TL;DR for operators
Generated explanations of customer preferences should be evaluated by whether they improve recommendations, not simply by how convincing they sound. MUSE, a reasoning-distillation framework, measures how each explanation changes the observed next item’s position in a full-catalog ranking and uses that measurement to select training examples.1
On Amazon Video Games, replacing utility-guided selection with random selection lowers Hit Rate@10 from 0.1020 to 0.0844 under the Qwen2.5-3B student backbone. The framework also combines heterogeneous teachers with student-generated candidates and trains reasoning generation separately from its downstream utilization.
For e-commerce recommendation developers, the approach offers a way to connect language-model supervision to measurable ranking performance. Teacher reasoning is cached during training, avoiding teacher calls at inference. However, better offline ranking does not establish truthful explanations, higher conversion, or favorable production economics.
A convincing preference explanation may not improve recommendations
A shopping recommender can examine a customer’s purchase history and generate several plausible explanations of what that customer might want next. One explanation might emphasize brand loyalty; another might identify an emerging interest in a product category. Both can sound reasonable while producing different rankings.
The difficulty is deciding which explanation deserves to influence the recommendation model. Linguistic fluency provides little direct information about whether the explanation helps the system recover the customer’s next selected item.
This matters when organizations use powerful language models as teachers for smaller recommendation models. Reproducing a teacher’s reasoning can transfer its language patterns without necessarily transferring information that improves item selection.
The MUSE study addresses that mismatch by treating downstream ranking improvement as the criterion for selecting reasoning trajectories. Its central contribution is not simply producing more explanations, but determining which generated explanations provide useful supervision for the recommendation task.
Measure what reasoning changes in the ranking
MUSE introduces Rank-Gain Utility Assessment (RGUA).
For each training instance, the system knows the user’s interaction history and the held-out next item. It first ranks the complete catalog without an additional reasoning trajectory. It then repeats the ranking with a candidate trajectory included.
The utility score is:
Here, $\rho_i(\varnothing)$ counts the catalog items ranked above the target without reasoning, $\rho_i(r)$ counts them with reasoning trajectory $r$, and $|\mathcal{V}|$ is the catalog size.
A positive score means the explanation moves the observed next item higher. A negative score means it pushes that item lower. Normalization makes the rank change relative to catalog size.
MUSE selects the highest-utility candidate as preferred reasoning and the lowest-utility candidate as dispreferred reasoning. Both become training signals.
The component ablation provides direct evidence that this selection rule matters. On Video Games with Qwen2.5-3B, full MUSE achieves H@10 of 0.1020, against 0.0844 when trajectories are selected randomly (Appendix B.4, Table 8). This is a 0.0176 absolute difference, approximately 20.9% relative to the random-selection variant.
The interpretation has an important boundary. RGUA measures usefulness under the current recommendation encoder and an observed target item. It cannot independently establish whether an explanation accurately describes the customer’s actual motivations. A trajectory could improve ranking while remaining an unreliable account of why the customer made a choice.
Three teachers supply candidates; the student supplies new ones
Selecting the best candidate is only effective when the candidate pool contains useful alternatives.
MUSE starts with three heterogeneous teacher models: DeepSeek-V3.2-Thinking, Gemini-3-Flash-Preview, and GPT-5-Mini. Their generated trajectories are cached. During a two-epoch warm-up, the student learns from selected teacher reasoning while the recommendation encoder remains frozen.
The next stage introduces progressively sampled student trajectories. Exploration increases toward six student-generated candidates by epoch ten.
This design addresses a limitation of fixed-teacher distillation: different teachers may produce better reasoning for different recommendation instances, while a student adapted to the recommendation task may eventually generate candidates that its original teachers did not provide.
The study tests this idea with a matched-candidate comparison against MTMR, which increases teacher-generated trajectories without student self-exploration. With total candidate counts held equal, MUSE leads on 14 of 18 reported metrics across three datasets (Section 4.4.3, Table 6).
The four exceptions matter. On Video Games, MTMR leads on NDCG@5 and Hit Rate@5; other exceptions occur on CDs and Vinyl and Musical Instruments. Thus, student exploration contributes beyond candidate quantity on most evaluated metrics, but its superiority is not universal.
The increasing selection frequency of student-generated trajectories during training offers additional evidence that the student’s contributions become more useful. It does not establish that the student surpasses its teachers on independent reasoning-quality benchmarks.
Generating helpful reasoning and using it require different training
After selecting preferred and dispreferred trajectories, MUSE optimizes two distinct components.
Contrastive Trajectory Distillation (ConDistill) trains the student generator to assign higher length-normalized likelihood to preferred reasoning than to dispreferred reasoning. The required likelihood separation increases with the measured utility gap. This supplements supervised fine-tuning on preferred trajectories.
Contrastive Reasoning Utilization (ConUtil) trains the recommendation encoder. It penalizes cases where the representation conditioned on dispreferred reasoning matches the target item more strongly than the representation conditioned on preferred reasoning.
The distinction matters because a generator can learn to produce task-relevant text without the ranking model effectively exploiting that information.
The component-removal ablations quantify the separate contributions on Video Games with Qwen2.5-3B (Appendix B.4, Table 8):
| Training configuration | Hit Rate@10 |
|---|---|
| Complete MUSE | 0.1020 |
| Without ConDistill | 0.0847 |
| Without ConUtil | 0.0915 |
| Without student exploration | 0.0998 |
| Random trajectory selection | 0.0844 |
Removing ConDistill produces a larger deterioration than removing ConUtil in this particular comparison. These are removal tests within the complete architecture, not estimates of each mechanism’s independent contribution in every possible configuration.
Other datasets provide qualifications. On CDs and Vinyl, the variant without ConUtil achieves H@20 of 0.0825 against 0.0804 for complete MUSE. Component contributions should therefore be assessed across the ranking cutoffs that matter to the application.
The benchmark improvements are measurable, but bounded
The main comparative evaluations cover three Amazon datasets, two student backbones, and established sequential, LLM-based, and reasoning-augmented recommenders. Evaluation uses leave-one-out prediction against complete item catalogs, reporting Hit Rate and NDCG at K=5, 10, and 20.
For Qwen2.5-3B, selected NDCG@10 results against the strong R²ec reasoning-recommendation baseline are:
| Dataset | R²ec | MUSE | Relative gain |
|---|---|---|---|
| Video Games | 0.0515 | 0.0536 | +4.1% |
| CDs and Vinyl | 0.0302 | 0.0348 | +15.2% |
| Musical Instruments | 0.0195 | 0.0255 | +30.8% |
Source: Section 4.2, Tables 2–3; Appendix B.3, Table 7. Relative gains are calculated from reported mean scores across three runs.
MUSE also improves the reported overall ranking metrics with Gemma-2-2B, although some differences are modest. On Video Games, Gemma H@10 rises from 0.0890 to 0.0902 against R²ec.
The results support the framework’s offline ranking effectiveness within these experimental settings. They do not establish statistical superiority for every comparison: the reported evidence includes three-run standard deviations but no comprehensive significance analysis.
Production value depends on the full recommendation workflow
For recommendation operators with historical next-item labels, MUSE suggests a development procedure: generate diverse candidate reasoning, evaluate its incremental ranking contribution, retain beneficial and detrimental examples, and separately train the generator and downstream encoder.
This is a Cognaptus inference from the architecture and ablations, not a demonstrated production outcome.
The inference architecture is relevant to deployment costs. MUSE invokes neither its teacher models nor RGUA at serving time. Instead, the student generates one trajectory, which the recommendation encoder combines with user history to rank catalog items.
That eliminates online teacher dependence, but not reasoning-generation latency. The reported efficiency analysis finds competitive inference latency, while R²ec remains faster.
Several questions therefore remain outside the study’s evidence:
- Commercial effectiveness: Offline Hit Rate and NDCG do not measure conversion, revenue, retention, or customer satisfaction.
- Explanation reliability: Rank improvement does not verify factual correctness or faithful representation of customer preferences.
- Operating economics: Cached teacher generation, repeated full-catalog candidate evaluation, student sampling, retraining, and catalog updates still incur costs that are not comprehensively quantified.
The evidence is also confined to three Amazon product categories. Different catalog sizes, interaction densities, languages, and user behavior could change the relative value of reasoning-guided selection.
A deployment decision would require product-specific offline comparisons, separate explanation-reliability checks where explanations are exposed or relied upon, end-to-end cost measurement, and online experimentation against the existing recommender.
Better supervision starts with a better definition of usefulness
MUSE demonstrates that selecting generated reasoning by its effect on recommendation ranking can improve a smaller model’s training signals. Heterogeneous teachers and student exploration expand the available alternatives, while dual-level contrastive learning addresses both generating beneficial reasoning and incorporating it into ranking representations.
Its broader methodological lesson is that the quality of an intermediate AI output should be judged against the downstream task it is meant to improve.
For recommendation systems, target-item rank gain provides a concrete way to apply that principle. It is a useful training criterion, not a substitute for evidence that explanations are faithful, customers benefit, or the complete system is economical to operate.
Cognaptus: Automate the Present, Incubate the Future.
-
Liangbo Ning and Siyuan Liang and Yi Zhou and Yifeng Zhang and Qing Li and Wenqi Fan and Dacheng Tao (2026). Multi-Teacher Reasoning Distillation for LLM Recommendation with Self-Exploration. arXiv:2611.02457. https://arxiv.org/abs/2611.02457 ↩︎