Self-edited context lifts Qwen3.5-9B from 28.8% to 42.5% on BrowseComp-Plus

In this design the tuning work moves into the prompt. The in-context learning variant gets its context strategy as a natural-language instruction. It then refines that strategy through an iterative skill-optimization loop ^[6]. Users can steer context management by telling the agent which strategy they want, and the model can evolve a skill document of context-management procedures to reuse later ^[5]. For a team that maintains a compaction pipeline, the thing under review becomes a prompt plus a document the model edits itself. The researchers say learned strategies may surpass "existing human priors" ^[19], a polite name for the predefined rules compaction usually runs on ^[3].

The reinforcement-learning figure is internally consistent. 42.5 minus 28.8 is 13.7 points, and 13.7 over 28.8 gives the 47.6% relative gain InfoQ reports ^[18]. The InfoQ summary does not say which baseline the zero-shot gains are measured against, or how the FLOPs counts treat cache misses. For the numbers to transfer, a team's tasks have to resemble the benchmarks, two of which run for 12 and 24 hours ^[7][8]. On the 24-hour multi-repository agent-swarm task, zero-shot CLMs reported 65% greater improvement at the same compute ^[8]. @Rennix7t wrote on X that "the accuracy and computational power results in the paper come from specified tasks" ^[14].

Serving cost is where we would push hardest. Combinatorilliance argued on Reddit that the approach is not production-ready because rewriting the context invalidates the model cache ^[16]. The researchers' caching trick "isn't ideal either and has some drawbacks that need to be taken into consideration," the same commenter wrote ^[17]. We'd expect the penalty to grow with how early an edit lands, since everything after the changed message has to be processed again. We think that favours fixed rules on cost. A compaction rule changes the prefix only when it fires. A model allowed to rewrite old messages ^[4] can change it on any turn.

The researchers flag the security cost themselves. The editable context "can become another channel through which prompt injections or self-generated instructions persist across turns," they wrote ^[12]. They also concede that more control over context does not necessarily produce better behavior, because a model can still choose badly what to keep, change or discard ^[13]. We think hand-built context pipelines stay the right production default on this evidence. They are also the control group a team needs to tell whether a model-managed context saves more compute than its cache misses cost on real work. @omarsar0 called the approach an interesting research direction but voiced reservations about trusting a model to manage its context end-to-end ^[15].