Research map / Multi-agent systems and games / Human-AI and goal-conditioned
How to talk so AI will learn: Instructions, descriptions, and autonomy
Theodore Sumers, Robert D. Hawkins, Mark K Ho, Thomas L. Griffiths, Dylan Hadfield-Menell
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 |
|---|---|---|---|
| Infer the speaker's horizon jointly with their reward function instead of fixing the horizon | Pragmatic listener · Behavioral experiment | future rewards: .93 vs M=0.94; mean difference 0.01 | Improved |
| Use pragmatic inference rather than literal interpretation | Listener · Behavioral experiment | future rewards: 0.15 | Improved |
| Use a fixed long horizon H=4 instead of inferring the horizon (ablation) | Pragmatic listener · Behavioral experiment | future rewards: 0.026 | Hurt |
| Use a fixed short horizon H=1 instead of inferring the horizon (ablation) | Pragmatic listener · Behavioral experiment | future rewards: 0.11 | Hurt |
| Use the social posterior as a prior for individual Thompson sampling | Individual reinforcement learner · Behavioral experiment | regret: Individual coefficient 2.60 | Improved |
| Use known H rather than latent H when integrating social information with RL (ablation) | Pragmatic reinforcement learner · Behavioral experiment | regret: β = -.12 | Improved |
| Apply pragmatic inference to instructions rather than interpreting them literally | Listener · Instruction utterances | mean pragmatic gain in future rewards: .42 ± .23 | Improved |
| Apply pragmatic inference to descriptions rather than interpreting them literally | Listener · Description utterances | mean pragmatic gain in future rewards: -.07 ± .22 | Hurt |
| Increase the speaker optimality parameter from βS1=2 to βS1=3 (ablation) | Pragmatic Known H · Behavioral experiment | future rewards: 0.9308 vs 0.9295 | No clear effect |
| Use instructions rather than descriptions at short horizons | Reward-maximizing speaker · Contextual bandit simulations, short horizons | speaker rewards: 1.75 | Improved |
| Use descriptions rather than instructions as the horizon lengthens | Reward-maximizing speaker · Contextual bandit simulations, long horizons | future rewards | Improved |
Closest papers by meaning
- ROSETTA: Constructing Code-Based Reward from Unconstrained Language Preference (ICLR 2026 · similarity 0.89)
- Bridging semantics and pragmatics in information-theoretic emergent communication (NeurIPS 2024 · similarity 0.89)
- Ground-Compose-Reinforce: Grounding Language in Agentic Behaviours using Limited Data (NeurIPS 2025 · similarity 0.88)
- Motif: Intrinsic Motivation from Artificial Intelligence Feedback (ICLR 2024 · similarity 0.87)
- REvolve: Reward Evolution with Large Language Models using Human Feedback (ICLR 2025 · similarity 0.87)
- Implicit Intelligence - Evaluating Agents on What Users Don’t Say (ICML 2026 · similarity 0.87)
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.
