SWE-Pruner Pro: The Coder LLM Already Knows What to Prune

TL;DR. Coding agents drown in long tool outputs (file dumps, test logs, grep results) that bloat the context window and the KV cache. Prior context-pruning work (SWE-Pruner) bolted on a separate classifier to decide which lines to keep. SWE-Pruner Pro shows that is unnecessary: the agent's own internal activations, computed while it reads a tool output, already encode which lines are relevant. A tiny keep-or-prune head reading those activations saves up to 39% of prompt and completion tokens while preserving task quality, and on one backbone it even raises SWE-Bench Verified resolve rate by +3.8%.

flowchart LR
    T[Tool output<br/>file / test log] --> A[Agent reads it<br/>internal activations]
    A --> H[Small keep/prune head<br/>+ length-aware embedding]
    H --> K{Per-line<br/>keep or prune?}
    K -->|keep| C[Pruned context<br/>into KV cache]
    K -->|prune| X[Dropped<br/>39% tokens saved]
    C --> R[Next agent turn]
    classDef input fill:#dbeafe,stroke:#3b82f6,color:#1e3a8a
    classDef decision fill:#fef3c7,stroke:#f59e0b,color:#78350f
    classDef output fill:#d1fae5,stroke:#10b981,color:#065f46
    classDef warn fill:#fee2e2,stroke:#ef4444,color:#7f1d1d
    classDef aux fill:#e0e7ff,stroke:#6366f1,color:#312e81
    class T input
    class A aux
    class H aux
    class K decision
    class C,R output
    class X warn

What it is

Context pruning for long-horizon coding agents. When an agent calls a tool (open a file, run tests, grep), the raw output is often hundreds of lines, most irrelevant to the task. Keeping it all inflates the prompt, the KV cache, and per-turn cost. SWE-Pruner Pro prunes tool outputs inside the agent: a small head turns the agent's own per-line internal representations into a keep-or-prune label, with a length-aware embedding keyed to each tool output's line count.

Core novelty

The finding that the relevance signal is already present in the base agent's activations while it reads the output, so no separate code classifier is needed. This removes the extra model call and the train/serve mismatch that a bolt-on classifier introduces. It is the coding-agent analogue of the "the model already knows" line the wiki has tracked in other settings: TIP (2026-04-16) found most teacher tokens carry no distillation signal and only ~10% matter; Make Each Token Count (2026-05-12) used the model's own attention to drive KV eviction. SWE-Pruner Pro extends the pattern to agent tool-output context.

Key results

Up to 39% of prompt + completion tokens saved across two open-weight backbones and four multi-turn benchmarks, with bounded inference overhead.
On MiMo-V2-Flash, +3.8% SWE-Bench Verified resolve rate and +2.2 points on the long-context Oolong benchmark — pruning noise improved accuracy, not just cost.
No separate classifier model to train, serve, or keep in sync with the agent.

How it relates to prior wiki knowledge

Extends the context-compression thread: TokenPilot (2026-06-16) cache-efficient agent context, LongAttnComp (2026-06-02), LCLM end-to-end context compression (2026-06-09). SWE-Pruner Pro is the first to do it with zero added model by reusing the agent's own activations.
Confirms the "internal representations are a free signal" pattern (kv-cache.md concept page), joining attention-driven eviction work.
Industrial angle: this is directly deployable on top of any open-weight coding agent (SWE-Bench-style harness) with a small trained head, no architecture change.

Gaps

Tested on two open-weight backbones only; whether the internal-relevance signal transfers to frontier proprietary agents (which may encode relevance differently) is untested. The +3.8% resolve-rate gain is on one backbone (MiMo-V2-Flash); the effect on other models is preservation, not improvement. Length-aware embedding is keyed to line count, so it may misbehave on tool outputs with very long single lines (minified files, base64 blobs).

Raw source: HuggingFace Daily Papers 2026-07-21 · arXiv 2607.18213