LLMs Leverage Self-Hard Negatives For Recommendation
Researchers (v1 submitted Feb 19, 2026) propose ILRec, a preference fine-tuning framework that improves LLM-based recommender systems by extracting self-hard negative tokens from intermediate layers. The two-stage method uses cross-layer preference optimization and distillation, plus a lightweight collaborative-filtering module to assign token-level rewards and reduce false-negative penalties. Experiments on three datasets show ILRec enhances recommendation performance versus sequence-level offline negatives.
Key Points
- 1Introduce ILRec extracting self-hard negative tokens from intermediate LLM layers for fine-grained negative supervision
- 2Enable discriminative preference learning via cross-layer optimization and distillation to improve negative signal quality
- 3Provide token-level rewards using lightweight collaborative filtering to avoid over-penalizing false negatives in training
Scoring Rationale
Novel cross-layer negative mining increases practical relevance, but arXiv preprint lacks peer review and broader benchmarks.
Sources
Public references used for this report.
Practice with real Logistics & Shipping data
90 SQL & Python problems · 15 industry datasets
250 free problems · No credit card
See all Logistics & Shipping problems
