AutoCompact preprint reports learned context management for coding agents
AutoCompact is a newly recorded arXiv preprint proposing learned context-compaction decisions for coding agents. Its authors report benchmark gains over a base model, but the supplied abstract does not establish peer review, independent replication, statistical certainty, or broad generalisation.
This edition passed Imananq's enhanced publication checks. Some material claims remain explicitly attributed to official or company sources because no independent source is currently bound to this edition. The engine continues checking approved sources and will add corroboration only through a new edition that passes the full gate.
AutoCompact is an arXiv preprint that proposes training long-horizon coding agents to decide when to compact their context, which working state to retain, and how to proceed afterward. Its authors report benchmark improvements over a base model, but the supplied record is an abstract-level account and does not establish peer review, reproducibility, or general performance. [1]
01
What we know now
- 01
arXiv records AutoCompact as version 1, submitted on 1 October 2026. [1]
- 02
The authors describe a learned policy for timing context compaction, preserving working state, and continuing afterward. [1]
- 03
The authors report absolute gains of 9.2 percentage points on SWE-bench Verified and 5.0 percentage points on SWE-PolyBench Verified versus their base model. [1]
02
Preprint status
The work is recorded as arXiv version 1, submitted on 1 October 2026. The supplied record does not establish peer review.Reported SWE-bench Verified change
The authors report an absolute pass-rate improvement over their base model; the abstract does not give the underlying pass rates or uncertainty measures.Reported SWE-PolyBench Verified change
The authors report an absolute pass-rate improvement over their base model; the supplied abstract does not provide task counts, variance, or independent replication.Evaluated context settings
The authors describe a 256K setting that did not overflow and a 16K setting where overflow triggered fallback compaction.Reported results and setup are attributed to the preprint authors, not independently verified findings. [1]
03
The method the authors describe
Long coding tasks can include inspection, search, editing, and testing over extended trajectories. AutoCompact is intended to handle the point at which earlier material has become less useful, rather than treating compaction only as a response to running out of context space. These are the authors’ design claims; the supplied abstract does not explain the judge’s design, correction rules, or full training protocol. [1]
- The preprint frames context management as a sequence of decisions: when to compact, what state to preserve, and how to continue after compaction.
- The authors say the agent is trained to make these choices as part of its policy during repository-level software-engineering tasks.
- For data collection, they say a judge reviews compaction decisions, summaries, and post-compaction actions. Outputs judged flawed are replaced with corrected outputs before the trajectory continues.
- The authors say these trajectories are used first for supervised fine-tuning and then for joint reinforcement-learning optimisation of coding and compaction using task-success rewards.
04
What the reported gains do and do not show
The reported figures support a narrow conclusion: in the authors’ evaluated setup, their method outperformed their stated base model by the reported absolute margins on those two benchmarks. They do not, from the supplied record alone, establish the size of the starting pass rates, the number of tasks evaluated, result variability, statistical significance, or whether the method would work similarly with another model, agent setup, or context-management technique. [1]
- On SWE-bench Verified, the authors report a 9.2 percentage-point absolute pass-rate improvement over their base model.
- On SWE-PolyBench Verified, they report a 5.0 percentage-point absolute pass-rate improvement over their base model.
- They say the improvements held across their evaluated inference budgets.
- The abstract describes a 256K context-window setting that never overflowed, and a 16K setting where overflow triggered fallback compaction.
05
Status and source
arXiv records the work as a preprint. That establishes the publication record and the authors’ abstract-level claims, but not external validation. The primary record is available at https://arxiv.org/abs/2610.02163v1. [1]
- The arXiv record lists the title as “AutoCompact: Learning When to Compact Context in Long-Horizon Coding Agents.”
- It names Xuan Zhang, Longtao Zheng, Cunxiao Du, Bo An, and Xin Dong as authors.
- The record is version 1 and shows a submission date of 1 October 2026.
06
How to read the result
Treat AutoCompact as a research claim documented in a newly posted preprint, rather than as independently established evidence of more reliable coding agents.
- 01
Read the authors’ reported gains as absolute pass-rate differences against their base model, not as a guarantee for another model, repository, or workflow.
- 02
Check the full paper for evaluation details before relying on the approach: the supplied record does not provide baseline pass rates, task counts, variation, statistical tests, or comparisons with other context-management methods.
- 03
Use the primary arXiv record to follow the preprint and any later versions: https://arxiv.org/abs/2610.02163v1
07
Limits of this edition
The source is an arXiv preprint record and abstract, not evidence of peer review or independent validation. [1]
The supplied material does not state baseline pass rates, task counts, statistical variation, confidence intervals, or significance tests. [1]
It does not provide judge criteria, implementation code, training data, ablations, failure cases, or reproducibility materials. [1]
The record does not establish comparisons with other context-management approaches or generalisation to other models and agent configurations. [1]
SRC
Source desk
Direct links to the material behind this selection. Seeing the source matters as much as reading the synthesis.
Keep exploring
21 Sept 2026
Preprint examines coding agents’ claims of completed file reviews
An arXiv preprint introduces a five-scenario benchmark for comparing coding agents’ final review summaries with file coverage. Its authors report frequent incomplete reading and inadequate disclosure in incomplete runs, but peer review, replication, and broader generalisability are not established by the supplied record. [1]
30 Aug 2026
Preprint claims selective training data can improve software-agent benchmarks
The authors of the SWE-Prime preprint propose filtering software-agent training data at trajectory and segment levels while retaining full sequence context. They report that a selected 10% trajectory subset beat full-dataset training on two SWE-Bench evaluations, but the supplied evidence is limited to a preprint abstract for which peer review is not established and does not support conclusions about reproducibility or broader performance.
30 Sept 2026
Preprint proposes test-time adaptation by revising an AI agent’s workflow
The authors of a new arXiv preprint propose “harness learning”: training a model to revise the executable workflow around a language-model agent from execution feedback. They report gains on reasoning and multi-hop question-answering tasks and transfer to unseen tasks, but the supplied evidence contains no scores, baselines, code, peer review or independent replication.