rsi.watch

Xu Huang et al. (3 authors) · 2026-10-29

Try Before You Evolve: Ranking-Scored Trials for Recursive Self-Improvement of Recommender Systems

“With two sequential recommenders on three datasets and the same number of full retrainings, RecRSI significantly outperforms existing self-evolving agents in every setting, and on both Amazon datasets reaches in its first round a ranking quality that none of them reaches in five.”
Who
Xu Huang et al. (3 authors)
Date
2026-10-29
Evidence
Very low certainty Starts at Low: the developer’s own numbers.
  • ↓ indirect measure. The "existing self-evolving agents" in the quote are the authors' re-implementations, all run with one LLM, GLM-5.3, on two backbones starting from recipes with hidden size 50 and two layers. The authors write that "the effects of using different models have not yet been investigated." The short-trial signal was developed on archived SASRec edits from CDs and Office, the datasets of the headline comparison; the authors call that analysis "retrospective rather than a held-out test". “The three outcome-reward frameworks are re-implemented in the shared recommender harness rather than run from their original code.”
  • ↓ imprecise. Each method runs one five-round trajectory per setting with one shared seed, and the significance tests resample Test users ("5,000 Test-user resamples"). The paper reports no repeated searches. “Within a backbone, all methods share the data split, the initial checkpoint, the seed, and the training setup.”
how we rate evidence
Replication
No rerun found. The paper links no code; it refers to "our released results" for LastFM without a link. We found no rerun by another group. Searched up to 2026-10-30. all claims
Checked
2026-10-30, against the source page
Links
Source Paper

Also on the page

  • The paper lists the affiliation University of Science and Technology of China. Each round, an LLM, GLM-5.3, reads a diagnosis of the current recommender, its training code and a memory of earlier rounds, and proposes 3 code edits to the training recipe (loss, negative sampler, optimizer, prediction or architecture). Each valid edit gets a short trial: 4,000 training steps, about 15% of the step budget, from the same initialization, scored by full-ranking NDCG@10 on probe users. The edit with the top trial score is fully retrained and replaces the parent if NDCG@10 on a separate set of select users rises by more than a fixed margin.
  • After five rounds, Test NDCG@10 is 16.1% to 102.0% above the strongest of AlphaEvolve, DeepEvolve and Self-EvolveRec in each of the six settings (SASRec and HSTU on CDs, Office and LastFM). On SASRec and Office, for example: 0.0079 for the initial recipe, 0.0171 for Self-EvolveRec, 0.0242 for RecRSI. 23 of RecRSI's 30 retrainings are kept, against 37 of 90 for the three baselines.
  • The baselines are the authors' re-implementations in their own harness, with the same LLM, editable files and promotion rule. Under an equal budget of one GPU hour on SASRec and CDs, RecRSI's Test NDCG@10 is 17.1%, 24.6% and 8.7% above AlphaEvolve, DeepEvolve and Self-EvolveRec; "the numbers of LLM tokens are not equalized."
  • Choosing among the same three edits by other means (SASRec, CDs and Office): an LLM judge is slightly higher than RecRSI on CDs (0.0608 against 0.0594 NDCG@10, not significant; on CDs HR@10 the judge, 0.1005, and a warm start, 0.1003, are above RecRSI's 0.0973) and 15.7% lower on Office; RecRSI is 26.3% and 103.4% above random choice. On 38 archived edits, the short trial's correlation with the retraining gain is 0.46 and a raw gradient signal's is 0.49; the authors call this analysis "retrospective rather than a held-out test".
  • The first-round comparison in the quote is on select NDCG@10, the validation score that decides promotion, and on SASRec only: on Office against the others' five-round results, on CDs against their one-hour runs, in which the outcome-reward agents complete 11 to 20 rounds.
  • The authors' limits: "the main comparison fixes five rounds per trajectory", with Self-EvolveRec reaching its best model "within 6–11 iterations in its own experiments"; "longer searches on larger datasets remain to be evaluated." Only one LLM was used, and "whether the trial length established in the current experiments can be generalized to larger-scale datasets remains to be explored."

In the self-improving AI catalog

  • RecRSI: what it changes, whether the change lasts, and how it was tested

How it got here

Earlier work in this catalog that RecRSI cites, oldest first. Under each: where that work sits on the map, then where RecRSI sits (changes in bold), and what RecRSI's own paper says about it. Its whole family on the citation map

  1. 2025 AlphaEvolve uses it

    Prompts, memory · gone when the run ends · does not improve the improver

    “AlphaEvolve (Novikov et al., 2025) keeps an archive of scored programs and samples parents and inspirations from it; DeepEvolve (Liu et al., 2025) adds literature retrieval and reflection; and Self-EvolveRec (Kim et al., 2026) adds numerical diagnosis of the parent, critiques from simulated users, and co-evolution of its diagnostic code.”
    “The three outcome-reward frameworks are re-implemented in the shared recommender harness rather than run from their original code.”

  2. 2026 RecRSI

    Prompts, memory · gone when the run ends · does not improve the improver

Also cites Darwin Gödel Machine 2025

← Ledger Report an error