All papers

The paper benchmarks twenty-three self-supervised 3D medical vision foundation models across segmentation, classification, and regression tasks.

The paper introduces an evaluation method that separates question answerability detection from answer span extraction in radiology report question answering.

The paper analyzes why dynamic graph neural networks fail on evolving nodes and introduces a change-aware classification framework.

The paper introduces an action-conditioned world model that separates observation representation learning from intervention learning for mobile network energy control.

The paper applies Hilbert geometry to constrain activation shifts in softmax layers for stable optimizer updates in language model heads.

The paper re-evaluates the pretraining compute scaling laws between diffusion and autoregressive language models using modern diffusion training designs.

The paper introduces a dataset and benchmark to evaluate vision-language models on recognizing and answering questions about Philippine cultural heritage sites.

The paper introduces a morphology-aware contrastive learning framework to predict five-year colorectal cancer survival from whole-slide pathology images.

The paper presents a monocular depth estimation method that uses image editing priors to improve geometric consistency under environmental perturbations.

The paper proposes a rubric-based reinforcement learning method combining selective teacher distillation with an adaptive hand-off to standard optimization.

The paper introduces a 4D latent cache to reuse static and dynamic visual representations for efficient conditioning in video generation models.

Presents SADDER, an inference-time text diffusion safety framework using directional embedding signals to steer unsafe tokens toward benign distributions.

Shows that activation steering evaluations can conflate bias reduction with model abstention, introducing a gradient-constrained direction design to isolate preferences.

Introduces RoTA-Bench to evaluate vision-language models on 3D spatial reasoning through interactive object rotation to match target 2D silhouettes.

Proposes PrT-Flow, an invertible flow matching method that enables exact action probability computation for reinforcement learning in robotics.

Introduces EncounterMob, an LLM-based framework that generates human mobility trajectories while preserving both individual routines and multi-person encounter patterns.

Proposes Dynamic K-Head, a question-conditioned budget selector that sets continuous latent reasoning steps based on input difficulty.

Proposes Token Window Pacing, a serverless scheduling framework that regulates request admissions at the token level to reduce inter-token latency.

Proposes MedEvo, an iterative training framework that repairs multimodal medical reasoning trajectories using minimal hints and taxonomy-guided feedback.

Introduces the PocketWatch benchmark and zPocket, an approach training pocket-prediction heads on frozen representations from protein co-folding models.

Presents LeanPRM, a method using the Lean proof assistant to generate step-level correctness labels for training natural-language mathematical proof verifiers.

Studies the trade-off between spatial smoothness and gene-level fidelity when mapping single-cell RNA sequencing profiles to images for neural networks.

Introduces a graph embedding and Procrustes alignment framework to provide a bounded metric for evaluating discovered process models against ground truth.

Proposes a training-free framework that mitigates vision-language model hallucinations by admitting claims only when supported relative to a visual-null input.

Develops SCRIM, a method quantifying state criticality in imitation learning by measuring downstream value degradation under bounded action perturbations.

Proposes a recursive self-improvement framework for electrocardiogram analysis models combining verified synthetic waveform edits with targeted intervention selection.

This paper introduces a benchmark to evaluate whether egocentric multimodal assistants understand who is harmed rather than solely tracking output refusal.

This paper proposes a credit-assignment method that uses rubric evaluation outcomes to allocate token-level advantages during reinforcement learning.

This paper presents a reinforcement learning approach that uses in-context dynamic memory during rollouts and self-distillation to accelerate reasoning exploration.

This paper introduces a posterior learning approach that improves multi-sample reasoning coverage by updating models using single posterior draws.

This paper presents a benchmark with an action taxonomy to evaluate the interactive conversation capabilities of full-duplex spoken language models.

This paper proposes an end-to-end framework integrating nonlinear dimension reduction with longitudinal counterfactual prediction for high-dimensional observational data.

This paper provides four auditing tests to determine whether failure-attribution judges actually perform multi-step checks or merely rely on spurious surface cues.

This paper presents a system combining interactive human corrections with iterative policy retraining for long-horizon whole-body mobile robot manipulation.

This paper proposes Gaussian process bandit algorithms with selective verification audits to optimize AI system configurations under noisy evaluations.

This paper presents an autonomous system combining LiDAR SLAM, visual imagery, and thermal sensing to inspect under-panel components of solar farms.

This paper presents an optimization method that trains models to decide when to invoke reasoning tokens, reducing unnecessary thinking overhead.

This paper proposes a flow-matching method that averages Bellman target directions under a posterior distribution to improve reinforcement learning critics.

This paper introduces a benchmark to evaluate whether forecasting models can incorporate forward-looking contextual information into accurate time-series predictions.

This paper conducts an adversarial evaluation comparing a probability-based safety judge against large language model judges and safety classifiers across multiple attack types.

This survey analyzes how verifier noise affects reinforcement learning with verifiable rewards across 58 studies and summarizes practical mitigation techniques.

Introduces QK-Edit, a training-free query-key attention manipulation method for text-guided image and video editing in multimodal diffusion transformers.

Proposes Diff Diffusion, a diffusion-based model that predicts localized code edit sequences rather than regenerating entire source files.

Provides a tutorial analyzing theoretical expressiveness gains and parameter tradeoffs when re-uploading data in quantum amplitude embedding architectures.

Introduces FASHION-RQ, an automated framework deriving hierarchical evaluation benchmarks for fashion retrieval directly from catalog structures.

Releases MMTA, a multimodal benchmark testing the ability of language models to jointly analyze time series, text, and visual data.

Proposes TStruct, a network that captures recurring temporal relationship structures across variables to improve long-term time-series forecasting.

Presents RT-J, a relational transformer pretrained on diverse databases to achieve strong in-context table predictions with few sample examples.

Shows that evil persona vectors extracted from low-resource languages transfer poorly and fail to steer model behavior effectively across languages.

Evaluates SHAP attribution methods on sparse autoencoder features, finding weak correspondence with direct causal feature intervention in language models.

Introduces PaceReasoner, using long reasoning outputs during test-time reinforcement learning to boost the accuracy of brief, low-latency model answers.

Analyzes answerability signals in frozen visual question answering models, finding standard confidence and alignment metrics fail to reliably detect unanswerable inputs.

Introduces DriveWLA, a mixture-of-transformers architecture unifying scene semantics, future dynamic prediction, and vehicle trajectory planning.

Finds that framing malicious coding prompts as engineering specifications systematically reduces language model refusal rates across multiple models.

Proposes LinMorph, a learnable weight transformation method to initialize linear attention models from pretrained video diffusion transformers efficiently.

Introduces CineMem, an attention-aligned memory retrieval framework using clustered KV caches for autoregressive multi-shot long video generation.

Presents a memory decomposition framework separating spatio-temporal beliefs to support autonomous robot navigation under evolving online goals.

Proposes MERIT, a framework using visual table reasoning and adapted encoders to extract information directly from table images.

The clinical study compares hyperspectral and RGB imaging for intraoperative detection of low-grade gliomas during brain tumor surgery.

HaloBlock allocates adaptive context halo regions across blocks to smooth context transitions and improve diffusion language model decoding.