Towards Efficient Reasoning in LLM-Based Recommender Systems via Model Merging
Linh Dieu Le, Tong Chen, Shazia Sadiq, Hongzhi Yin, Ming Jin, Junliang Yu
cs.IR, cs.AI
2026-08-11
Merging a slow- and fast-thinking recommender model with per-attention-head coefficients cuts reasoning length up to 24.3% while slightly improving rating accuracy.
LLM-based recommenders come in two flavors. Slow-thinking models write out a chain of reasoning before scoring an item, accurate but verbose, burning two to three hundred tokens per prediction. Fast-thinking models score directly, cheaper but slightly worse. What production wants is the middle ground: keep the accuracy, cut the verbosity.
Existing compression routes either retrain the model (expensive) or cap tokens and prompt for brevity at inference time (brittle, hard to scale). This paper takes the model-merging route: fuse the slow and fast models in a shared parameter space, with no training.
Conventional model merging applies one uniform coefficient across whole layers or models. REAM drops the granularity down to the individual attention head. The authors find reasoning behavior concentrates in a sparse subset of heads, and since heads serve different roles, per-head weighting is more faithful.
Each head gets its own coefficient from three signals:
The three signals feed a constrained optimization whose solution is water-filling in form: heads that matter most to reasoning and are least sensitive to change get touched lightly; the rest absorb more of the fast model's conciseness.
On Amazon Book, Music, and Yelp (Qwen2.5-3B-Instruct base; slow = RecZero, fast = TALLRec):
| Dataset | Reasoning length down | MAE | RMSE |
| Book | 24.3% | 0.6338 | 0.9116 |
| Music | 21.1% | 0.5348 | 0.8324 |
| Yelp | 17.9% | 0.7564 | 1.0649 |
Shorter traces and accuracy did not drop; it nudged up. Book MAE went from RecZero's 0.6650 to 0.6338. A causal check ablated the top 5% critical heads and saw KL/JSD divergence jump 3.29 to 11.57x, far above random ablation, confirming the signals locate the core of the reasoning.
Model merging is a train-free, plug-and-play direction, and taking it to attention-head granularity is a genuine step. For recsys practitioners it means deploying slow-thinking recommenders more cheaply without a retrain. The per-head merging idea is not bound to recommendation; it applies anywhere you want to blend two models that behave differently.
The authors concede that at Qwen2.5-7B, REAM (MAE 0.7910) still trails pure slow-thinking (0.7626); selective merging does not fully close the gap at larger scale. Evaluation is rating prediction (MAE/RMSE) only, with no Top-K recall metrics. Computing the three criticality signals needs a calibration set, an extra step in the pipeline.