{"slug":"learning-long-horizon","label":"Learning Long-Horizon","item_count":8,"day_count":7,"source_count":4,"first_seen":"2026-10-19T08:07:35+00:00","last_updated":"2026-11-03T17:49:14+00:00","generated_at":"2026-11-05T00:13:10.775833+00:00","sources":["arxiv_agent_systems_research","arxiv_cs_ai","arxiv_cs_cl","arxiv_cs_lg"],"days":[{"date":"2026-10-19","items":[{"title":"ClueFlow: Learning Long-Horizon Search through Attributed Research-State Transitions","url":"http://arxiv.org/abs/2610.25980v1","source":"arxiv_cs_ai","type":"paper","summary_1line":"Long-horizon search requires agents to acquire relevant evidence and integrate it into an evolving research state. Existing methods maintain research states to guide exploration or assign turn-level progress signals,...","why_it_matters":"Matches feed focus: agent, eval.","sid":"736cb3afaf8cf4b3","published":"2026-10-19T08:07:35+00:00","editor_note":"ClueFlow trains search agents on attributed research-state transitions."}]},{"date":"2026-10-25","items":[{"title":"Continual Learning Exam: Benchmarking Long-Horizon RSI and Experiential Learning in LLM Agents","url":"http://arxiv.org/abs/2610.33956v1","source":"arxiv_cs_cl","type":"paper","summary_1line":"Continual learning is the ability of LLM agents to improve over time by acquiring, refining, and applying knowledge from experience while retaining useful prior knowledge. It is essential for agents learning over long...","why_it_matters":"Matches feed focus: agent, harness, evaluation.","sid":"55edc2615764e1b5","published":"2026-10-25T00:00:25+00:00","editor_note":"Continual Learning Exam benchmarks experiential learning in LLM agents."}]},{"date":"2026-10-26","items":[{"title":"LongSearchEval: Evaluating Long-Horizon Web Search Capabilities of Large Language Models via a Trajectory-Aware Agent Framework with Failure-Node Localization","url":"http://arxiv.org/abs/2610.35150v1","source":"arxiv_agent_systems_research","type":"paper","summary_1line":"Long-horizon web search requires LLM-based agents to integrate information across web pages through extended interactions under task-specific constraints. Evaluating this capability is important not only for measuring...","why_it_matters":"Matches feed focus: agent, evaluation.","sid":"5f97c21b490b1ce9","published":"2026-10-26T10:15:38+00:00","editor_note":"LongSearchEval localizes failure nodes in long-horizon web search trajectories."},{"title":"FOCUS: Active Critic Learning for Long-Horizon Language Model Agents","url":"http://arxiv.org/abs/2610.35801v1","source":"arxiv_cs_lg","type":"paper","summary_1line":"Sparse outcome rewards make it difficult for long-horizon language-model agents to assign credit to earlier actions. Learned critics provide value estimates for intermediate states, but these estimates can be inaccura...","why_it_matters":"Matches feed focus: agentic.","sid":"3ca73f33ea96bd6a","published":"2026-10-26T17:34:47+00:00","editor_note":"FOCUS uses an active critic to address credit assignment under sparse rewards."}]},{"date":"2026-10-28","items":[{"title":"CALM: Conflict-Aware Lifecycle Memory Governance for Long-Horizon Language Agents","url":"http://arxiv.org/abs/2610.38573v1","source":"arxiv_cs_ai","type":"paper","summary_1line":"A memory can be outdated for present-day reasoning yet indispensable for reconstructing the past. Relevance-based retrieval overlooks this tension: semantic relevance, evidence-grounded validity, and query-specific el...","why_it_matters":"Matches feed focus: agent, eval.","sid":"423c688c0621c177","published":"2026-10-28T10:06:57+00:00","editor_note":"CALM governs memory lifecycle, separating outdated from historically needed memories."}]},{"date":"2026-10-29","items":[{"title":"Goal-Aware Dynamic Multi-Agent Collaboration for Long-Horizon Tasks","url":"http://arxiv.org/abs/2610.39193v1","source":"arxiv_agent_systems_research","type":"paper","summary_1line":"LLM-based agents are increasingly used to complete complex, long-horizon tasks without human intervention. To make progress toward the task goal, agents need to perform different duties, such as planning and verificat...","why_it_matters":"Matches feed focus: agent.","sid":"0b7565240a3894fa","published":"2026-10-29T02:15:35+00:00","editor_note":"Proposes goal-aware dynamic multi-agent collaboration for long-horizon tasks."}]},{"date":"2026-11-01","items":[{"title":"Is Your Long-Horizon Benchmark Really Deep? Measuring and Synthesizing Depth in Terminal Tasks","url":"http://arxiv.org/abs/2611.01927v1","source":"arxiv_cs_ai","type":"paper","summary_1line":"Long-horizon terminal tasks are used to evaluate and train agents. Human-curated datasets are scarce, so many synthesis methods have emerged. They usually measure the length of a task by the turns a model needs to pas...","why_it_matters":"Matches feed focus: agent, eval.","sid":"52a3c89c275c2670","published":"2026-11-01T14:05:27+00:00","editor_note":"Questions whether turn count measures true depth in terminal benchmarks and synthesizes deeper tasks."}]},{"date":"2026-11-03","items":[{"title":"JHiMA-WebRL: A Joint Hierarchical Multi-Agent Reinforcement Learning Framework for Long-Horizon Web Tasks","url":"http://arxiv.org/abs/2611.05263v1","source":"arxiv_agent_systems_research","type":"paper","summary_1line":"Multimodal web agents have shown promising capabilities for interacting with websites through graphical user interfaces (GUIs), yet realistic web tasks often require long-horizon, multi-step planning that remains chal...","why_it_matters":"Matches feed focus: agent, eval.","sid":"08c314bb237b4b82","published":"2026-11-03T17:49:14+00:00","editor_note":"JHiMA-WebRL trains hierarchical multi-agent policies jointly for long-horizon web tasks."}]}],"editorial":{"tldr":"Mid-October arXiv papers began framing long-horizon agent work around research-state tracking and continual-learning benchmarks. Late October added search evaluation with failure localization, critic-based credit assignment, and memory governance.","stale":false,"whats_new":"JHiMA-WebRL applies joint hierarchical multi-agent reinforcement learning to long-horizon web tasks, following a paper questioning whether long-horizon terminal benchmarks are actually deep.","why_it_matters":"Long-horizon failures (lost state, stale memory, weak credit assignment) are what break production agents; this cluster shows research moving from raw task length toward diagnosing where trajectories fail.","take_for_builders":"If you evaluate long-running agents, track failure nodes and memory staleness rather than turn counts alone; these are unreplicated preprints, so rerun on your own tasks before adopting any method.","beats":[{"kicker":"FRAMING","tone":"launch","headline":"Research state and continual learning define the problem","summary":"ClueFlow models search as attributed state transitions; Continual Learning Exam benchmarks experiential learning.","sids":["736cb3afaf8cf4b3","55edc2615764e1b5"]},{"kicker":"EVALUATION","tone":"rising","headline":"Failure localization, critics and memory governance","summary":"Papers target where long trajectories break: search failure nodes, sparse-reward credit, and outdated memories.","sids":["5f97c21b490b1ce9","3ca73f33ea96bd6a","423c688c0621c177"]},{"kicker":"THE TURN","tone":"turn","headline":"Multi-agent collaboration and a benchmark-depth critique","summary":"Goal-aware collaboration arrives alongside a challenge to how benchmark length is measured.","sids":["0b7565240a3894fa","52a3c89c275c2670"]},{"kicker":"NOW","tone":"now","headline":"Hierarchical multi-agent RL for web tasks","sids":["08c314bb237b4b82"]}],"open_questions":["Do any of these methods report results on a deployed agent rather than a research benchmark?","Does the benchmark-depth critique change how later long-horizon papers measure task length?","Do JHiMA-WebRL or CALM release code or cost and latency measurements?"],"generated_at":"2026-11-05T00:12:43+00:00"}}