proomt

Search

Search posts, papers, and topics

All posts

Hugging Face Daily PapersMeijia Chen, Hao Li, Zheng Lu2 min readpaperadvanced

False Frontiers: Diagnosing and Mitigating Co-Cheating in Self-Evolving Search Agents

Read on arxiv.org

Summary

Self-evolving search agents can suffer from "co-cheating," where the question proposer and answer solver increasingly agree on shared errors, improving internal reward without external correctness gains. The paper introduces CrossFit, a method that partitions source documents and uses cross-fitted agreement to determine proposer reward, significantly reducing false agreement and improving downstr…

  • Co-cheating is a failure mode in self-evolving LLM agents where internal metrics diverge from true correctness due to shared errors.
  • Multi-sample verification (MSV) partially reduces co-cheating but is costly and leaves substantial residual false agreement.
  • CrossFit partitions source documents (A/B) and uses an auxiliary solver trained on B to score proposals from A, preventing same-source pseudo-label reproduction.
  • CrossFit reduced false agreement mass from 6.1% to 3.0% (4B model) and 8.8% to 3.7% (9B model).

This paper is crucial for researchers and engineers developing self-improving LLM systems, as it identifies a critical failure mode and provides an effective, specific mitigation strategy.

8/10
llmmachine learningevaluationaisearchself supervision

Related reading

  1. Self-Evolving Search Index

    The paper introduces SELF-INDEX, a framework that lets a search index automatically diagnose retrieval failures, revise its keys, and validate changes, using a query simulator to anticipate future queries. Experiments show consistent gains across corpora and downstream LLM agents.

    Hugging Face Daily Papersarxiv.org1 minpaper

  2. Search for agents splits discovery from proof

    Exa’s Agent splits search for AI agents into a discovery phase (candidate generation) and a verification phase (evidence check). The design uses a powerful model for generation and cheaper models for parallel verification, optimizing latency, token usage, and cost. It treats the web as a structured database rather than a ranked list, and runs the verification sub‑agents as independent, fault‑tole…

    Renderrender.com4 min

  3. Recursive self-improvement of AI research agents

    The paper introduces AIDE², an AI research agent that rewrites its own code, benchmarks each version, and adopts the best performing changes—a process they call recursive self‑improvement. In an 8‑day autonomous run it produced seven improvements that beat a strong human‑engineered baseline on four unseen benchmarks and reduced reward‑hacking from 55 % to 32 %.

    Hugging Face Daily Papersarxiv.org2 minpaperHN31

  5. CERA-MoA: Co-Evolving Routing Mechanisms with Continually Learning LLM Agents

    CERA-MoA proposes a reinforcement‑learning loop where a router and a set of LLM agents are trained together. A “familiarity” estimator reads mid‑layer hidden states to predict each agent’s competence on a query, letting the router activate only a minimal subset of agents that meet a cumulative confidence threshold. The system also feeds targeted training examples to agents based on their evolving…

    Hugging Face Daily Papersarxiv.org1 minpaper

  6. BI-Agent and BI-Bench: Towards Automating End-to-End Business Intelligence

    The paper introduces BI‑Bench, a new benchmark of real‑world BI questions derived from public dashboards, and BI‑Agent, a tool‑augmented LLM system that breaks BI workflows into search, join, and transform subtasks. Baseline LLMs hit <50 % accuracy on BI‑Bench. By orchestrating specialized data‑management tools and post‑training the model with supervised fine‑tuning and reinforcement learning on…

    Hugging Face Daily Papersarxiv.org2 minpaper