* equal contribution
tl;dr — Instead of committing to a single decoded action, we sample multiple candidate actions and use a trained generative verifier to select the best one. Up to 36% relative improvement on long-horizon embodied tasks without retraining the policy.
tl;dr — Under a fixed inference budget, we compare scaling solutions (self-consistency) vs. scaling verification chains (GenRM). Self-consistency is more compute-efficient at most practical budgets; GenRM requires up to 8× the compute to match it. We derive inference-scaling laws for the optimal trade-off.
tl;dr — Centaur is a language model fine-tuned on Psych-101, a dataset of over 10 million choices from 60,000+ participants across 160 experiments. It predicts held-out human behavior better than existing cognitive models and generalizes to new cover stories, structural task modifications, and new domains.
tl;dr — Concept Bottleneck Models treat concepts independently during human interventions. We realign concept assignments post-intervention using the statistical relationships between them, improving both classification and concept accuracy with fewer human edits per image. The method integrates into existing CBM architectures without modification.
tl;dr — CLIP is susceptible to data-poisoning backdoor attacks: as few as 75 poisoned pairs in 3M can manipulate its behavior. Fine-tuning with a combination of multimodal contrastive and per-modality self-supervised losses re-decorrelates the trigger from the target label and mitigates the attack while preserving clean performance.
🏆 Oral (Top 2%) · Best Paper, RTML @ ICLR 2023
tl;dr — We formulate a resource-rational theory of (self-)management by goal-setting and computationally derive optimal subgoals from a model of bounded human goal-pursuit. The derived subgoals improve problem-solving performance for both simulated agents and human participants.
🏆 Oral · Diversity and Inclusion Award
tl;dr — Re-analyzing three previously published mixed-gamble datasets with the drift-diffusion model, we find that a pre-valuation rejection bias plays a substantial role alongside valuation bias. A leaky competing-accumulator model further shows that an attentional bias toward losses can produce behavior indistinguishable from valuation bias.
tl;dr — Using fMRI on a game-like, intrinsically motivating task, we find that the striatum activates after successful outcomes regardless of whether the goal is framed as approach or avoidance — suggesting it encodes general motivation or effort rather than positive motivational state. The hippocampus instead tracks salient outcomes (success under approach, failure under avoidance).