All papers

The paper measures inference energy efficiency across 86 open-weight language models, showing large efficiency gaps among models of similar capability.

The paper presents a simulated benchmark across eleven unconstrained environments to evaluate vision-language models on documenting real-world human physical work.

The paper proposes an uncurated multi-stream latent memory framework that augments vision-language-action robotic policies while maintaining low inference latency.

The paper introduces an evaluation framework to measure the trade-off between task quality and computation cost when replacing iterative neural modules.

The paper introduces a benchmark and framework to evaluate how multimodal models track entities and reason about future scene evolution in videos.

The paper proposes a data-adaptive importance sampling method to stabilize variance and prevent sample size collapse in asynchronous RL post-training.

The paper introduces a testing framework that turns documented model safety claims into specifications to find counterexamples via mutated prompts.

The paper introduces INT4 KV cache quantization with selective INT8 protection to accelerate asynchronous rollouts in reinforcement learning.

The paper presents a framework that jointly learns query rule assignment and rule-constrained evidence retrieval over structured knowledge bases.

The paper proposes a multi-winner banner ad auction mechanism using diffusion models guided by bids and text descriptions to generate ads.

The paper introduces a two-stage reasoning framework to generate personalized, safe, and explainable medication recommendations using language models.

The paper presents two geometric methods to evaluate where and why representational structures align across different large language models.

The paper develops a text-to-speech model that simulates realistic clinical dysfluency in aphasia speech across multiple speakers for training healthcare professionals.

The paper demonstrates that generative recommenders are vulnerable to targeted poisoning attacks using fewer than 1% malicious synthetic interactions.

The paper presents a pretraining framework aligning wearable PPG signals with text across both individual pulse morphologies and broader demographic contexts.

The paper examines positive activation bias in neural networks, showing that reducing this bias slows initial training but improves final model performance.

The paper introduces a flow-matching model that enforces exact linear identity constraints among EEG channels to generate physically valid clinical signals.

The paper presents a feature-proxy method to selectively erase combined object-style concepts from diffusion models without degrading general generative performance.

The paper specifies a verifier-governed protocol for panvascular research and benchmarks a baseline model on intervention imaging tasks.

The paper proposes an energy-anchoring technique to reliably separate known and novel classes during continual test-time model adaptation.

The paper presents a neural-symbolic method for multi-robot systems to evaluate robot and object hypothesis trust purely via inter-robot observation consistency.

The paper introduces a diagnosis framework to evaluate whether multimodal benchmark items genuinely require multi-image reasoning rather than unimodal shortcuts.

The paper introduces a finite-window theoretical framework to determine what historical state information dynamical systems and distributed AI workflows must retain.

The paper introduces a large language model-based metric for offline evaluation of user-product relevance in dynamic product advertising retrieval.

The paper presents a certificate-based search method paired with a symbolic checker to resolve complex, distributed dependencies in protocol specifications.

The paper proposes a distillation framework to remove specific multimodal knowledge without causing hallucinations or leaving residual queryable information.

The paper analyzes theoretical computational limits of sequential sampling circuits in autoregressive language model token generation.

The paper releases a multimodal dataset and dual-encoder model to align natural language descriptions with global Earth system meteorological data.

The paper evaluates a framework that replaces spoken patient identifiers and alters voice identity while preserving clinical conversational evidence.

The paper studies expert width and dimension balance in latent mixture-of-experts architectures to optimize parameter capacity across model layers.

The paper introduces a multi-head prediction method that injects future long-range supervision to improve autoregressive video generation consistency.

The paper introduces a hybrid kinematic graph neural network and elastic transformer framework for efficient Indian sign language recognition from video.

The paper presents a framework to learn dexterous robotic hand control directly from human monocular video demonstrations without teleoperation.

This paper presents a parameter-efficient video diffusion adaptation method combining point tracking with gradient conflict resolution to preserve visual priors.

This paper proposes a dual-alignment framework to keep speech-conditioned audio generation robust and consistent across clean and noisy acoustic environments.

This paper presents a multimodal diffusion framework that jointly designs protein sequences and physical structures to ensure interface complementarity.

This paper demonstrates a security risk where dormant backdoors trigger safety failures only after merging models that passed individual safety audits.

This paper introduces an aligned language model family trained with structured reinforcement learning for clinical reasoning and traditional Chinese medicine prescribing.

This paper introduces a pairwise evaluation benchmark and protocol designed to measure subjective aesthetic taste while controlling for shortcut artifacts.

This paper proposes a framework using language model attributes to ground and label novel visual classes discovered dynamically at test time.

This paper integrates adversarial inverse reinforcement learning into a fuzzy control framework to estimate reward functions directly from demonstration data.

This paper introduces a hierarchical multi-timescale world model to reduce compounding rollout errors and drift in sequential decision-making systems.

This paper introduces a self-improving synthetic data generation pipeline to train embedding models across diverse retrieval domains without extensive manual annotation.

This paper introduces energy-based channel grouping and bit allocation to preserve accuracy when quantizing large language models to two bits.

This paper proposes an evaluation framework to test whether credit assignment techniques in modular financial reinforcement learning truly improve economic decisions.

This paper shows recurrent layers can predict future attention importance, enabling memory eviction in hybrid language models before queries arrive.

This paper introduces a multimodal unlearning method using counterfactual inputs and on-policy self-distillation to remove sensitive data without damaging general knowledge.

This paper introduces a benchmark evaluating whether coding agents produce functionally correct code that also avoids real-world security vulnerabilities.

This paper investigates using frozen, pre-trained Transformer scorers instead of fixed mathematical formulas for multimodal contrastive representation learning.

This paper introduces a compact text-to-speech model trained with a realism critic to improve prosodic naturalness during zero-shot voice cloning.

This paper develops a graph edge attention neural solver for dynamic pricing and vehicle routing in same-day delivery operations.

This paper presents a multi-year network traffic time-series dataset to evaluate long-term anomaly detection and forecasting models.

This paper introduces a method to enable composable driving-style control in end-to-end autonomous driving planners using task vectors.

This paper proposes a depth-progressive framework using 2D generative diffusion models to build 3D seismic velocity models efficiently.

This paper evaluates an edge-cloud hybrid architecture using AWS IoT Greengrass for real-time machine learning inference in predictive maintenance.

This paper introduces an Indian language text-to-speech benchmark and an inference-time duration estimation method for alignment-free synthesis.

This paper introduces an accounting framework to measure recursive self-improvement and audits machine-authored operator generation in a physics runtime.

This paper proposes a curriculum-guided reinforcement learning framework and an agentic reward model for spatial text-to-image generation.

This paper introduces an adapter that aligns operational numerical weather prediction inputs with ERA5 reference states for pretrained forecasting models.

This paper introduces a JAX-native benchmark for evaluating heterogeneous and competitive multi-agent reinforcement learning algorithms.