Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Continual Learning for Instruction Following from Realtime Feedback
Alane Suhr, Yoav Artzi
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 |
|---|---|---|---|
| Train continually on realtime user feedback | Instruction-following agent · CerealBar, long-term experiment | instruction execution accuracy: 66.7 ± 1.5 -> 82.1 ± 1.1 | Improved |
| Deploy the continually trained final agent rather than the initial agent | Instruction-following agent · CerealBar, side-by-side comparison | instruction execution accuracy: 70.4 -> 83.6 | Improved |
| Add heuristically propagated rewards instead of using only simple rewards (ablation) | RewardProp · CerealBar, Round 3 | instruction execution accuracy: 78.1 ± 1.0 -> 80.3 ± 1.0 | Improved |
| Discard negative feedback and reboots from training (ablation) | RewardProp · CerealBar, Round 4 | instruction execution accuracy: 81.0 ± 0.9 -> 76.8 ± 1.0 | Hurt |
| Reduce initialization demonstrations to 144 human-human interactions (ablation) | Instruction-following agent · CerealBar, initialization | instruction execution accuracy: 70.6 ± 1.1 -> 34.8 ± 1.8 | Hurt |
| Reduce initialization demonstrations to 144 human-human interactions (ablation) | Instruction-following agent · CerealBar, after five rounds | instruction execution accuracy: 34.8 ± 1.8 -> 64.2 ± 1.5 | Improved |
| Use feedback data from human-agent interaction rather than an equivalent amount of added human demonstrations (ablation) | RewardProp · CerealBar, five-round comparison | instruction execution accuracy | No clear effect |
| Propagate feedback to previously unrewarded actions within the heuristic window (ablation) | RewardProp · CerealBar, action reward assignment | proportion of actions assigned reward: 63.1% -> 82.2% | Improved |
Closest papers by meaning
- Generalizing Verifiable Instruction Following (NeurIPS 2025 · similarity 0.88)
- Policy Improvement using Language Feedback Models (NeurIPS 2024 · similarity 0.87)
- GUIDE: Real-Time Human-Shaped Agents (NeurIPS 2024 · similarity 0.87)
- MultiScale Contextual Bandits for Long Term Objectives (NeurIPS 2025 · similarity 0.86)
- Towards Better Robot Learners: Leveraging Implicit and Explicit Human Feedback Together in Human Robot Interactions (AAAI 2025 · similarity 0.86)
- Checklists Are Better Than Reward Models For Aligning Language Models (NeurIPS 2025 · 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.
