- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper studies catalog-support drift in LLM-based generative recommendation: DPO and Softmax-DPO can improve relative preferences among candidate items without preserving the probability mass assigned to valid catalog items. Consequently, better candidate ranking can coexist with substantially worse free-generation validity.
The proposed DEHALL-SDPO augments Softmax-DPO with full-slate listwise policy ranking (LPR), primal–dual sequence-mass control (Dual Mass) using an SFT-calibrated budget, and on-policy prefix-support regularization (Prefix) based on a catalog trie. These objectives modify training while leaving inference unmasked. Experiments use Mistral-7B-Instruct-v0.3 on five recommendation datasets and evaluate candidate ranking and generation under multiple decoding protocols.
Critical strengths
-
The paper identifies a meaningful evaluation blind spot. Candidate ranking, relative preference margins, and free-generation validity measure different properties. The checkpoint-level measurements illustrate failures that ranking metrics alone can conceal, providing a useful perspective on training and evaluating generative recommenders.
-
The recovery in generation validity over the evaluated Softmax-DPO baseline is substantial. On Amazon CD, for example, greedy catalog violation decreases from 59.19% to 1.56%, while greedy Hit increases from 22.19% to 38.69%. These results demonstrate the practical importance of support preservation under the evaluated setup.
Critical weaknesses
-
Improved per-output correctness does not establish better recommendation-list quality. On Amazon CD, sample-10 mean Hit improves from 22.66% to 37.52% relative to SFT, but Any Hit decreases from 59.50% to 47.56%, and Unique decreases from 53.48% to 17.09%. MovieLens also exhibits lower Any Hit. These results indicate that improved per-output correctness accompanies a more concentrated output distribution.
-
The connection between the theoretical interpretation and achieved optimization behavior remains incomplete. The selected models do not satisfy the validation mass budgets. For example, the selected MovieLens model has mean negative-log slate mass of 0.02729 versus an SFT budget of 0.00840; Amazon CD has 0.09462 versus 0.05825. It is therefore unclear whether adaptive dual weighting provides more effective support control than an ordinary penalty.
Questions for authors
-
Which rows in Tables 1–2 use validation versus held-out test evaluation? Can all main methods be evaluated on the same independent test set under a unified selection protocol?
-
At the selected λr, does the complete method outperform Softmax-DPO + chosen-NLL, Prefix + chosen-NLL, and fixed-weight mass regularization? Can these comparisons control both candidate exposure and computation?
-
Why does Dual Mass fail to meet the SFT-calibrated budget? Is this due to incomplete optimization, a training–validation distribution difference, or another factor?