Allocates token-level credit in rubric-conditioned GRPO by counterfactually replaying the same response under rubric and criteria-free prompts, using tokenwise log-likelihood contrasts to compute bounded, response-normalized weights that redistribute GRPO advantages without training an auxiliary scorer.
Evaluates multimodal context learning across grounding, new information application, and knowledge acquisition using a 3,443-instance benchmark spanning science, finance, long documents, spatial reasoning, and web VQA; finds current multimodal models perform poorly (best score 0.2847) and analyzes failure modes.
Recovers editable design files from raster images by growing an editable layer hierarchy via an agentic pipeline that selects and composes modality-specific tools. Introduces graceful verification (accept/prune/retry) to prevent error accumulation and presents the Figma Edit Replay Benchmark (909 files, 14,796 edits) to measure editability across layout, color, and text edits.
Directly maps visual observations and language instructions to continuous robot actions, replacing LLM-centric V→L→A pipelines. Uses separate visual and language encoders with lightweight bidirectional interaction and a compact decoder to cut inference cost and VRAM, achieving ~31 ms latency and <1 GB VRAM on an RTX 4090; suited for real-time robotic manipulation under tight compute budgets.
Evaluates whether vision-language models can make actionable decisions for a physical body by decoupling decision-making from low-level motor execution. Introduces HumanCLAW-Bench with 1,218 long-horizon egocentric episodes across 41 indoor scenes and diagnoses a lack of embodied self-awareness in current VLMs.
Learns a discrete “physical language” from unlabeled videos and uses a reason-then-render pipeline: predict compact state-transition tokens, then decode them into future video. Separates dynamics inference from pixel synthesis to improve physical fidelity, controllable simulation, and zero-shot motion transfer.
Autonomously proposes, modifies, executes, and evaluates ML experiments to study recursive self-improvement in machine learning engineering. Implements an open stack (OpenMLE-Gym, -RL, -Evo) and post-trains Frontis-MA1 (35B) around four evolution operators (Draft, Improve, Debug, Crossover); releases model weights and the full codebase.
Estimates the visually attributable portion of a privileged teacher’s next-token corrections and reconstructs student-anchored training targets for multimodal on-policy distillation. Uses counterfactual teacher queries and a signed proxy to raise supported tokens and suppress refuted ones, improving fine-grained visual knowledge transfer across model scales.
Introduces AISPA, a user-centric framework to audit system prompts in LLM applications, and applies it to 3,249 instructions from 88 commercial products to classify protective versus problematic instructions. Highlights design variability, growing prompt length/protection, persistent problematic directives, and calls for transparency and oversight.
Regularizes latent world models by replacing the Epps–Pulley Gaussianization objective with a quantile–quantile matching loss that aligns projected latent samples to rank-matched Gaussian quantiles, improving tail correction and planning success via cross-batch ranking.
Evolves persistent, stateful environments to red-team tool-using AI agents — provides 10K+ validated scenarios across 50 domains and a feedback-driven attack policy (EMHA) to surface long‑horizon safety failures.
Retrieves short speech segments from MEG recordings with a compact interpretable neural decoder trained against wav2vec 2.0 embeddings, and maps decoder weights to cortical source space to reveal which acoustic and linguistic features drive retrieval.