# Context Management

10 related articles

Subscribe to updates

arXiv Nov 5, 2026
CORE retools the deep-research harness: only the best snippet enters context, and on a 553M-document corpus its answer accuracy leads by 12.28 points
CORE, a training-free harness from Qualcomm AI Research, splits retrieval into coarse and fine stages: a dense retriever returns whole passages, then token-level similarities from a late-interaction reranker are reused to hand the agent only the most relevant snippet from each passage. In controlled closed-search comparisons over fixed corpora, CORE reaches the highest answer accuracy and evidence recall in every model-benchmark group across three agent models and three benchmarks, with corpora ranging from 100K web documents to 553M ClimbMix documents.
CORE, a training-free harness from Qualcomm AI Research, splits retrieval into coarse and fine stages: a dense retriever returns whole passages, then token-level similarities from a late-interaction reranker are reused to hand the agent only the most relevant snippet from each passage. In controlled closed-search comparisons over fixed corpora, CORE reaches the highest answer accuracy and evidence recall in every model-benchmark group across three agent models and three benchmarks, with corpora ranging from 100K web documents to 553M ClimbMix documents.
Hoang Tran Vuong（First author）
CORE Coarse-to-Fine Search Context Management Deep Research Agent
arXiv Oct 13, 2026
Frozen models improve without weight updates: TextualRL turns repeated same-task attempts into context patches, gaining up to 24.4 points on a six-benchmark average
TextualRL critiques a group of rollouts from the same task under the same context as a unit, rather than an isolated trajectory, and writes reviewed textual patches back into the shared context while model parameters stay frozen. Across five target models and six benchmarks it reaches the highest six-benchmark average for every target, gaining 11.4 to 24.4 percentage points over the no-skill baseline and 0.9 to 3.2 points over the strongest baseline, SkillOpt.
TextualRL critiques a group of rollouts from the same task under the same context as a unit, rather than an isolated trajectory, and writes reviewed textual patches back into the shared context while model parameters stay frozen. Across five target models and six benchmarks it reaches the highest six-benchmark average for every target, gaining 11.4 to 24.4 percentage points over the no-skill baseline and 0.9 to 3.2 points over the strongest baseline, SkillOpt.
Fudan University Shanghai Innovation Institute East China Normal University
Tianyu Huai（First author）
Context Management TextualRL Gradient-free Reinforcement Learning Groupwise Policy Critique Cross-Group Policy Update
Claude 产品博客 Oct 22, 2026
More agents isn't always better: folding specialist subagents into one loop often raised quality and cut latency
In enterprise agent deployments, folding routers, planners and domain subagents into a single main loop often beat both a one-prompt-for-everything design and a subagent orchestration design on quality, frequently at lower cost and latency per task. Subagents still fit when context bloat degrades a single loop, when tasks need parallel fresh context windows, or when tool choice and instructions conflict. A guided workflow lets teams try the migration on a mock system first.
In enterprise agent deployments, folding routers, planners and domain subagents into a single main loop often beat both a one-prompt-for-everything design and a subagent orchestration design on quality, frequently at lower cost and latency per task. Subagents still fit when context bloat degrades a single loop, when tasks need parallel fresh context windows, or when tool choice and instructions conflict. A guided workflow lets teams try the migration on a mock system first.
Agent Harness multi-agent systems Context Management Claude API
langchain.com Oct 15, 2026
Long-running agents stop drowning in context: LangChain moves memory into files that survive a model swap
LangChain's open source Deep Agents harness applies five context engineering techniques against context rot in long tasks: offloading oversized tool results and user messages to the filesystem, summarizing history near the limit, loading skills on demand, isolating noisy work in subagents, and keeping long-term memory as a plain file. Information is moved out of the active window rather than deleted and can be read back, so memory lives in storage the user controls and survives a model or provider switch.
LangChain's open source Deep Agents harness applies five context engineering techniques against context rot in long tasks: offloading oversized tool results and user messages to the filesystem, summarizing history near the limit, loading skills on demand, isolating noisy work in subagents, and keeping long-term memory as a plain file. Information is moved out of the active window rather than deleted and can be read back, so memory lives in storage the user controls and survives a model or provider switch.
Context Engineering Deep Agents Context Management
langchain.com Oct 13, 2026
Long-running agents stop paying frontier prices for context chores: ScaleDown models land in LangChain
Long-running agents previously made every context-management decision with a general-purpose LLM, which is slow and expensive at the volume such agents need. ScaleDown's constrained decoding models (CDMs) are trained for one task and only return answers that task allows, so they can be smaller, faster and cheaper; a new LangChain and Deep Agents integration ships three middlewares for summarization, tool-result compression and structured extraction, and the LangSmith LLM Gateway now lists ScaleDown as a direct model provider so calls are traced and count toward spend and rate limits. ScaleDown reports its models are 200x faster and 467x cheaper than frontier models on these narrow tasks.
Long-running agents previously made every context-management decision with a general-purpose LLM, which is slow and expensive at the volume such agents need. ScaleDown's constrained decoding models (CDMs) are trained for one task and only return answers that task allows, so they can be smaller, faster and cheaper; a new LangChain and Deep Agents integration ships three middlewares for summarization, tool-result compression and structured extraction, and the LangSmith LLM Gateway now lists ScaleDown as a direct model provider so calls are traced and count toward spend and rate limits. ScaleDown reports its models are 200x faster and 467x cheaper than frontier models on these narrow tasks.
langchain-scaledown integration Constrained Decoding Context Management LangSmith LLM Gateway
arXiv Oct 1, 2026
PoS manages long-horizon agents with explicit belief states, achieving the highest overall performance on four benchmarks across three LLM backbones
The work introduces PoS, an inference-time framework that constructs and continually maintains explicit belief states as the agent's decision context, where each belief combines an estimate of the current world state with unresolved task requirements; PoS validates belief consistency and monitors task progress to detect Belief Trapping, then tailors recovery to the trapping pattern and the type of unresolved requirement. On four benchmarks spanning execution and diagnosis, PoS achieves the highest overall performance with all three LLM backbones; ablations show the importance of consistency validation and recovery, and context-scaling experiments show resilience to context growth.
The work introduces PoS, an inference-time framework that constructs and continually maintains explicit belief states as the agent's decision context, where each belief combines an estimate of the current world state with unresolved task requirements; PoS validates belief consistency and monitors task progress to detect Belief Trapping, then tailors recovery to the trapping pattern and the type of unresolved requirement. On four benchmarks spanning execution and diagnosis, PoS achieves the highest overall performance with all three LLM backbones; ablations show the importance of consistency validation and recovery, and context-scaling experiments show resilience to context growth.
Yu Luo（First author）
PoS explicit belief state framework LLM agent Context Management Belief Trapping detection and recovery
arXiv Oct 2, 2026
Alibaba team's PoS adds explicit belief states plus Belief Trapping detection and recovery, topping all four benchmarks across three backbones
The work introduces PoS, an inference-time framework that replaces interaction history as the agent's decision context with a continually maintained explicit belief state (world state plus unresolved epistemic and achievement gaps), validates belief updates through a Belief Sentinel, detects Belief Trapping (stagnation, cycles, drift) from window-level signals, and composes recovery constraints tailored to the trapping pattern and blocked gap type; across four benchmarks (ALFWorld, LOCA-Bench, RCA-100, ClinDiag) and three backbones (Qwen3.7-Plus, Kimi-K3, GLM-5.3), PoS achieves the highest overall performance in all 12 benchmark-backbone combinations, with relative gains over the strongest same-backbone baseline reaching 22.68% on ALFWorld and 37.89% on RCA-100 joint accuracy.
The work introduces PoS, an inference-time framework that replaces interaction history as the agent's decision context with a continually maintained explicit belief state (world state plus unresolved epistemic and achievement gaps), validates belief updates through a Belief Sentinel, detects Belief Trapping (stagnation, cycles, drift) from window-level signals, and composes recovery constraints tailored to the trapping pattern and blocked gap type; across four benchmarks (ALFWorld, LOCA-Bench, RCA-100, ClinDiag) and three backbones (Qwen3.7-Plus, Kimi-K3, GLM-5.3), PoS achieves the highest overall performance in all 12 benchmark-backbone combinations, with relative gains over the strongest same-backbone baseline reaching 22.68% on ALFWorld and 37.89% on RCA-100 joint accuracy.
Yu Luo（First author）
PoS (Progression of States)Explicit Belief States Belief Trapping detection and recovery LLM Agents Context Management
arXiv Sep 30, 2026
Turning context into an editable file: CLM lets models manage their own context, raising BrowseComp-Plus accuracy by 11.4% with 21.5% fewer FLOPs
The work introduces Context Language Models (CLMs), which mirror a model's live context as a freely editable file that the model can rewrite with Bash, shifting context management from external harness control to an intrinsic model capability; applied zero-shot to existing models such as Qwen3.6-27B and GPT5.6-Sol, CLMs achieve 11.4% higher accuracy with 21.5% fewer prefix-reuse FLOPs than the strongest baseline on BrowseComp-Plus, 5% higher scores with 59% fewer FLOPs on 12-hour EdgeBench, and 65% greater end-to-end speedup at the same compute on a 24-hour multi-repository agent-swarm task, and can further learn context-management strategies through natural-language instructions, textual evolution, and online reinforcement learning.
The work introduces Context Language Models (CLMs), which mirror a model's live context as a freely editable file that the model can rewrite with Bash, shifting context management from external harness control to an intrinsic model capability; applied zero-shot to existing models such as Qwen3.6-27B and GPT5.6-Sol, CLMs achieve 11.4% higher accuracy with 21.5% fewer prefix-reuse FLOPs than the strongest baseline on BrowseComp-Plus, 5% higher scores with 59% fewer FLOPs on 12-hour EdgeBench, and 65% greater end-to-end speedup at the same compute on a 24-hour multi-repository agent-swarm task, and can further learn context-management strategies through natural-language instructions, textual evolution, and online reinforcement learning.
University of Washington Massachusetts Institute of Technology
Rulin Shao（First author）
Context Language Models Context Management Suffix Cache Reuse ContextBench GRPO
arXiv Sep 29, 2026
ACG lifts long-horizon agent success from 44.5% to 50.2%, with an 11.1-point gain on BrowseComp-Plus
The work introduces the Adaptive Consistency Graph (ACG), which incrementally organizes execution evidence and its provenance into a persistent graph and builds a temporary requirement-centered view for each decision under a bounded context budget, raising GPT-5.6-luna's equal-weight average success across three benchmarks from 44.5% with ReAct to 50.2% and BrowseComp-Plus from 62.4% to 73.5%, without replacing the base planner or tool executor.
The work introduces the Adaptive Consistency Graph (ACG), which incrementally organizes execution evidence and its provenance into a persistent graph and builds a temporary requirement-centered view for each decision under a bounded context budget, raising GPT-5.6-luna's equal-weight average success across three benchmarks from 44.5% with ReAct to 50.2% and BrowseComp-Plus from 62.4% to 73.5%, without replacing the base planner or tool executor.
Beihang University
Jiecong Wang（First author）
Adaptive Consistency Graph Agent Memory Context Management Long-Horizon Reasoning
arXiv Sep 23, 2026
Recursive Self-Improvement of AI Research Agents: AIDE² Finds Seven Successive Improvements in an 8-Day Autonomous Run
The work presents AIDE², a two-loop system that implements recursive self-improvement at the harness layer of an AI research agent: an inner-loop agent optimizes code on AI R&D tasks while an outer-loop agent rewrites the inner-loop agent itself, and in one autonomous 8-day run it accepted seven successive improvements that transferred to four held-out benchmarks (including an out-of-distribution physics-based weather-forecasting task) and reduced reward hacking from 55% to 32% on a behavior the loop never optimized for.
The work presents AIDE², a two-loop system that implements recursive self-improvement at the harness layer of an AI research agent: an inner-loop agent optimizes code on AI R&D tasks while an outer-loop agent rewrites the inner-loop agent itself, and in one autonomous 8-day run it accepted seven successive improvements that transferred to four held-out benchmarks (including an out-of-distribution physics-based weather-forecasting task) and reduced reward hacking from 55% to 32% on a behavior the loop never optimized for.
Dhruv Srikanth（First author）
Recursive Self-Improvement AIDE^2 Self-Evolving Agents Context Management Reward Hacking

You’re all caught up