Why Pay for Long Reasoning When Rewrite Do Trick? Causal Effects of Reasoning History Compaction on Future Generation and Agent Cost

Why Pay for Long Reasoning When Rewrite Do Trick? Causal Effects of Reasoning History Compaction on Future Generation and Agent Cost. By Anton Hristov, from the Technological School Electronic Systems. Published on the second of November.

Reasoning makes language models better at difficult problems, and long reasoning traces make that improvement expensive. Efficient reasoning would express the useful analysis in fewer tokens, with less repetition and less text that adds little to solving the problem. Proprietary models have made progress toward this, but their scratchpads are hidden, which limits direct study of how it is done. The question taken up here is whether open-weight models can be led to reason less verbosely, without retraining, by changing the form of their own earlier thoughts. The answer reported is that, usually, they can. The model is not changed. Only the record of what it has already thought is, and that alone moves how much it goes on to think.

The intervention rewrites prior reasoning into compact state records, at the point before pending tool results are revealed. Generation then continues from the original history and from the compact one, and the two continuations are compared. That pairing is what lets the effect be read causally: one model, one moment, two versions of its past. Across the tested models, accepted compact histories usually shorten the reasoning that follows. A coding case study adds a narrower pattern: less repeated verification after a candidate solution that both histories share.

The claim comes with conditions. The shortening is reported for accepted compact histories, and as usual rather than universal. In software-agent deployments, compaction can reduce total spending, but task quality depends on the target model and on the compressor doing the rewriting, and so does the time taken once the rewrite itself is counted. A shorter future does not by itself make the whole run cheaper or better, since the rewriting has a cost of its own. Nothing is retrained, so the effect concerns open-weight models as they already are.

Reasoning history emerges as one lever on how many reasoning tokens a model spends next. Because the outcome shifts with target and compressor, the author argues for compaction policies fitted to specific models, and judged on both efficiency and the capability they retain.