mlsanityData to 7 Oct 2026

DRSR: Learning Set-Level Deletion Risk for Efficient Long-Horizon Agents

arXiv 2609.27276 ↗First posted 23 Sep 2026cs.AIcited by 0

On the map: LLM reasoning → Reasoning for agents, planning and code also fits: Test-time compute and efficient thinking

also in 1 more
Abstract (arXiv)

Long-horizon language-model agents accumulate reasoning traces, tool exchanges, and observations whose relevance changes with the current decision. Existing compression strategies often score historical units independently, but the safety of deleting several units is generally not determined by their singleton scores: redundant evidence, accumulated small effects, and the information that remains after deletion all matter. We introduce Direct Relational Set-Risk Pruning (DRSR), which formulates agent-history compression as risk-constrained selection over deletion sets. Offline, DRSR constructs exact counterfactual supervision by jointly deleting protocol-valid history Blocks and measuring the change in teacher-forced likelihood of the same recorded next output. A lightweight scorer then predicts set-level harm from online-visible relations between candidate history and the current pre-action state, together with deleted-retained and pairwise set structure. At deployment, DRSR evaluates a small set of structurally valid deletion candidates with the lightweight scorer and removes the largest feasible set under recency, protocol, budget, and learned-risk constraints, abstaining when no set is sufficiently safe. On WorkBuddyBench Full260, DRSR increases mean reward from 0.699 to 0.802 while reducing total model tokens by 20.820%. On the fixed Eval40 comparison, it obtains 0.794 reward at 1.211M tokens per task, using 35.850% fewer tokens than the uncompressed agent. Mechanistic analyses and ablations further show that decision-conditioned relations, retained-context information, pair interactions, and abstention each contribute to reliable pruning.

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Placed in Reasoning for agents, planning and code, one of 17 directions in LLM reasoning. Builds on 42 of its 51 references that are in the map. First posted on arXiv 23 Sep 2026.

Builds on

42 of 51 references.

Closest ancestors · 1

Same direction, newest first

  1. 2026SAM: State-Adaptive Memory for Long-Horizon Reasoning Agent23 May

Other references · 41

  1. 2015Distilling the Knowledge in a Neural Networkcited by 10,076
  2. 2022ReAct: Synergizing Reasoning and Acting in Language Modelscited by 8,355
  3. 2023Lost in the Middle: How Language Models Use Long Contextscited by 3,383
  4. 2023SWE-bench: Can Language Models Resolve Real-World GitHub Issues?cited by 3,193
  5. 2023Efficient Streaming Language Models with Attention Sinkscited by 2,115
  6. 2023MemGPT: Towards LLMs as Operating Systemscited by 1,295
  7. 2024SnapKV: LLM Knows What You are Looking for Before Generationcited by 848
  8. 2023Large Language Models Can Be Easily Distracted by Irrelevant Contextcited by 820
  9. 2023H2O: Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Modelscited by 805
  10. 2023Semantic Uncertainty: Linguistic Invariances for Uncertainty Estimation in Natural Language Generationcited by 795
  11. 2024AppWorld: A Controllable World of Apps and People for Benchmarking Interactive Coding Agentscited by 359
  12. 2024LLMLingua-2: Data Distillation for Efficient and Faithful Task-Agnostic Prompt Compressioncited by 343
29 more
  1. 2019SelectiveNet: A Deep Neural Network with an Integrated Reject Optioncited by 333
  2. 2022Conformal Risk Controlcited by 290
  3. 2025Chain of Draft: Thinking Faster by Writing Lesscited by 290
  4. 2018Consistent Individualized Feature Attribution for Tree Ensemblescited by 279
  5. 2023LLMLingua: Compressing Prompts for Accelerated Inference of Large Language Modelscited by 279
  6. 2021Distribution-Free, Risk-Controlling Prediction Setscited by 278
  7. 2025TokenSkip: Controllable Chain-of-Thought Compression in LLMscited by 246
  8. 2024Ada-KV: Optimizing KV Cache Eviction by Adaptive Budget Allocation for Efficient LLM Inferencecited by 230
  9. 2015Unifying distillation and privileged informationcited by 218
  10. 2020Explaining by Removing: A Unified Framework for Model Explanationcited by 218
  11. 2024Kernel Language Entropy: Fine-grained Uncertainty Quantification for LLMs from Semantic Similaritiescited by 156
  12. 2019The Shapley Taylor Interaction Indexcited by 143
  13. 2025Scaling Long-Horizon LLM Agent via Context-Foldingcited by 136
  14. 2024Language Models with Conformal Factuality Guaranteescited by 128
  15. 2025ACON: Optimizing Context Compression for Long-horizon LLM Agentscited by 126
  16. 2025AgentFold: Long-Horizon Web Agents with Proactive Context Managementcited by 95
  17. 2022Faith-Shap: The Faithful Shapley Interaction Indexcited by 68
  18. 2025Memory as Action: Autonomous Context Curation for Long-Horizon Agentic Taskscited by 62
  19. 2023SHAP-IQ: Unified Approximation of any-order Shapley Interactionscited by 50
  20. 2026SWE-Pruner: Self-Adaptive Context Pruning for Coding Agentscited by 40
  21. 2025Making Slow Thinking Faster: Compressing LLM Chain-of-Thought via Step Entropycited by 37
  22. 2025Sculptor: Empowering LLMs with Cognitive Agency via Active Context Managementcited by 17
  23. 2025Git Context Controller: Manage the Context of LLM-based Agents like Gitcited by 13
  24. 2026Self-Compacting Language Model Agentscited by 11
  25. 2026Self-GC: Self-Governing Context for Long-Horizon LLM Agentscited by 8
  26. 2026ACM: Agentic Context Management for Long Horizon Taskscited by 8
  27. 2026ForesightKV: Optimizing KV Cache Eviction for Reasoning Models by Learning Long-Term Contributioncited by 7
  28. 2026CoMem: Context Management with A Decoupled Long-Context Modelcited by 5
  29. 2026Learning What Not to Forget: Long-Horizon Agent Memory from a Few Kilobytes of Learningcited by 4

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026StateComp: Learning When to Compress History in Long Horizon Agents
  2. 2026When Can Agents Forget Their Reasoning? ICLR for Long-Horizon Agent Context Compression
  3. 2026The Missing Complement: State-Conditioned Minimal Sufficient Evidence for Coding Agents
  4. 2026FOCUS: Training-Free Decision-Preserving Context Compression for LLM Agents
  5. 2026Learning What to Remember: Observability-Safe Memory Retention via Constrained Optimization for Long-Horizon Language Agents
  6. 2026Beyond Memory: Harnessing Long-Horizon Agents with Explicit Belief States
  7. 2026Memory Control Signals Emerge Before Action in Long Horizon Agents
  8. 2026Agent-BRACE: Decoupling Beliefs from Actions in Long-Horizon Tasks via Verbalized State Uncertainty