SCOReD: Student-Aware CoT Optimization for Recommendation Distillation

Abstract: Chain-of-thought (CoT) distillation in the recommendation domain is a necessary precursor to RL training, but raw teacher traces are ill-suited to this task. Large teachers approach the recommendation task with unusually high reasoning uncertainty, repeatedly rechecking their answers without revising them; supervised fine-tuning on such traces produces verbose students that never revise their initial guess. Furthermore, due to the novelty of the recommendation domain, the teacher's reasoning traces are highly out-of-distribution for the small student LLM. We propose Student-Aware CoT Optimization for Recommendation Distillation (SCOReD), a CoT optimization framework tailored to recommendation that first parses each teacher trace into typed segments and uses the student LLM's attention to score the importance of each segment. Then SCOReD dynamically selects a per-segment edit (KEEP / REWRITE / FUSE / PRUNE) based on the output length and comparative log probability lift of the answer given the edit as per the student. Therefore, SCOReD prunes redundant sections of the reasoning trace while preserving information-dense sections and adapts raw teacher traces to the student's output distribution. Training on SCOReD-optimized CoTs provides a cleaner learning signal to the student model and improves over baseline SFT by 1.56% NDCG and 1.9% Recall@5, while reducing reasoning length by 27.3%.

• The paper introduces SCOReD, a student-aware method that segments teacher reasoning, scores segment importance through student attention, and selectively keeps, rewrites, fuses, or prunes content.
• SCOReD improves NDCG from 0.7786 to 0.7908 and Recall@5 from 0.7108 to 0.7243 while shortening traces by 27.3% and reducing parse failures from 2.82% to 1.52%.
• The findings show that recommendation CoT compression must prioritize student likelihood, perplexity, and output-format reliability, while DAPO and OPSD provide little additional benefit after strong SFT initialization.

Motivation and problem setting

Generative recommendation systems such as OneRec, OneRec-Think, and GR2 increasingly cast reranking as an explicit reasoning task in which a LLM reasons over user history and candidates before emitting a ranked list. Because production recommenders require small models for latency and cost reasons, chain-of-thought (CoT) distillation from a large teacher into a small student is a necessary precursor to reinforcement learning. The authors observe that this distillation step is unusually difficult in the recommendation domain: labels derive from noisy user behavior rather than expert-verified derivations, so teacher traces exhibit a pathology rarely seen in math or coding — repetitive verification that almost never changes the answer.

The empirical characterization of this pathology is one of the paper's strongest contributions. Across 11,350 teacher traces generated by Gemma-4-26B on Amazon Beauty data, the teacher rechecks its initial answer 3.2 times per trace on average, and 79% of these rechecks leave the final answer unchanged. Segmentation over 289,541 annotated blocks shows that Verification (33.9%) and Ranking Construction (29.4%) together account for roughly 63% of all segments; raw traces contain $8.65 \pm 4.17$ verification stages each, yet only 47.15% of traces produce a final ranking that differs from the initial ranking, and 92.89% of individual refinement loops are no-ops. Notably, the number of refinement attempts is essentially unpredictable from lexical features: prompt length, purchase-history length, and purchase–candidate overlap correlate with verification counts at Pearson coefficients of only −0.08, −0.05, and −0.11 respectively. Under naive SFT, this becomes the dominant supervision signal: reason verbosely, recheck repeatedly, revise rarely.

Method

SCOReD builds on CRISP's attention-based compression but adds two recommendation-specific and student-aware components. First, each teacher trace is segmented by a second LLM pass into six typed stages — Purchase History Review, User Interest Modeling, Candidate Analysis, Intermediate Ranking, Verification, and Final Ranking — reflecting the observation that reasoning LLMs control their CoT structure in only 2.7% of cases, making structured reasoning a post-hoc extraction problem. Second, segment importance is scored using the average attention that each segment's tokens receive from thedelimiter under the *pre-SFT target student* ([Qwen3-0.6B](https://www.emergentmind.com/topics/qwen3-0-6b)), exploiting CRISP's finding that post-delimiter tokens attend almost exclusively to`, which thus acts as a learned summary representation of the trace.

Segments are bucketed into Low/Medium/High importance (thresholds at the bottom/top 10%), and each bucket defines a restricted action set: High segments may be Kept or Rewritten; Medium segments Rewritten or Fused; Low segments Fused or Pruned. The final edit is selected per segment by a reward combining the student's log probability of producing the teacher's final ranking conditioned on the edited prefix, a length penalty ($\alpha = 0.005$), and a perplexity penalty ($\beta = 0.1$) under the student:

$$R_i(a) = \log P(\text{answer} \mid C_i, \widetilde{S}_i^a) - \alpha\,\text{Len}(\widetilde{S}_i^a) - \beta\,\text{PPL}(\widetilde{S}_i^a \mid C_i).$$

The perplexity term directly addresses the out-of-distribution risk of compressed traces, motivated by evidence that high-perplexity tokens contribute to catastrophic forgetting during fine-tuning. This is the key contrast with prior CoT compression work (C3oT, TokenSkip, Compress-Distill), where compression is performed without regard to the student and has consistently traded accuracy for brevity — across all prior work, raw uncompressed traces remained the best SFT targets.

Main results

Training Qwen3-0.6B on SCOReD-optimized traces improves over raw-trace SFT by 1.56% NDCG (0.7786 → 0.7908) and 1.9% Recall@5 (0.7108 → 0.7243) while reducing average trace length by 27.3% (8.5K → 6.2K characters) and cutting parse failures from 2.82% to 1.52%, a relative reduction of about 46%. The paper claims, apparently for the first time in this domain, that compressed CoT can surpass raw traces in distillation. The 0.6B student also outperforms the much larger Qwen-3.6-35B-A3B reference (NDCG 0.7908 vs. 0.7879) and approaches its own Gemma-4-26B teacher (0.8030).

The comparison against a one-shot LLM summarization baseline is instructive and somewhat contradictory to human intuition. The summarizer produces shorter traces (4.4K chars) that read as coherent and pedagogical, but degrades performance severely (NDCG 0.7342) with an 8.74% parse-failure rate. A parse-failure taxonomy shows LLM-compressed traces induce disproportionate Duplicate Indices errors (18.4% of its failures vs. 0.8% for baseline SFT), indicating the compressed style interferes with even basic format adherence. Critically, on the correctly parsed subset (3,834 samples), all three students converge to nearly identical NDCG (0.8019–0.8035): the headline gains of SCOReD over one-shot compression stem primarily from output-format reliability rather than ranking quality per se. The paper states this plainly — minimizing trace length alone does not yield an effective distillation target.

Post-SFT optimization

The authors additionally evaluate DAPO with a verifiable NDCG reward and On-Policy Self-Distillation with privileged information (OPSD), both initialized from the SCOReD SFT checkpoint. Standard on-policy distillation is inapplicable here because the stronger teacher (Gemma-4-26B) does not share the Qwen student's vocabulary, while the vocabulary-compatible Qwen-3.6-35B-A3B is not consistently stronger on this task; OPSD sidesteps this by conditioning the same model on a privileged reference ranking constructed by stably moving ground-truth items to the front of the teacher ranking.

Both methods fail to improve the SFT initialization. OPSD checkpoints show small, non-monotonic fluctuations, with NDCG declining from 0.7923 to 0.7878 and parse failures rising from 1.39% to 1.81% by checkpoint 100. DAPO training oscillates within a narrow NDCG range over six epochs. The authors attribute this to the SFT student already approaching teacher-level performance, leaving sparse, low-variance rewards with limited positive advantage. This is an honest negative result with practical weight: it suggests that for small-student reranking distillation, the marginal return of post-SFT policy optimization may be negligible once the distillation target itself is well constructed.

Limitations and open questions

Several constraints bound the generality of these findings. The evaluation uses a single dataset (Amazon Beauty) with a fixed candidate set of ten items and three ground-truth items per query, and a single teacher–student pair (Gemma-4-26B → Qwen3-0.6B); whether the redundancy statistics and the superiority of student-aware editing transfer to larger candidate pools or other domains is untested. The segmentation pipeline depends on LLM-based annotation, and 2,201 training samples were discarded as unparseable, introducing potential selection bias. Hyperparameters $\alpha$, $\beta$, and the importance thresholds were set without an ablation reported in the main text. The failure of DAPO and OPSD is explained post hoc via reward sparsity, but alternative reward designs (e.g., process-level or format-aware rewards) remain unexplored. Finally, since correctly parsed samples perform equivalently across compression strategies, an open question is whether SCOReD's ranking-quality advantage exists at all beyond format reliability, or whether a cheaper intervention targeting output formatting alone could capture most of the benefit.

Conclusion

SCOReD demonstrates that recommendation CoT traces require domain- and student-aware treatment before distillation: teacher traces are dominated by unproductive verification loops, and effective compression must be guided by the student's attention, likelihood, and perplexity rather than by length reduction alone. The resulting 27% shorter traces improve NDCG, Recall@5, and format reliability simultaneously, contradicting the established pattern in which any CoT compression degrades downstream performance. The equally informative negative results on DAPO and OPSD indicate that well-targeted SFT data may saturate the gains available to post-hoc policy optimization in this regime.