GENERATIVE · 2026-10-26

Generative — Paper Daily, 2026-10-26

The generative track has 2 papers today. TCBSA tackles generative recommendation's beam search pruning away emerging interests too early: before each pruning step it borrows a classic scoring model's item scores as prefix bonuses, without retraining, and has a deep dive. MPQ is a post-training quantization solver that keeps several candidate values per weight block before committing, which is more accurate at low bit widths. Teams retrieving with semantic IDs or shrinking generative models for deployment should read both.

4.45

1. From Item Scores to Semantic Tokens: Test-Time Collaborative Adaptation for Generative Recommendation

In generative recommendation with semantic IDs, the model writes an item's code one token at a time, and beam search keeps only the highest-scoring prefixes at each step. Once a prefix is pruned, every item beneath it can never be generated. The problem is sharpest when a user has just developed a new interest that shows up in only one or two recent interactions. This paper proposes TCBSA (test-time collaborative beam search adaptation), which turns a scoring model's item scores into bonuses on code prefixes before every pruning step. The scoring model scores the whole catalog and keeps the top few hundred items; each item's margin above the cutoff is spread over every step of its code. Before each beam search ranking step, these bonuses are added to the path's cumulative score. Neither model is retrained. Added to TIGER, it raises NDCG@10 by a relative 4.9% to 30.4%, depending on the public dataset. With a narrower beam, it is also more accurate and faster than TIGER with a wider beam.

📖 Read the deep dive →
3.05

2. MPQ: A Message-Passing View of Post-Training Quantization

When an LLM is quantized after training to very low bit widths, common solvers such as GPTQ decide block by block which codebook value each group of weights takes. Once a block is fixed it never changes, and near-tied alternatives are dropped. This paper proposes MPQ (message-passing quantization), which keeps probabilities over each block's candidate values and commits only at the end. It alternates between two steps. One uses the layer's output error to estimate each block's weights as continuous values along with how reliable that estimate is. The other gives each candidate a probability: candidates closer to the estimate score higher, and the less reliable the estimate, the more even the probabilities. After a few rounds each block takes its most probable value, and blockwise coordinate descent polishes the result. On Llama, Qwen and Mistral models it improves perplexity and zero-shot accuracy, with larger gains at lower bit widths. With Llama-3-8B at 2 bits, its perplexity is about 11.6, against 12.76 for the baseline.