English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Envs-FORGE: Customizing Agent RL Training Environments with Mixed-Integer Programming

Forum topic · ✨步子哥 · 2026-08-17

Summary

Envs-FORGE introduces a mixed-integer programming (MILP) framework that customizes environment synthesis for agent reinforcement learning, replacing one-size-fits-all rewriting recipes such as Few-shot, Self-Instruct, and Evol-Instruct. The framework estimates each seed task's pass rate against the current policy, then selects one of six operations—three projection types (Increase, Reduce, Diversify) crossed with two evolution directions (In-depth, In-breadth)—to match the policy's capability frontier while covering target skill gaps and preserving verifier feasibility. A key concept is the bridge task: for seeds far beyond the policy's ability, Envs-FORGE creates intermediate-difficulty tasks aligned with Vygotsky's zone of proximal development. Evaluated with GRPO on Qwen models (4B-35B), Envs-FORGE gains 9.2 percentage points on tb-core (40.0% to 49.2%), 6.4pp on tb-2.0 (23.0% to 29.4%), and 3.7pp on SWE-bench Verified (73.4% to 77.1%), using the same 100 verified environments and 2.27M-2.88M synthesized tokens as baselines—showing the advantage comes from better strategy selection, not more data.

Overview

Envs-FORGE is a method for synthesizing training environments in agent reinforcement learning (RL). Rather than applying the same rewriting recipe to every seed task—as Few-shot, Self-Instruct, and Evol-Instruct do—Envs-FORGE estimates each seed's pass rate under the current policy and selects a customized rewriting strategy via mixed-integer programming (MILP).

Reference: arXiv:2608.14312 | Code: DataArc-SynData-Toolkit

The Problem with Fixed Recipes

Current environment-synthesis methods apply identical transformations regardless of whether a seed is too easy (90% pass rate) or too hard (5% pass rate). The pass-rate signal is ignored, so easy seeds get harder and hard seeds get even harder—or are diversified into tasks that may still be unreachable.

Envs-FORGE's core insight: estimate seed pass rate first, then decide how to rewrite it.

The 3×2 Action Space

Each seed task receives one of three projections:

  • Increase — raise complexity
  • Reduce — lower complexity, producing bridge tasks that step back from seeds beyond the current policy's reach
  • Diversify — preserve difficulty while changing the scenario
  • Each projection runs along one of two directions:

  • In-depth — deepen within the same domain
  • In-breadth — extend across domains
  • This yields 3 × 2 = 6 combinations. The choice is computed by an MILP solver.

    MILP as Strategy Selector

    Envs-FORGE models environment synthesis as a MILP that optimizes simultaneously for:

    1. Frontier matching — synthesized difficulty distribution should align with the policy's capability frontier 2. Skill coverage — synthesized environments should fill gaps in the target skill graph 3. Verifier feasibility — every synthesized environment must pass an executable test-based reward verifier

    Fixed recipes are recovered as special cases via action masks: Evol-Instruct always Increases; Self-Instruct always Diversifies. The solver outputs 100 verified environments that feed directly into GRPO training.

    Bridge Tasks and the Zone of Proximal Development

    When a seed's pass rate is far below the policy's ability, training on it yields no learning signal. Envs-FORGE's Reduce projection creates intermediate bridge tasks—stepping back from an unreachable seed to a difficulty the policy can actually learn from. This mirrors Vygotsky's concept of the zone of proximal development: learning happens at the edge of capability, not far beyond it.

    Results

    GRPO training on Qwen 3.5 35B:

    | Benchmark | Base | Envs-FORGE | Strongest Baseline | Gain | |-----------|------|-----------|--------------------|------| | tb-core | 40.0% | 49.2% | 46.8% | +9.2 pp | | tb-2.0 | 23.0% | 29.4% | 27.3% | +6.4 pp | | SWE-bench Verified | 73.4% | 77.1% | — | +3.7 pp |

    Across 4B–35B model sizes, tb-core improves by 6.8–9.2 pp.

    Crucially, all four synthesis methods produce 100 verified environments using 2.27M–2.88M synthesized tokens. Envs-FORGE's advantage comes from better strategy selection—not more data or compute.

    Costs and Limitations

    The paper reports MILP solver overhead transparently but does not fully quantify it. Appendix B provides a full audit trail of the branch-and-cut search, fallback boundaries, and solver decisions.

    Evaluation scope is limited to Terminal-Bench subsets (CLI tasks) and SWE-bench Verified (code repository repair). Generalization to web interaction, multimodal agents, and broader task types remains open.

    Broader Pattern

    Envs-FORGE is an instance of a recurring principle: optimization granularity should match the granularity of what is being optimized. Where Heddle and CodeRescue raise decision granularity from single calls to trajectory/recovery actions, Envs-FORGE raises environment-synthesis granularity from fixed recipes to per-seed customization. Fixed recipes are uniform processing; MILP is on-demand allocation.

    Resources

  • Paper: arXiv:2608.14312
  • Code: github.com/DataArcTech/DataArc-SynData-Toolkit
  • HTML: arxiv.org/html/2608.14312v1

Tags

#agent-rl#environment-synthesis#mixed-integer-programming#grpo#llm-training#reinforcement-learning#bridge-task#vygotsky

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178633600