Contextual Information Policy Optimization for Search Agents¶

简体中文

Core-mechanism mini-suite. Runs the paper's search-agent RL operator; a small deterministic evaluation does not reproduce the full paper.

Paper information¶

Field	Details
Paper	arXiv 2608.06128
Organization	Beihang University
First public release	2026-08-06 (arXiv v1)
Authors' code	No public author code was found on 2026-08-09.
Adapter / method	`cipo`
Local reproduction code	`src/auto_research/agent_research/latest_20260809.py`

Original paper¶

Motivation and method¶

Topic: search-agent RL. Rewarding only the final answer can reduce retrieval to confirmation bias. CIPO identifies later actions influenced by external evidence, gives dense turn-level credit, and jointly optimizes it with global outcome reward.

Architecture¶

flowchart LR
    N0["Search request"] --> N1
    N1["External evidence"] --> N2
    N2["Evidence influence"] --> N3
    N3["Turn-level credit"] --> N4
    N4["Global outcome"]

Key figure from the paper¶

Figure 2: Core architecture, modules, and connections. From the original paper; copyright remains with the authors.

Core equation¶

\(A_t=\lambda A_t^{evidence}+(1-\lambda)A^{outcome}\).

Paper results¶

Across seven in-domain and out-of-domain benchmarks, the method reduces prior-driven reasoning and achieves best or competitive results on most tasks.

Local reproduction¶

See the stable metrics. Checkpoints and raw run directories are not committed.

auto-research agent-research --method cipo --benchmark planbench-mini --episodes 120 --seed 42

Local comparison: Compare cipo with a control that disables the mechanism in the same public mini-suite. Report only local artifact metrics; do not treat large-model or real-environment paper results as local gains.

Reproduction boundaries¶

Implements the paper-specific state, credit, or selection operator rather than merely renaming a method.
No original large model, proprietary environment, or expensive judge was run; these are not marked integrated.
cipo participates in unified evolve candidate discovery. Composed genomes remain subject to public evaluation and budget constraints.