic: A Diagnosis of Long-Horizon World-Model Failure — shows imagined rollouts fail kinematically, not dynamically; the sharpest negative result in this folder	2607.05966
Jun 2026	Foresight: Failure Detection with Action-Conditioned World Model Latents — detects failure from world-model latents using only task-level labels + conformal thresholds	2606.23085
May 2026	World Model for Robot Learning: A Comprehensive Survey — frames subgoal generation in VLAs as "predictive branches" vs native video world models	2605.00080
Mar 2026	Do World Action Models Generalize Better than VLAs? A Robustness Study	2603.22078
Feb 2026	H-WM: Robotic Task and Motion Planning Guided by Hierarchical World Model — understanding + prediction experts map symbolic plans to visually grounded subgoals	2602.11291
Feb 2026	BagelVLA — interleaves textual reasoning + visual forecasting into the action loop	2602.09849
Jan 2026	PALM: Progress-Aware Policy Learning via Affordance Reasoning — future affordance prediction + progress estimation; SOTA on CALVIN, LIBERO-LONG	2601.07060
Jul 2025	DreamVLA — predictive "dreaming" branch with comprehensive world knowledge	2507.04447

Benchmarks/ — evaluating long-horizon manipulation

Date	Paper	arXiv
Jul 2026	RoboDojo: A Unified Sim-and-Real Benchmark for Generalist Robot Manipulation Policies — evaluates 30 policies along generalization / memory / precision / horizon axes	2607.04434
Apr 2026	LongBench: Evaluating Robotic Manipulation Policies on Real-World Long-Horizon Tasks — real-world, not sim	2604.16788
Mar 2026	RoboMME: Benchmarking and Understanding Memory for Robotic Generalist Policies — 16 tasks taxonomized by temporal/spatial/object/procedural memory	2603.04639
Aug 2025	Long-VLA — first end-to-end long-horizon VLA (phase-aware input masking); introduced the L-CALVIN benchmark	2508.19958

Classic baselines: CALVIN (5-subtask chains) and LIBERO-Long (100 long-horizon tasks).

Recovery/ — verification, failure detection & correction

HELM names three execution-loop gaps — memory, verification, recovery — but the first three folders only answer the memory one. This folder collects the mid-2026 cluster on the other two: noticing that a long rollout has gone wrong, and doing something about it without restarting.

Date	Paper	arXiv
Jul 2026	Diagnosing Semantic Handoff Failures in Agent-Orchestrated VLA Skill Composition — the failure taxonomy paper: skills that pass in isolation break at their boundaries when chained	2607.06256
Jul 2026	VLA-Corrector: Lightweight Detect-and-Correct Inference for Adaptive Action Horizon — detects visual deviation mid-chunk and triggers online replanning	2607.01804
Jun 2026	PhysReflect-VLA: Physical Feasibility and Self-Reflective Regulation — execution-time feasibility check + LLM reflection for error correction	2606.27146

See also Foresight (WorldModel/), which does failure detection from world-model latents.

Trends (as of Jul 2026)

• LLM side: shift from prompting/planning tricks to memory systems (indexed experience, continuum/meta-cognitive memory) and RL with dense milestone rewards.
• VLA side: four competing strategies — (1) put a VLM planner on top, (2) build memory into the policy, (3) imagine subgoals with a world model, (4) end-to-end with phase-aware tricks. HELM's gap taxonomy (memory / verification / recovery) is a useful organizing lens.
• Memory is consolidating on selection, not capacity. The Jun–Jul 2026 wave (EventVLA, NativeMEM, Chronos, HiMe) is less about storing more history than about choosing what to keep — learned keyframe evidence, native-encoder compression, selective state spaces. μVLA is the useful control: how much of this does plain recurrence already give you?
• Recovery is the new front. Through ~May 2026 the field answered "don't fail." Since June it has started answering "notice you failed, then fix it" — see Recovery/ and Foresight. The handoff-failure diagnosis paper argues the damage concentrates at skill boundaries, which is exactly the seam hierarchical VLAs create.
• World-model foresight took a hit: "Imagined Rollouts are Kinematic, Not Dynamic" (Jul 2026) is a pointed negative result for the subgoal-imagination strategy.
• Benchmarks are shifting from sim chains (CALVIN) to memory-probing (RoboMME), real-world (LongBench), and unified sim+real with an explicit horizon axis (RoboDojo).

52 papers total · 51 with PDFs · 1 withdrawn (KLong).

PDFs are gitignored (same policy as paper-shelf); re-download via the PDF: link in each source.txt.