When a frontier reasoning teacher is distilled into a smaller deployable student, which reasoning capabilities are most often lost or weakened, even if the student still mimics chain-of-thought formatting?
Distilled students keep surface CoT format but often lose hard-problem depth, adaptive thinking calibration, and out-of-distribution robustness. Easy metrics hold while competition-level accuracy drops.
Copying a chess master's written notes into a smaller player's notebook does not make them a master. The student learns how the notes look (headings, step labels). But they do not learn when to think longer, or how to handle weird board positions that were never in the notes. Distillation works like that. You get cheaper and faster answers on easy problems. But skill on the hardest puzzles, and knowing when to spend extra thinking, usually fade first.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Reasoning distillation is central to deployable reasoning models such as the DeepSeek-R1-Distill family. Yet teams routinely over-read CoT formatting as capability parity. What erodes matters because product and eval design depend on honest tail expectations.
What distillation actually transfers
Distillation (SFT on teacher traces, sometimes multi-stage) copies token sequences the teacher emitted on a finite prompt set. Students learn delimiters, step labels, language, and heuristics that are frequent in training logs.
They generally do not inherit the teacher's RL policy, online verifier loop, or adaptive stop rules. The exception is when those behaviors are consistently expressed in traces and the student capacity supports them.
Report tail metrics separately: AIME, LiveCodeBench hard, GPQA diamond. A blended MATH average hides distill gaps that matter for "reasoning SKU" marketing claims.
Capability vs format eval: run the same prompts through teacher and student, and have blind human raters score trace quality. Format scores correlate, but correctness diverges on the tail.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
Real products, models, and research that use this idea.
- DeepSeek-R1-Distill-Qwen-32B trails full R1 on hardest AIME subsets despite similar CoT style
- OpenAI's smaller reasoning models trade peak reasoning for lower cost. OpenAI does not publish how they are trained
What an interviewer would ask next. Try answering before peeking at the approach.
QCan student-side GRPO recover tail performance after distillation?
Partially. Student RLVR is cheaper than training the teacher, but it needs verifiers and compute. It does not fully close the gap on the hardest tasks.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Assuming matching CoT formatting means the student retained the teacher's reasoning depth on competition-hard tasks.
60 second bullets to scan on the way to the call.
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.