Microsoft reveals cost-effective method to enhance GPT-5.4-mini's reasoning through skill distillation

Microsoft researchers propose a novel approach to reduce costs and boost performance in AI reasoning by converting agent experiences into reusable skills, challenging traditional reliance on costly reasoning modes and demonstrating promising results across multiple domains.

Microsoft researchers say they can strip much of the cost out of test-time reasoning by turning past agent runs into a short reusable prompt, allowing a cheaper version of GPT-5.4-mini to match or beat its reasoning mode on some tasks instead of paying that premium on every run. The paper, submitted to arXiv on 8 August and listed for the COLM 2026 Efficient Reasoning Workshop, reports that the distilled-prompt approach closed 55% to more than 100% of the performance gap between non-reasoning and reasoning settings across four benchmarks, while cutting output-token use by 2.7 to 6 times and using no reasoning tokens at all. (arxiv.org)

The claim matters because the paper argues that reasoning mode is often re-spending computation on procedures the model has already had to discover before. In the authors’ framing, standard reasoning is “deep” search inside one episode, whereas their alternative is “wide” search across many previous episodes in the same domain. The resulting skill is attached to the system prompt as a fixed text prefix, which both the paper abstract and secondary summaries say can be cached rather than regenerated step by step. (alphaxiv.org)

The implementation is deliberately narrow. For each domain, the team collected 35 to 50 trajectories from the training split, then used a coding agent to inspect tool calls, observations, actions and outcomes, and to write analysis scripts looking for recurring failure modes, loops and win-loss contrasts. In the paper’s reported setup, the coding agent was Claude Code using Claude Sonnet 5. It then produced a human-readable skill file of roughly 40 to 130 lines of markdown. No model weights were updated, no fresh environment roll-outs were gathered for distillation, and the one-off production cost was reported at $1.28 to $2.44 per domain, broadly matching secondary coverage that rounded this to about $1 to $3. (arxiv.org)

The strongest headline result came on ALFWorld, a text-based household-task benchmark. There, GPT-5.4-mini in reasoning mode scored 0.713 with an average 3,723 output tokens per episode, while the no-reasoning model with a distilled skill reached 0.787 using 832 tokens. The same pattern appeared in tau-squared retail, where the skill-equipped version scored 0.408 against 0.350 for reasoning mode, at 565 tokens rather than 1,615. The picture was more mixed elsewhere: in SpreadsheetBench-Verified the distilled version rose from 0.447 to 0.560 but still trailed the reasoning setting at 0.613, and in tau-squared telecom it improved from 0.192 to 0.333 while remaining below the reasoning score of 0.450. (arxiv.org)

One of the more commercially interesting findings is that the distiller did not need reasoning traces to work. The paper says skills built only from non-reasoning trajectories remained competitive with those built from paired reasoning and non-reasoning corpora. In ALFWorld the paired corpus did somewhat better, at 0.813 versus 0.787, and retail also favoured the paired setup, 0.458 versus 0.408. But SpreadsheetBench-Verified moved the other way: the no-think-only skill scored 0.560, ten points above the paired version at 0.460. That suggests the valuable signal often lies in the record of what actually happened in the environment, not in a model’s internal narrative about why. (chapterpal.com)

The paper gives concrete examples of what is being captured. In retail support, it says one recurring error, calling an authentication tool with a fabricated argument, appeared in 59% of non-reasoning training roll-outs and accounted for 94% of observed tool errors. In ALFWorld, the authors describe a task requiring the agent to put a cool tomato in a microwave. Without the distilled rules, the model reportedly treated “cool” as a description rather than an action, then looped until it hit the step limit. With the skill inserted, it issued the cooling command and completed the task. The researchers say stall loops in ALFWorld fell from 28.7% to 5.3%, while missed transformation failures dropped from 35.9% to 11.5%. (arxiv.org)

The team also tested whether this passive distillation pipeline could outperform more active prompt tuning. Against GEPA, described in the paper as a state-of-the-art reflective prompt evolver, the distilled-skill approach scored higher on both tau-squared domains while costing less to produce. The reported comparison put retail at 45.8% against 39.2% and telecom at 32.5% against 30.8%, with production cost of $3.72 rather than $15.28. The authors attribute that gap to GEPA’s need for repeated optimisation roll-outs, whereas the distilled-skill method makes a single pass over logs that many agent systems already store. (arxiv.org)

The paper is more careful than the headline result might suggest. It says skills were distilled once per model-domain pair, so variance in the distillation process itself was not measured, and cross-model transfer was not tested. The remaining gap in telecom and spreadsheet work is also central to the authors’ argument: some tasks still depend on long, instance-specific chains of reasoning that a fixed prompt cannot anticipate. Secondary commentary has drawn the same practical conclusion, namely that skill distillation looks most useful for high-volume agent deployments with repeatable procedures, where a cheaper default agent can be augmented with cached domain rules and the expensive reasoning mode reserved for the harder minority of cases. (arxiv.org)