---
title: "CalibForge：通过对抗性求解器校准扩展可学习终端任务"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2608.06352v1"
permalink: "https://feed.astralor.com/items/itm_a6532616d3e4c3cc6e27d2cb"
published: "2026-08-06T17:53:18.000Z"
score: 75
channel: "ai"
tags: ["Agent 训练数据合成","arXiv","CalibForge","Terminal-Bench 2.0","求解器校准"]
---

# CalibForge：通过对抗性求解器校准扩展可学习终端任务

## AI 点评

> 这篇论文值得关注，因为它把任务难度从“人来定义”转成“求解器行为来定义”，对数据合成自动化很有启发。做 Agent 训练数据、尤其是终端操作类 Agent 的团队，应该仔细看看它的校准策略和迁移效果。

## 速览

CalibForge 提出一种自主终端任务合成系统，利用验证后的求解器行为对候选任务进行对抗性校准，构建求解器相对的“可学习区”。多求解器校准针对异构求解器池的分歧，对比校准则指定强过/弱败关系。研究构建了 5,431 个校准任务，相比仅作者验证或单求解器反馈带来更有效监督，在 Terminal-Bench 2.0 上达到 32.58% 和 47.57%，相对基线最高提升 24.71 个百分点，并在 SWE-bench Pro 和 Doc2Repo 上分别提升 27.68 和 30.04 个百分点，证明求解器相对可学习性可生成有效且可迁移的 Agent 训练数据。

## 关键点

- 提出对抗性求解器校准机制，利用多求解器分歧与强过/弱败对比关系定义任务难度。
- 构建 5,431 个校准终端任务，显著优于仅作者验证和单求解器反馈。
- 在 Terminal-Bench 2.0、SWE-bench Pro 和 Doc2Repo 上分别带来最高 24.71、27.68 和 30.04 个百分点的提升。

## 全文

Training terminal agents requires executable and verifiable tasks that are not merely solvable, but appropriately challenging for learning. Executable validation establishes feasibility, yet does not reveal how a task behaves relative to a given solver setting. In this paper, we present CalibForge, an autonomous terminal-task synthesis system that uses verified solver behavior to revise candidate tasks through adversarial solver calibration. Multi-solver calibration targets disagreement within a heterogeneous solver pool, whereas contrastive solver calibration targets a designated strong-pass/weak-fail relation; both operationalize a solver-relative learnable zone anchored in demonstrated solvability. Using CalibForge, we construct 5,431 calibrated terminal tasks. Our ablations show that both strategies yield more effective supervision than authoring and validation alone or ordinary single-solver feedback. Models trained on the full collection achieve 32.58% and 47.57% on Terminal-Bench 2.0. The largest improvements over the corresponding base model reach 24.71 percentage points on Terminal-Bench 2.0, 27.68 points on SWE-bench Pro, and 30.04 points on Doc2Repo. Together, these results support solver-relative learnability as a practical target for constructing effective and transferable agent training data.

---

引用前请核对原文;本文由星际流动整理:https://feed.astralor.com/items/itm_a6532616d3e4c3cc6e27d2cb