ReflectRec adds a verification loop to LLM-based recommendation
The arXiv paper proposes checking a recommender's written preference note by round-tripping it against user history, then regenerating weak notes and borrowing signals from similar users.
In plain words
Some recommenders built on large language models first write a short note about what a user likes, then use it to rank items, but nobody checks whether that note is accurate. ReflectRec tests the note by seeing if user history can be recovered from it. When the check scores poorly, the system rewrites the note and also looks at what similar users watched.
What was announced
The authors describe a preference-to-history round-trip verifier whose recall score measures how well a written preference can be recovered from history. Low scores trigger conditional regeneration, and the system selects among regeneration attempts. The pipeline also adds a Stage-1 semantic prompt and Forward-Look retrieval, which injects future watches of behaviorally similar users as a collaborative hint, addressing the paper's observation that chain-of-thought rankers never consult similar users.
On MovieLens-100K and MovieLens-1M, the full pipeline reportedly outperforms CoT4Rec-ICL by +16.25 and +4.46 percentage points on Hit@10, with gains across all ValScore buckets and three LLM backbones: Qwen3-30B, Qwen2.5-72B, and Llama-3.3-70B. The authors report a similar verifier-guided regeneration pattern on Amazon Electronics. These figures come from the paper's own abstract and have not been independently confirmed.
What it does not show
- Evidence is limited to the paper's abstract; full experimental setup, baselines beyond CoT4Rec-ICL, and ablations are not available for verification.
- Reported gains are the authors' own claims and have not been independently confirmed.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.