AI Paper Atlas

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

NeurIPS 2022 · Multi-agent systems and games › Human-AI and goal-conditioned · approach: Theory-of-mind social reasoning · 5 citations (OpenAlex)

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 changedWhereMeasuredEffect
Infer the speaker's horizon jointly with their reward function instead of fixing the horizonPragmatic listener · Behavioral experimentfuture rewards: .93 vs M=0.94; mean difference 0.01Improved
Use pragmatic inference rather than literal interpretationListener · Behavioral experimentfuture rewards: 0.15Improved
Use a fixed long horizon H=4 instead of inferring the horizon (ablation)Pragmatic listener · Behavioral experimentfuture rewards: 0.026Hurt
Use a fixed short horizon H=1 instead of inferring the horizon (ablation)Pragmatic listener · Behavioral experimentfuture rewards: 0.11Hurt
Use the social posterior as a prior for individual Thompson samplingIndividual reinforcement learner · Behavioral experimentregret: Individual coefficient 2.60Improved
Use known H rather than latent H when integrating social information with RL (ablation)Pragmatic reinforcement learner · Behavioral experimentregret: β = -.12Improved
Apply pragmatic inference to instructions rather than interpreting them literallyListener · Instruction utterancesmean pragmatic gain in future rewards: .42 ± .23Improved
Apply pragmatic inference to descriptions rather than interpreting them literallyListener · Description utterancesmean pragmatic gain in future rewards: -.07 ± .22Hurt
Increase the speaker optimality parameter from βS1=2 to βS1=3 (ablation)Pragmatic Known H · Behavioral experimentfuture rewards: 0.9308 vs 0.9295No clear effect
Use instructions rather than descriptions at short horizonsReward-maximizing speaker · Contextual bandit simulations, short horizonsspeaker rewards: 1.75Improved
Use descriptions rather than instructions as the horizon lengthensReward-maximizing speaker · Contextual bandit simulations, long horizonsfuture rewardsImproved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy