简体中文 · English

Agent-G²: Online Gaussian estimation of guidance depth¶

Fidelity: core-mechanism mini-suite. Samples expert prefix depth for each task from a Gaussian updated online, without extra probe rollouts.

Paper information¶

Field	Details
Paper	arXiv 2608.23318
Institution	Baidu (first-listed affiliation of the first author)
First public release	2026-08-24 (arXiv v1)
Author code	Released: ZJU-REAL/Agent-G2
Local adapter or method key	`agent-g2`
Local reproduction code	`src/auto_research/agent_research/latest_20260825.py`

Original paper summary¶

Background and main changes¶

Hint-based agent RL retains an expert trajectory prefix and then lets the policy explore. A fixed depth ignores task difficulty; per-example probes waste rollouts. Agent-G² estimates the center and variance of a guidance band for each difficulty cluster from existing policy rollouts, then samples a prefix depth per task.

flowchart LR
  R["Existing rollouts"] --> C["Difficulty-cluster statistics Ak,Vk"]
  C --> G["N(μi,σi²)"]
  G --> D["Sample guidance depth"]
  D --> P["Expert prefix + on-policy continuation"]
  P --> U["GRPO + auxiliary BC"]

Key figure from the paper¶

Figure 4: The overall workflow, stages, and data flow in the paper. From the paper; copyright belongs to the authors. Click for the source.

Core equations¶

\[ \mu_i=\operatorname{clip}(\mu_{global}+\lambda(p_{target}-A_k),0,1),\quad \sigma_i=\max(\gamma V_k,\sigma_{min}), \]

\[ z_i\sim\mathcal N(\mu_i,\sigma_i^2),\qquad r_i=\operatorname{clip}(z_i,0,1). \]

Offline and online results¶

On ALFWorld, Agent-G² improves over the strongest hint-based, hint-free, and Aux-RL baselines by 2.3 / 3.9 / 7.4 points, respectively. Rollout cost is less than one third of per-example probing. The paper also evaluates WebShop and reports no industrial online A/B test.

Local reproduction¶

On PlanBench-mini, 120 episodes, seed 42: joint success 1.0000, 120 online guidance samples, average cost 0.8253. See metrics/planbench-mini-seed42.json and the batch index.

auto-research agent-eval --method agent-g2 --benchmark planbench-mini --episodes 120 --seed 42

Scope¶

The mini-suite verifies online per-cluster estimates, Gaussian sampling, and the expert-prefix/policy-continuation control flow. It does not run Qwen2.5, ALFWorld/WebShop, or actual GRPO.

Unified L2.1 isolated capability result¶

On the isolated toolroute-l2.1-v1 test split (60 episodes/seed, seeds 42/43/44), joint success is 0.7778, plan-step F1 0.8276, and failure recovery 0.6667. The public-evidence verifier uses only public reliability, reversibility, and cost metadata. See metrics and the L2.1 isolated capability protocol.