{"total":14,"items":[{"citing_arxiv_id":"2607.05378","ref_index":8,"ref_count":2,"confidence":0.9,"is_internal_anchor":false,"paper_title":"CompactionRL: Reinforcement Learning with Context Compaction for Long-Horizon Agents","primary_cat":"cs.LG","submitted_at":"2026-07-06T17:55:12+00:00","verdict":"CONDITIONAL","verdict_confidence":"MODERATE","novelty_score":5.0,"formal_verification":"none","one_line_summary":"CompactionRL trains LLM agents to generate context summaries during RL rollouts, enabling long-horizon task completion under fixed context budgets with consistent gains on SWE-bench Verified and Terminal-Bench 2.0.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.31650","ref_index":14,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"ECHO: Prune To Act, Trace To Learn With Selective Turn Memory In Agentic RL","primary_cat":"cs.LG","submitted_at":"2026-06-30T13:29:58+00:00","verdict":"CONDITIONAL","verdict_confidence":"MODERATE","novelty_score":6.0,"formal_verification":"none","one_line_summary":"ECHO stores each agent turn as a source-indexed memory, reconstructs bounded contexts by selecting useful records, and routes RL credit through the same selection trace — reaching 43.4% on BrowseComp-Plus vs 28.9% (GRPO) and 36.1% (SUPO).","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.28566","ref_index":29,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"R$^2$-Searcher: Calibrating Retrieval and Reasoning Boundaries for Agentic Search","primary_cat":"cs.IR","submitted_at":"2026-06-26T19:44:34+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":4.0,"formal_verification":"none","one_line_summary":"R²-Searcher introduces fine-grained evidence modeling, retrieval reflection, and R²PO RL to calibrate retrieval-reasoning boundaries and improve multi-hop QA performance.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.28434","ref_index":26,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"SWE-MeM: Learning Adaptive Memory Management for Long-Horizon Coding Agents","primary_cat":"cs.SE","submitted_at":"2026-06-26T04:55:24+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"SWE-MeM introduces adaptive memory management for coding agents via synthesized trajectories and Memory-aware GRPO, reporting 43.4% and 60.2% resolve rates on SWE-Bench Verified for 4B and 30B models while beating baselines on performance and token use.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.21943","ref_index":126,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"Modularized Reinforcement Learning on LLMs: From MDP Creation to Exploration and Learning","primary_cat":"cs.LG","submitted_at":"2026-06-20T08:20:41+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"Survey mapping RL techniques onto LLM training and highlighting gaps in value-based, off-policy, and bootstrapping methods.","context_count":1,"top_context_role":"background","top_context_polarity":"background","context_text":"distinct variant is the process reward model (PRM) [113, 201], which provides step-level supervision, offering finer- grained feedback suited to multi-step reasoning tasks such as mathematics. Lightman et al. [113] introduce this approach at scale through dense human annotation of each reasoning step, producing the PRM800K dataset. To reduce annotation cost, Wang et al. [211] and Luo et al . [126] replace human labels with automatically derived step-level signals via outcome-based rollouts and Monte Carlo estimation respectively. Guan et al. [58] further eliminate explicit step-level scoring by training a process preference model through iterative MCTS-based self-evolution. However, Zhang et al. [261] find that Monte Carlo estimation often yields inferior labels compared to human annotation, and propose a"},{"citing_arxiv_id":"2606.28376","ref_index":6,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"When Does Overlap Help? OSU-Mem and a Cell-Conditional Analysis of Trajectory Memory for LLM Agents","primary_cat":"cs.IR","submitted_at":"2026-06-19T04:23:04+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"OSU-Mem shows overlapping memory helps retrieval when evidence shares tools or entities but hurts when steps are heterogeneous, with benefits on synthetic benchmarks vanishing on mixed real ones due to query mixing.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.11680","ref_index":21,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"Organize then Retrieve: Hierarchical Memory Navigation for Efficient Agents","primary_cat":"cs.AI","submitted_at":"2026-06-10T05:49:14+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"HORMA builds a hierarchical memory structure from agent experiences and trains a lightweight RL navigator to retrieve minimal sufficient context, yielding better task performance with at most 22.17% of baseline token usage on ALFWorld, LoCoMo, and LongMemEval.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.04536","ref_index":9,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"Scaling Self-Evolving Agents via Parametric Memory","primary_cat":"cs.AI","submitted_at":"2026-06-03T07:18:31+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"TMEM lets LLM agents evolve their policy mid-episode by absorbing distilled supervision into online LoRA updates, outperforming summary and retrieval baselines on several long-context benchmarks.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.24468","ref_index":20,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"SAM: State-Adaptive Memory for Long-Horizon Reasoning Agent","primary_cat":"cs.AI","submitted_at":"2026-05-23T08:37:16+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"SAM is a standalone memory framework for long-horizon LLM agents that creates state-adaptive cues from interactions, preserves raw trajectories for intent-driven recall, and optimizes the module via expert supervision and RL, outperforming baselines on BrowseComp and related benchmarks.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.15315","ref_index":30,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"Context Pruning for Coding Agents via Multi-Rubric Latent Reasoning","primary_cat":"cs.AI","submitted_at":"2026-05-14T18:30:10+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"LaMR decomposes code context pruning into two rubrics using dedicated CRFs, a mixture-of-experts gate, and AST-derived labels to filter noise and often match or beat full-context baselines on coding benchmarks.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.08580","ref_index":81,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"Slipstream: Trajectory-Grounded Compaction Validation for Long-Horizon Agents","primary_cat":"cs.MA","submitted_at":"2026-05-09T00:47:43+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"Slipstream uses asynchronous compaction with trajectory-grounded judge validation to improve long-horizon agent accuracy by up to 8.8 percentage points and reduce latency by up to 39.7%.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.05191","ref_index":6,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"LongSeeker: Elastic Context Orchestration for Long-Horizon Search Agents","primary_cat":"cs.AI","submitted_at":"2026-05-06T17:54:16+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"Context-ReAct enables agents to dynamically manage context via five atomic operations, and LongSeeker fine-tuned on 10k trajectories achieves 61.5% and 62.5% on BrowseComp benchmarks, outperforming prior agents.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.03804","ref_index":21,"ref_count":2,"confidence":0.9,"is_internal_anchor":false,"paper_title":"ScrapMem: A Bio-inspired Framework for On-device Personalized Agent Memory via Optical Forgetting","primary_cat":"cs.AI","submitted_at":"2026-05-05T14:30:30+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"ScrapMem reports SOTA 51.0% Joint@10 on ATM-Bench with up to 93% memory reduction and 70.3% Recall@10 via optical forgetting and EM-Graph.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2604.27859","ref_index":54,"ref_count":3,"confidence":0.9,"is_internal_anchor":false,"paper_title":"Rethinking Agentic Reinforcement Learning In Large Language Models","primary_cat":"cs.AI","submitted_at":"2026-04-30T13:43:25+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":3.0,"formal_verification":"none","one_line_summary":"The paper reviews conceptual foundations, methodological innovations, effective designs, critical challenges, and future directions for LLM-based Agentic Reinforcement Learning.","context_count":1,"top_context_role":"background","top_context_polarity":"background","context_text":"suffer from unstable credit assignment, the HiPER framework [67] is proposed. By factorizing the agent into a high-level planner and a low-level executor, HiPER facilitates hierarchical optimization via a novel technique called Hierarchical Advantage Estimation (HAE). Complementing this, the fundamental bottleneck of context length explosion in tool-use scenarios is tackled by a summarization-based training regime [54]. This approach compresses tool usage histories into A Brief Overview: Agentic Reinforcement Learning In Large Language Models 11 task-relevant summaries, enabling the derivation of a policy gradient representation that seamlessly optimizes both tool- use behaviors and summarization strategies within a compact, end-to-end RL infrastructure. In work [114], key insight"}],"limit":50,"offset":0}