RecHarness: A Bandit-Routed Agentic Harness for Self-Evolving Recommender Systems¶

Fidelity: core mechanism. Paper results, local public-data measurements, and omissions are separate.

Paper information¶

Field	Details
Paper	arXiv 2607.29241
Organization	Huazhong Agricultural University (Kuaishou internship)
First public release	2026-07-31 (arXiv v1)
Authors' code	Public implementation
Adapter / method	`recharness`
Local reproduction code	`src/auto_research/reproductions/recharness/`

Original paper¶

Motivation and method¶

Define architectural changes as a finite set of arms. A bandit allocates trials from historical gains and uncertainty, then writes winning feedback into the next candidate round.

flowchart LR
 A["Candidate architectural arms"] --> B["Bandit-routed experiments"] --> C["Feedback update"]

Key figure from the paper¶

Figure 1: Overview of the paper's method and key components. From the original paper; copyright remains with the authors.

Core equations¶

\[ a_t=\arg\max_a[\hat\mu_a+\sqrt{2\log t/n_a}] \]

Paper results¶

The main online result is ADVV +2.08% on 10% traffic for seven days. Offline tables and online metrics use private or paper-specific protocols. The private paper protocol is not directly comparable with local MovieLens measurements.

Local reproduction¶

Local comparison: The baseline is the shared transition + content scorer. The treatment adds only recharness on the same users, items, full-catalog candidates, and seed.

MovieLens-100K, 220 users and 360 items, seed 42: NDCG@10 0.0540 → 0.0491 (-9.02%); Hit@10 0.1091 → 0.1000. Validation selects mixture weights only; the test set is not used for tuning.

auto-research reproduce --paper recharness --dataset-dir data --seed 42

See the fixed metrics.

Reproduction boundaries¶

The core mechanism runs on MovieLens-100K with the same split, fixed full-catalog candidates, and seed. Private features, the production foundation model, and online traffic are unavailable. The implementation has independent model state and scoring, rather than mapping a paper name to a shared weighted formula. See the adapter's omitted_core_components for omissions. Local changes and paper A/B gains are reported separately.