ContextPilot's snapshot-level RL reaches 69.4 average on long-context QA at 32K

ContextPilot: Teaching Agents for Proactive Context Management via Fine-grained RL

Zhuoshi Pan, Qizhi Pei, Junru Lu, Honglin Lin, H. Vicky Zhao, Di Yin, Xing Sun

EMNLP 2026 (Main Track)

cs.CL

2026-08-29

ContextPilot adds planning, long-term memory, and soft offloading tools, then spends extra RL rollouts on high-impact context edits. ContextPilot-8B-RL averages 69.40 on four long-context QA tasks, 3.55 points above StateLM-8B-RL.

What problem this solves

Long-horizon agents grow a working context every turn. Rule-based truncation treats the model as a passenger. Proactive managers such as StateLM let the model search, delete, and summarize, but the toolset is thin: little planning, little long-term memory, little reversible compression. Training is coarser still. A context edit rewrites later history, and different edits have very different effects on final success, yet the same trajectory-level reward is copied onto every intermediate action.

ContextPilot targets three gaps: tools, exploration budget, and credit.

Method

On top of StateLM-style search, deletion, and summarization, the toolset adds planning (plan, analyzeText, checkBudget), long-term memory (memorize writes entity-time-event items with edges; readMemory returns a node plus neighbors), and soft offloading (summarizeContext, LLMLingua-2 compressContext, and foldHistory, which collapses history into keywords plus a summary that search can recover). History-changing calls split a trajectory into snapshots used as the training unit.

SFT trajectories come from Qwen3.5-397B-A17B under a harness whose hints never enter the kept data. RL has two custom pieces. Context-aware partial rollout scores each management action by relative context-length change and entropy change versus the start of the trajectory, then spends leftover snapshot budget branching at the most sensitive edits. Fine-grained credit averages terminal rewards of every sampled continuation through a snapshot, then runs GRPO inside the snapshot group of one query.

Long-context QA starts from Qwen3-8B/14B and Gemma4-E4B-it. Deep search skips SFT and RL-tunes WebSailor-7B and WebExplorer-8B on 1,000 OpenSeeker items. Inference uses a 32K input cap.

Results

Long-context QA, mean of three runs:

MethodWindowNovelQA∞BenchLongMemEval-SBrowseComp+Avg
StateLM-8B-RL32K84.1573.0759.7346.4465.85
ContextPilot-8B-RL32K83.8875.2564.2754.1869.40
ContextPilot-14B-RL32K84.8181.0867.4055.5072.20
Qwen3.5-397B + tools32K91.9492.1383.6080.9687.16

The 8B-RL model is 3.55 points above same-size StateLM and 7.74 points higher on BrowseComp+. RL barely moves NovelQA, whose other split was in SFT, but adds 5.34 points on BrowseComp+, whose inputs average about 552K tokens. Bare Qwen3-8B at 128K averages 45.93; the same tools without training drop to 27.61.

On deep search, ContextPilot on WebExplorer-8B averages 50.10 versus 49.09 for SUPO and 47.92 for OpenSeeker trained on the same 1K items without context tools. Working context on BrowseComp stabilizes near 8K-10K tokens per turn, against a baseline that grows toward 30K. Adding planning, soft offloading, then long-term memory on the teacher model lifts the average from 77.89 to 87.16, and BrowseComp+ from 63.49 to 80.96. Entropy-only branching is unstable; adding context variation plus snapshot credit takes Qwen3-8B from 65.78 SFT to 69.40.

Why it matters

A 32K managed window beating a 128K raw backbone says the scarce resource is often how the agent edits its own working memory, not another context-length SKU. Teams already training ReAct search agents can drop the toolset and snapshot credit into an existing pipeline.

The gains are incremental. The deep-search edge over SUPO is about 1.5 points. The clearer win is a shorter working context and tool use that looks like management rather than retrieval-only. Code and weights are public.

Limitations

The authors say the toolset still misses some editing needs, hyperparameters were not searched hard, and evaluation never leaves long-context QA and deep search: no coding agents, no GUI agents. SFT used a constrained teacher; production will not. BrowseComp+ sits in the QA suite because it uses a fixed corpus with no live web, so it is not full internet search. Partial rollout costs extra samples, and that bill is not itemized.

Terms

Source

Related papers

All paper explainers