Research map / Multi-agent systems and games / Human-AI and goal-conditioned
DataEnvGym: Data Generation Agents in Teacher Environments with Student Feedback
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces DataEnvGym, a testbed of teacher environments in which LLM-based data generation agents iteratively create training data for a student model using student feedback across math, programming, VQA, and tool-use tasks.
Problem
Automating the human-driven loop of diagnosing student model weaknesses and creating targeted training data lacks a simulated, feedback-driven environment for testing data generation agents.
Approach
Frames data generation as sequential decision-making: a policy plans data creation from student feedback states, an engine renders the plan into data, and the environment trains and evaluates the student each iteration. Three environments (Open-Ended, Skill-List, Skill-Tree) differ in state structure and action space, with skills automatically inferred by an LLM.
Key results
- With State policies outperform No State by 3.5% (Open-Ended), 2.05% (Skill-List), and 1.08% (Skill-Tree) on average across GQA, MATH, LiveCodeBench.
- Open-Ended with State improves average accuracy by 4.60 points; GQA improves from 44.18 to 47.90 and MATH from 15.78 to 23.44.
- Oracle skills improve GQA student to 53.02 vs 48.18 with inferred skills in Skill-List.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| condition the teacher policy on student feedback instead of using a state-free policy (ablation) | Open-Ended environment · GQA | accuracy: 43.48 -> 47.90 | Helped |
| condition the teacher policy on student feedback instead of using a state-free policy (ablation) | Skill-List environment · MATH | accuracy: 19.78 -> 19.48 | No clear effect |
| condition the teacher policy on student feedback instead of using a state-free policy (ablation) | Skill-Tree environment · GQA | accuracy: 49.53 -> 49.76 | No clear effect |
| condition data generation on the student's state (ablation) | Open-Ended and Skill-List environments · NaturalBench | Group Score: 0.00 -> 6.00 | Helped |
| condition data generation on the student's state (ablation) | Skill-List environment · MnMs | Plan Accuracy: 62.13 -> 75.96 | Helped |
| use oracle skills instead of inferred skills (ablation) | Skill-List · GQA | accuracy: 48.18 -> 53.02 | Helped |
| use oracle skills instead of inferred skills (ablation) | Skill-List · MATH | accuracy: 18.25 -> 19.52 | Helped |
| replace GPT-4o with GPT-4o-mini as the teacher (ablation) | Open-Ended environment · GQA | accuracy: 47.90 -> 45.51 | Hurt |
| replace GPT-4o with GPT-4o-mini as the teacher (ablation) | Open-Ended environment · MATH | accuracy: 23.44 -> 15.78 | Hurt |
| replace GPT-4o with GPT-4o-mini as the teacher (ablation) | Open-Ended environment · LiveCodeBench | accuracy: 18.91 -> 17.50 | Hurt |
| use less generated data and compensate with more training epochs (ablation) | DataEnvGym training · GQA | accuracy: 47.90 -> 42.79 | Hurt |
| use less generated data and compensate with more training epochs (ablation) | DataEnvGym training · MATH | accuracy: 23.44 -> 13.98 | Hurt |
Surprising findings
- On LiveCodeBench, policies without state information produce no improvement at all.
- Teaching with a weaker GPT-4o-mini teacher still yields gains, though smaller.
Limitations
- Student models are already instruction-tuned, so improvements are small and hard to obtain.
- Skill-Tree performance is sensitive to the skill discovery quality; MATH uses approximated oracle skills.
- Results rely on GPT-4o as the teacher, which is costly and proprietary.
Experimental setup
Datasets: GQA NaturalBench MATH LiveCodeBench MnMs
Benchmarks: GQA MATH LiveCodeBench NaturalBench MnMs
Metrics: accuracy relative improvement over pre-teaching student group score plan accuracy tool F1
Baselines: No State policy GPT-4o-mini teacher Oracle skills
Builds on: EnvGen Alpaca LoRA Llama-Factory
Closest papers by meaning
- Benchmarking Generative Models on Computational Thinking Tests in Elementary Visual Programming (NeurIPS 2024 · similarity 0.87)
- Agent4Edu: Generating Learner Response Data by Generative Agents for Intelligent Education Systems (AAAI 2025 · similarity 0.87)
- MASTER: Enhancing Large Language Model via Multi-Agent Simulated Teaching (NeurIPS 2025 · similarity 0.86)
- DistiLLM-2: A Contrastive Approach Boosts the Distillation of LLMs (ICML 2025 · similarity 0.86)
- GenAL: Generative Agent for Adaptive Learning (AAAI 2025 · similarity 0.85)
- Star-Agents: Automatic Data Optimization with LLM Agents for Instruction Tuning (NeurIPS 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.