Research map / Large language models / Agents, code and math
A Systematic Study of Behavioral Cloning for Scientific Data Annotation
Ishaan Singh Chandok, Core Francisco Park
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Pretrain on 9 tasks then fine-tune, versus training from scratch (ablation) | Multi-task BC model · SHAPE MATCHING (held-out) | accuracy: 0% -> 76.6% | Improved |
| Provide in-context demonstrations or action prefixes instead of fine-tuning (ablation) | Pretrained multi-task model · SHAPE MATCHING (held-out) | accuracy: 0% | No clear effect |
| Increase fine-tuning data from 500 to 7800 sequences (ablation) | Pretrained multi-task model · SHAPE MATCHING | accuracy: 76.6% at 500; similar at 7800 | No clear effect |
| Scale model parameters 10x (25M to 320M) (ablation) | Multi-task BC · Training loss vs tokens | sample efficiency: roughly 3x better sample efficiency | Improved |
| Scale model from 25M to 320M at equal FLOPs (ablation) | Multi-task BC · Training loss at equal FLOPs | loss | Hurt |
| Larger model at equal training loss (ablation) | Multi-task BC · Decision actions (done, undo, navigate) | accuracy | Hurt |
| Larger model at equal training loss (ablation) | Multi-task BC · Motor actions | accuracy | No clear effect |
| Use the 95M BC model instead of prompting VLMs | BC vs Gemini 3 Flash / Qwen3-VL-32B · COLORED DOT TRACKING placement @5px | accuracy: 97.4% vs 80.0% / 25.0% | Improved |
| Apply beta-DAgger on failing tasks (ablation) | BC model · Animal behavioral and limb tracking | autoregressive success | No clear effect |
| Fine-tune briefly on 8 GUI layout variants | Pretrained BC model · COLORED DOT TRACKING GUI variants | performance | Improved |
| Train on data containing mistakes then undo | BC model · Generative mistake rate | mistake rate: 9.2% in data vs 1.0% generated | Mixed |
Closest papers by meaning
- Rethinking Latent Redundancy in Behavior Cloning: An Information Bottleneck Approach for Robot Manipulation (ICML 2025 · similarity 0.86)
- A Smooth Sea Never Made a Skilled SAILOR: Robust Imitation via Learning to Search (NeurIPS 2025 · similarity 0.84)
- Interactive and Hybrid Imitation Learning: Provably Beating Behavior Cloning (NeurIPS 2025 · similarity 0.84)
- Self-Supervised Keyframe Discovery for Horizon-Invariant Behavior Cloning (NeurIPS 2026 · similarity 0.84)
- Self-Improving Embodied Foundation Models (NeurIPS 2025 · similarity 0.83)
- What Makes Good Data for Alignment? A Comprehensive Study of Automatic Data Selection in Instruction Tuning (ICLR 2024 · similarity 0.83)
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.
