DeepSearch QRAE — Attention-entropy evidence scoring — What does it mean?
The news. On November 5, 2026 the QRAE paper was reported as showing that a DeepSearch agent can compress its growing retrieval context using only attention signals from a model it already runs. Across multiple benchmarks it reports about 45% fewer tokens, 30% lower end-to-end latency and a 2.35 percentage-point average gain in task success, and the authors say it carries over to a model of nearly 100B parameters. The full scoring formula is not reproduced in the summary this article draws on, so the details below stay at the level of the mechanism.
Picture the researcher again. A librarian keeps handing over photocopies, and a careless researcher stacks them all on the desk until the useful pages are buried. A DeepSearch agent has the same problem: every round of search adds chunks to a context window that is already a scarce resource. Long contexts cost tokens, slow every later step, and are one of the named failure modes of context. Typical remedies, such as summarizing old results or asking a model which chunks to keep, use a separate generation.
QRAE works the way a careful researcher would with a blank page for comparison. It reads each chunk twice with the same model: once with the question, once without. The two reads are a paired prefill, and the second, query-free read is the control. Subtracting the control removes the attention response that any text would trigger, so what is left is the part of the response caused by the question. The method then uses a few evidence-assessment heads, a sparse set of attention heads whose pattern tracks relevance, and measures a normalized entropy of that leftover response inside each chunk: how widely it spreads over the chunk's tokens. The exact way entropy and strength combine into a score is not stated in the source summary.
Scoring does not force a keep-or-drop decision. Chunks with a high score enter the whitelist, which stays in context across later rounds. Chunks the score cannot settle go to the candidate cache: uncertain chunks stay outside the reasoning context but are not thrown away, and the source says they remain retrievable from the cart. This is the same instinct as chunking and retrieval in a RAG pipeline, applied a second time, inside the agent's own working memory. The scoring runs on bounded batches of chunks.
Worked example (illustrative). Suppose an agent runs 5 search rounds, retrieving 8 chunks of about 1,000 tokens per round. Keeping everything gives 5 × 8 × 1,000 = 40,000 tokens of retrieval context. If the score whitelists 2 chunks per round and parks the other 6 in the candidate cache, the reasoning context holds 5 × 2 × 1,000 = 10,000 tokens, a 75% cut. The paper's reported average is 45%, smaller than this made-up case, which shows the saving depends on how many chunks the score lets through and on everything else that sits in the context. The paired prefills are extra work too, since each scored chunk is read twice, once with the question and once as the control.
| Approach | How it picks evidence | What it needs |
|---|---|---|
| Keep everything | Nothing is dropped; context grows each round | No extra work, but a large context |
| Summarize or filter with a model | A separate generation rewrites or ranks results | An extra generative pass per round |
| QRAE | Paired-prefill attention scores; whitelist plus candidate cache | Two prefills per chunk, no extra generation (reported by the authors) |
Goes deeper in: AI Agents → Context Engineering → Context as Scarce Resource
Related explainers
- UNREAL: model-native evidence selection — another way a frozen model ranks evidence from its own internal states.
- RIKA: evidence-based memory admission — the same question of what deserves to stay in an agent's context, asked about memory notes.
- RECAST: adaptive evidence routing — a trained router that builds evidence round by round.