- Source
- arXiv
- Published
- Runtime
- 0:00
- Snippets
- 4
A conversation between
RecHarness: A Bandit-Routed Agentic Harness for Self-Evolving Recommender Systems
§02
Snippets
-
RecHarness splits optimization into two: a bandit algorithm picks *which direction* to explore next, while an LLM generates the *specific code change* within that direction.
Separating strategy selection from implementation stabilizes search; LLMs alone often waste budget on conflicting or incoherent ideas.
-
A jump-basin mechanism detects stagnation and activates a 'structural-jump' arm—larger, riskier architectural changes—to escape plateaus and sustain long-horizon exploration.
Prevents the system from converging prematurely; enables discovery of genuinely better model families, not just marginal gains.
-
On a large-scale short-video ads platform, RecHarness improved ADVV by 2.084%, Revenue by 0.534%, and Exposure by 0.559% in a 7-day online test.
Demonstrates real-world impact; automated optimization translates to measurable revenue and user-engagement gains without manual engineering overhead.
-
RecHarness achieves more stable performance improvements and uses limited trial budgets more effectively than LLM-only reasoning search across multiple tasks and model backbones.
Structured human-AI collaboration—routing + generation—outperforms end-to-end LLM reasoning; a blueprint for autonomous ML optimization.
§03
Synthesis
The Problem: Manual Trial-and-Error Wastes Engineering Effort
Building better recommender systems today still relies on engineers manually tweaking model architectures, loss functions, and training strategies. Even with LLM-based agents that can propose changes automatically, letting them freely choose what to optimize and how to do it produces erratic results when experiment budgets are tight. The authors find that unguided LLM exploration wastes limited trials chasing dead ends.
RecHarness: Splitting the Decision into Two Stages
The core insight is to separate optimization into two tightly coupled steps. A bandit router (a statistical decision-maker that learns from past failures and successes) picks the direction of the next change—for instance, "modify the loss function" or "adjust training hyperparameters." Only then does the LLM generate a concrete hypothesis and write the actual code edit within that direction.
This division of labor matters because the bandit router is stateless and interpretable: it tracks which modification categories have paid off before and allocates future trials accordingly. The LLM stays focused on engineering within a narrower scope, reducing hallucination and instability.
The Jump-Basin Mechanism: Escaping Local Plateaus
One hurdle in any automated search: local stagnation. If incremental edits stop improving validation metrics, the system risks spinning its wheels. RecHarness adds a "jump-basin" trigger that activates a special arm in the bandit—one dedicated to structural jumps (larger, qualitative changes like swapping a component entirely). When the bandit detects sustained flat performance, it can propose a structural leap rather than another small tweak. This lets the search escape plateaus and explore new regions of the design space.
Real-World Validation
The authors test RecHarness across multiple recommendation datasets and model backbones, comparing it to baseline LLM-reasoning search. The key results: more stable improvements and better sample efficiency when trial budgets are limited—a practical constraint in industry.
The strongest evidence comes from a 7-day A/B test on a large-scale short-video advertising platform. The model selected by RecHarness improved:
- ADVV (advertisement views): +2.084%
- Revenue: +0.534%
- Exposure: +0.559%
These gains are modest but real, and they matter at scale. For a platform with millions of users, even 0.5% lift on revenue translates to significant value.
Why This Matters
Recommender systems are expensive to optimize—each training run consumes resources and takes time. RecHarness automates the tedious search process while maintaining stability through the bandit router's principled, feedback-driven allocation. By decoupling direction selection from hypothesis generation, the system becomes both more reliable and more sample-efficient. The jump-basin mechanism adds a safety valve for escaping local optima, a persistent challenge in automated machine learning.
The online results suggest the method generalizes beyond benchmarks to real production settings, where model latency, data drift, and business metrics all matter. This makes RecHarness a practical tool for teams scaling recommender systems without proportional increases in engineering headcount.
Mine your own.
Lode is a workbench, not a feed. Paste a YouTube URL. The model proposes a transcript, a set of quote-grounded snippets, a synthesis essay, and the fan-out. You decide what stays.