All papers

Presents a benchmark based on real criminal cases showing how correction framings and cultural adaptations bypass safety guardrails.

Introduces a causal-beam transfer framework that extracts relevant subtask representations from human videos to guide robotic manipulation.

Proposes a bi-level working and long-term memory framework to balance knowledge retention and acquisition in incremental learning.

Learns instruction-conditioned multimodal representations to dynamically cluster objects based on user-specified partition criteria without fixed cluster counts.

Proposes a structural scaling method based on transformer unit roles to convert language models to ternary weights effectively.

Develops pessimistic value-iteration algorithms with theoretical suboptimality bounds for offline risk-sensitive reinforcement learning using optimized certainty equivalents.

Presents a unified framework with synthetic data generation and benchmarks to train models on embodied spatial reasoning and actions.

Develops a solvable topic model theory explaining when sequence models adopt in-context learning rather than specialised memorisation during training.

Analyzes strategic identifiability in multi-agent games, showing when inverse reinforcement learning fails to recover true underlying agent rewards.

Introduces a framework that aligns interaction regions across objects to improve 3D affordance grounding on unseen objects.

Investigates how user trust influences delegation and verification choices during human interaction with multimodal video question-answering systems.

Presents a theoretical orthogonal polynomial framework explaining why shallow neural networks learn low-frequency sinusoidal components before high frequencies.

Analyzes how continual learning methods affect demographic fairness over time across eleven deepfake generation techniques.

The paper aggregates irregular electronic health records into temporal graphs to predict risks of gastrointestinal cancers.

The paper optimizes batched task and motion planning to generate high-quality autonomous demonstrations for robotic manipulation training.

The paper separates task generation from model update inheritance to avoid compounding errors in self-evolving language model loops.

The paper tests how language models weigh price momentum, commentary tone, and earnings surprises when forecasting financial outcomes.

The paper aligns functional brain scans with multimodal representations to decode and retrieve perceived natural speech segments.

The paper evaluates single-pass decision models against rational choice axioms and identifies ordering and rival-option sensitivity failures.

The paper enables multimodal models to proactively ask clarifying questions when ambiguous inputs carry cross-modal safety risks.

The paper compresses video history into evolving lightweight model weights instead of competing for limited input context length.

The paper develops a self-distillation approach enabling models to learn from their own unverified alternative reasoning paths without ground-truth answers.

The paper introduces a commercially permissive multilingual instruction dataset covering five European languages to train open language models.

The paper presents a tiny learned policy controller to optimize token unmasking and remasking schedules in diffusion language models.

The paper extends integrated gradients to trace feature attributions along causal paths for high-dimensional models.

The paper personalizes mixture-of-experts models and analyzes which specialized experts encode distinct user-specific features.

The paper proposes a decoupled per-token scoring method to efficiently manage key-value cache memory during long-context generation.

The paper coordinates projector and language model adaptation using task-specific routing to reduce catastrophic forgetting in multimodal models.

The paper introduces a framework to recalibrate, audit, and safely deploy confidence scores for language model outputs.

The paper uses retrieved in-context exemplars to improve rhetorical structure discourse parsing without requiring model fine-tuning.

The paper demonstrates specification gaming in a simulated cleaning task and tests a constrained reinforcement learning mitigation.

The paper proposes an alternating context optimization and logit steering framework to adapt language model outputs at test time.

The paper develops a multilingual retrieval-augmented tutoring system providing bilingual responses and glossaries for international computing students.

The paper trains feedforward 3D reconstruction models using privileged motion and gravity cues to improve dynamic scene understanding.

The paper introduces a reinforcement learning pipeline enabling humanoid robots to manipulate bulky objects using their arms and torso.

The paper evaluates how choices in rating scale design systematically alter text evaluation scores produced by large language models.

The paper introduces an adaptive connectivity objective to improve deep learning segmentation of thin, continuous anatomical or physical structures.

The paper proposes a graph neural network method to speed up layout decomposition for advanced semiconductor manufacturing.

The paper introduces a skill-decoupled on-policy distillation method to merge vision-language experts without conflicting gradients or reward hacking.

The paper presents a closed-form, training-free approach to merge specialized fine-tuned models into a single student using forward passes.

The paper introduces a framework to detect spurious shortcuts in safety benchmarks used to evaluate tool-using AI agents.

The paper evaluates 19 safety guards across multiple non-English languages to measure how translation alters step-level safety decisions for agents.

The paper develops an unsupervised routing method to escalate tasks from weak language models to stronger models without ground-truth labels.

The paper presents a method for robots to generate targeted preference queries to learn underspecified rewards from humans.

The authors introduce EliteReplay, combining reinforcement learning with quality-diversity search replay to retain solution diversity during post-training.

The authors propose Subspace Tuning, a fine-tuning method restricting weight updates to bottom singular subspaces to mitigate catastrophic forgetting in language models.

The authors present a random projection method to protect sensitive datastore embeddings against extraction attacks in retrieval-augmented generation systems.

The authors develop a clinical benchmark evaluating language model pipelines for generating multidisciplinary second opinions on complex medical cases.

The authors study why certain linearly represented facts in language models support generalization to new relations after fine-tuning while other recalled facts do not.

The authors present AVC-Reward, an event-centric framework to evaluate temporal, narrative, and attribute consistency in generated audio-video content.

The authors propose combining line plots and normalized delay embeddings to provide complementary visual representations for time-series reasoning in language models.

The authors propose ACEdit, a training and distillation scheme that stabilizes image editing models against accumulating errors across multiple editing turns.

The authors introduce LangDiT, an adapted transformer architecture designed specifically for continuous diffusion language models.

The authors evaluate how language models alter their refusal behavior when informed they are being monitored in English and five Indic languages.

The authors introduce DSP, a framework where robots learn from brief human play and autonomously collect contact-rich data for dexterous manipulation.

The authors introduce GeoState, a view-decoupled 3D world modeling approach separating scene dynamics from camera motion.

The authors propose MirrorMoE to reconcile an exploration dilemma between fine-grained expert segmentation and shared experts in mixture-of-experts models.

The authors present CoDA, a framework where a simulation data engine and a robot action policy co-evolve using memory of past failures.

The authors propose T2-Z, a method optimizing learnable environment latents at test time to help robotic world models adapt to unobserved physical factors.

The authors introduce an evolutionary framework using large language models to design and optimize resource-efficient quantum circuits.