All papers

Proposes an inference-time unlearning method that detects leaks and injects corrective embeddings into large reasoning model traces.

Aligns independently pretrained unimodal encoders using a few paired anchors and distilled entropic optimal transport.

Uses reinforcement learning-based neural architecture search to discover compact models for audio-visual speech enhancement.

Introduces a counterfactual-guided few-shot distillation framework to train smaller question-answering language models using limited training samples.

Proposes a hierarchy of modularity to identify functional modular organization in neural networks lacking single-neuron specialization.

Proposes formal mathematical postulates and definitions to measure and operationalize computational creativity across domains.

Investigates mappings between generative model latent states and simulator states to enable controllable environment simulation.

Introduces representation geometry regularization to prevent capability degradation during on-policy self-distillation in language models.

Applies a reward-directed diffusion model to optimize localized ship hull geometry evaluated by automated fluid dynamics simulations.

Proposes a sequence-drifting framework to generate full text sequences in a single forward pass from Gaussian noise.

Presents a benchmark using real urban data to evaluate multimodal language models on spatial reasoning and constrained editing.

Proposes a federated pairwise ranking framework for institutions to evaluate decentralized finance participants without sharing transaction data.

Studies the impact of AI shopping assistant mediation on retail product assortment planning and business revenue.

Uses natural language class descriptions as supervision to train neural decoders for brain-computer interfaces across multiple modalities.

Presents an uncertainty-guided framework to compress reasoning supervision trajectories and reduce reasoning post-training costs.

Proposes matrix-entropy and kernel alignment criteria to select optimal layer subsets of hidden states for speculative decoding.

Presents a benchmark to evaluate language models on generating mathematical figures and diagram-assisted reasoning across frontier problems.

Proposes a multimodal agent providing grounded prompts and video clips to guide video editing models in professional sports motions.

Proposes continuous-query generation paradigms to eliminate reachability constraints caused by finite-beam decoding over discrete semantic IDs in recommendation.

Analyzes layer-wise visual interventions in multimodal language models to identify where visual edits effectively reduce hallucinations.

Introduces an information-theoretic framework to reweight supervised fine-tuning data based on conditional information gain.

Compares hyperbolic and Euclidean embeddings for unsupervised patent representation to improve cross-category prior-art discovery.

Presents a scalable method to maximize agent empowerment by guiding agents toward states affording numerous executable skills.

Introduces a training objective enabling language models to update beliefs using report lineage rather than raw citation repetition.

Proposes a generalized f-divergence trust-region framework to unify and stabilize reinforcement learning algorithms for LLM post-training.

Introduces a benchmark to evaluate how effectively LLMs clarify ambiguous user requirements in repository-level software development tasks.

Proposes spatial-semantic token compression and layer routing to accelerate 3D vision-language model inference across multi-view inputs.

Presents a conversational benchmark testing whether models can edit text under counterfactual changes while preserving unaffected factual details.

Proposes an advertising image editing method that identifies and selectively refines visual attributes driving click-through rates.

Investigates internal transformer mechanisms, showing attention updates decode harmfulness while MLP layers primarily drive refusal token generation.

Presents a difficulty-aware framework that allocates reasoning steps across sections to improve efficiency in large audio-language models.

Audits audio-language model hallucination mitigations by testing localized acoustic evidence ablation against baseline methods across model architectures.

Proposes a synthetic self-distillation method using a slightly larger aligner to enable continual post-training in small language models.

Introduces an anticipatory generative modeling approach for road networks to improve realistic next-transition selection in trajectory generation.

Proposes a memory-guided differential privacy framework enabling private inference over million-token contexts in LLM-as-a-service settings.

Introduces an execution-grounded dataset and method to repair geometric and engineering constraint violations in LLM-generated industrial scene programs.

Introduces an adaptive curriculum reinforcement learning method to train multimodal models on long-horizon visual Sudoku reasoning tasks.

Investigates whether frozen recurrent networks can provide sufficient memory for partially observable reinforcement learning without end-to-end training.

Introduces AFCE, training vision-language-action policies to predict future physical effects rather than raw action chunks directly.

Introduces ReSFlow, rectifying categorical flows along geodesic trajectories on statistical manifolds to enable one-step discrete generation.

Studies uncertainty estimation for abstract meaning representation parsers using top-k beam candidates and prediction rejection ratios.

Analyzes tabular foundation models and shows most transformer layers can be replaced with affine translators without sacrificing performance.

Proposes BoundaryEvo, dynamically adapting training data around a student model's evolving capability boundary during long-chain reasoning distillation.

Proposes PRIME, an index-side optimization framework using semantic restructuring and deterministic metadata synthesis to improve document retrieval.

Evaluates a label-free rubric clarification method, finding it increases consensus among LLM judges without improving agreement with human evaluations.

Evaluates whether robotic manipulation agents can learn complex multi-step assembly skills purely from sparse binary task rewards.

Develops a probabilistic transformer meta-learning architecture to perform zero-shot system identification directly from long training sequence contexts.

Combines a CNN-Transformer network with a continuous sigmoid label encoding to overcome grid quantization errors in direction-of-arrival estimation.

Evaluates neural embedding and diagnostic strategies to interpret detected anomalies in high-luminosity collider particle physics data.

Incorporates marginal-cost rewards into multi-agent reinforcement learning to reduce autonomous vehicle travel times without worsening human driver delays.

Introduces SqueezeBench, a saturation-resilient benchmark for formal mathematical reasoning where new accepted proofs dynamically raise the target bounds.

Investigates multimodal LLM failures on squint-readable low-frequency text and provides a diagnostic benchmark alongside training interventions.

Introduces BranchOPD, which branches student rollouts from shared prefixes at uncertain decision points to improve on-policy distillation efficiency.

Presents Explanatory World Models, using language model causal explanations to redistribute reinforcement learning rewards across distant actions.

Proposes PRISM, a multimodal clinical prognostic model combining patient records with retrieved population-level embedding memory banks.

Formalizes a theoretical hierarchy of identifiability criteria for estimating causal effects across equivalence classes of causal graphs.

Proposes Road MOPD, using non-routed teachers as cross-domain critics to resolve gradient conflicts in multi-teacher on-policy distillation.

Introduces an unsupervised method to identify and remove subliminally transferred traits from model parameter updates during knowledge distillation.

Develops Step-wise Truncation Probing and termination strategies to quantify and mitigate accuracy degradation from overthinking in reasoning models.

Proposes AudioTETRIS, a training-free token pruning framework that uses acoustic transition signals to compress audio representations efficiently.