All papers

The paper shows that whole-context KV cache rehearsal harms compression under tight budgets, proposing targeted rehearsal to preserve answer-relevant entries.

The paper trains 123 models across 54 tokenizers to show tokenizer choice matters more for languages with less training data.

Analyzes interaction failures between KL regularization and rewards in group policy optimization and proposes a calibrated surrogate objective.

Introduces an instruction-following benchmark and training recipe to improve goal adherence in full-duplex conversational models.

Releases BERT models pre-trained with yearly chronological cutoffs to reduce lookahead bias in scientific and patent text mining.

Presents a rolling masked diffusion method that accelerates full-duplex spoken dialog models by forecasting multiple future interaction frames.

Evaluates whether language models can act as world models to predict empirical experimental outcomes across AI research workflows.

Designs an inference serving system optimized for dynamic, token-level routing across multiple language models.

Proposes a scope-then-rank retrieval method using document topology and relational structure rather than pure semantic similarity.

Develops a framework for egocentric multimodal assistants to decide both when to speak and what guidance to provide.

Identifies distribution shift during autoregressive decoding and proposes a decoding-aware pruning method to accelerate model inference.

Presents a symmetry-aware value cache merging method to compress key-value memory during long-context decoding without architectural changes.

Proposes a method to select a compact subset of procedural skills that provide effective training signals during model distillation.

Introduces a multilingual tokenizer that uses structural discovery and linguistic constraints to improve downstream model representations.

Introduces a benchmark evaluating how well vision-language models predict unseen physical outcomes following scene interventions.

Proposes a framework that encodes visual trajectories into reusable skill cards to reinforce vision-language agent planning and rewards.

Presents a benchmark and dataset using generative video rollouts to train and evaluate visual transition reasoning in multimodal models.

Introduces a benchmark evaluating streaming vision-language models on fast-moving, high-dynamic events in real-world video streams.

Introduces a lightweight correction mechanism to restore token dependencies during parallel draft generation in speculative decoding.

Proposes an unsupervised framework using contrastive learning and clustering to induce, name, and annotate event types from text.

Presents a training method for latent recurrent language models that avoids performance penalties at single loops while scaling with compute.

Studies reflective prompt optimization in research agents, demonstrating that gains on held-out tasks can stem from prompt-level memorization.

Proposes a diagnostic framework to study how language models compress fluid trans-border ethnic identities into rigid nation-state categories.

Introduces a benchmark of 5,525 QA pairs from intelligence reports to evaluate RAG document- and chunk-level retrieval.

Presents a stateless service and library providing isolated, semantics-preserving text perturbations for language model robustness testing.

Tests whether priming annotators with scenario vignettes and micro-narratives improves the operationalization of lived experience in subjective annotation.

Describes an ontology-constrained ensemble system of encoders and fine-tuned language models for directed Arabic relation extraction.

Generates synthetic multi-agent dialogues to benchmark conversational agent robustness against non-cooperative and adversarial user behaviors.

Introduces an evaluation benchmark measuring how effectively language models provide supportive responses during self-harm and crisis disclosures.

Examines methodological and conceptual implications of using large language models and agentic workflows for machine-assisted cultural analytics.

Evaluates whether real-time speech models decide to speak appropriately during natural pauses and turn ends in unscripted conversations.

Introduces a benchmark evaluating how misleading visualizations and charts affect the multi-step execution and decisions of web agents.

Tests cross-corpus emotion classification transfer across three datasets, finding significant loss in discrimination and conceptual disagreement on labels.

Evaluates whether framing explanations around unreliable narrator theory enables open-weight models to explain satirical news cost-effectively.

Proposes a sequence modeling framework using dynamic nonlinear time delays based on semantic similarity between current and past inputs.

Presents a temporal memory architecture using hierarchical tree structures to trace evolving multi-party conversational histories for agents.

Proposes separating continuous diffusion language generation into high-level flow-generated predictive states and block-level autoregressive token decoding.

Introduces a preview-guided admission framework allowing search agents to selectively incorporate retrieved candidates into the ongoing reasoning context.

Introduces a benchmark evaluating vision-language models on graph-theoretic reasoning where evidence is split across multiple visual graph views.

Derives a formula to forecast disagreement between masked text reconstruction models fitted on independent samples from the same population.

Identifies internal attention heads, neurons, and sparse autoencoder features that causally influence narrative representation in language models.

Formalizes language model creativity evaluation as long-tail input-output distribution performance and tests it across multiple benchmarks and models.

Proposes a co-evolution framework that coordinates agent harness prompt modifications with model weight updates using self-generated experience.

Analyzes quantization error propagation in recurrent-depth language models and proposes adaptive mixed-precision schedules for faster prefill inference.

Presents a two-stage visual-textual alignment and retrieval-augmented framework for generating structured radiology reports from medical scans.

Develops an open-source testbed to evaluate segment-wise key-value cache reuse beyond exact prompt prefixes during language model serving.

Evaluates speech self-supervised learning models on Finnish broadcast audio using ASR confidence filtering and controlled architectural comparisons.

Presents a multi-dimensional annotated corpus of Bengali political YouTube discussions to support research on contextual cyberbullying detection.

Introduces an end-to-end open pipeline, pretraining corpus, and family of decoder-only models trained from scratch for Romanian.

Constructs a temporal and multilayer social dialogue network dataset based on the film adaptation of Les Misérables.

Proposes a reinforcement learning framework that tracks external versus internal knowledge sources to reduce search agent fabrication.

Presents a predictive subtree prefetching method to accelerate power sampling for test-time language model reasoning.

Introduces speculative decoding using self-sparse attention to speed up rollout generation during reinforcement learning.

Proposes a prompt-guided text augmentation framework to assist fake news detection in low-resource settings.

Designs diagnostic probes to determine whether audio-language evaluation models judge music based on audio or prompt content.

Tests whether self-supervised vision transformers represent semantic meanings of rendered words without paired text data.

Proposes a chunk-level credit assignment method for on-policy distillation across language models with different tokenizers.

Introduces a retrieval method combining token-level matching and document-level embeddings for memory-augmented language models.

Evaluates whether automated prompt optimizers inadvertently cause regressions in unscored prompt constraints and requirements.

Studies host-level budget allocation and crawling policies to collect pretraining web data for language models.