All papers

Develops a GPU-accelerated tokenizer using CUDA graphs to reduce time-to-first-token in agentic language model serving.

Analyzes the emergence and behavior of attention sinks in text-to-image diffusion and autoregressive Transformers.

Presents a framework that defers memory construction to test time to answer queries across long interaction histories.

Proposes category-conditioned expert routing biases to mitigate data imbalance in mixture-of-experts language models.

Analyzes rubric construction mechanisms and introduces a tree-search method to refine rubrics for evaluating language model outputs.

Presents a modular prompt optimization framework and benchmark targeting stance detection in text.

Surveys twenty-two years of Tibetan natural language processing literature across corpora, speech processing, and language models.

Investigates continual learning regularization techniques for speech recognition during domain adaptation on financial earnings calls.

Evaluates DistilHuBERT-based accent classification under speaker-disjoint and cross-corpus conditions to measure generalization to unseen speakers.

Implements open-vocabulary speech recognition on a low-cost ESP32-S3 microcontroller without dedicated neural accelerators.

Proposes a workflow to enrich design-time cyber-physical system models using empirical evidence collected from simulations and operational runs.

Analyzes representational similarity metrics and demonstrates that input category selection has a larger impact on comparisons than the choice of metric.

Introduces POrchestra, an agentic orchestration framework enabling sub-agents to proactively communicate with the orchestrator during dynamic execution.

The paper identifies a linear spatial workspace subspace in multimodal language models that encodes continuous 3D coordinate representations.

The paper introduces a benchmark to measure demographic robustness, stereotyping, and intentional personalization in speech language models.

The paper demonstrates that machine-generated text detectors rely heavily on document formatting and PDF extraction artifacts rather than text quality.

The paper proposes compression-layer modifications to reduce accuracy loss when using dynamic LZW vocabularies during language model inference.

The paper presents a multilingual benchmark of minimal pairs to evaluate language model sensitivity to negative polarity items across 33 languages.

The paper trains conversational user simulators equipped with latent evaluation thoughts and scalar ratings to provide intrinsic feedback.

The paper introduces a benchmark evaluating agent susceptibility to prompt and script injections across multi-tool shared workspace environments.

The paper presents an importance-weighted downsampling protocol to lower evaluation costs on transaction graph datasets without distorting precision metrics.

The paper extends attention residuals across model depth by introducing multi-stream depth attention with causal convolutions.

The paper evaluates linguistic obfuscation patterns and reality-monitoring mechanisms across fraudulent and genuine scientific publications.

The paper evaluates knowledge-graph-informed chunking and retrieval strategies over German legal code to support verifiable retrieval-augmented generation.

The paper proposes multi-task pre-training guided by knowledge graphs to improve zero-shot legal reasoning in language models.

The paper presents a benchmark evaluating how effectively vision-language models recover fully occluded objects using remaining visual cues.

The paper studies language model sensitivity to discourse particles such as 'just' and 'only' in natural conversations.

The paper introduces a transformer architecture incorporating semantic role labeling into attention mechanisms to detect sentiment, emotion, politeness, and formality.

The paper analyzes phoneme-level error patterns from audio and text data augmentation techniques in low-resource automatic speech recognition.

The paper proposes a training-free query expansion method that filters hallucinated terms using corpus-conditioned cross-context reproducibility.

The paper studies policy learning and answer-selection mechanisms to help language models decide when to abstain on factual questions.

The paper presents an automatic pronunciation assessment method using Whisper cross-attention to predict word- and utterance-level scores without forced alignment.

The paper proposes accepting-path decoding using structural automata to enforce valid prefix completions within length limits for neural semantic parsing.

The paper evaluates vision-language models' spatial representations using causally controlled perturbations, showing spatial relations degrade compared to lexical features.

The paper introduces a domain-general framework using predictive world models to guide action selection and policy improvement in recursive self-improvement settings.

Presents a hypernetwork that maps document contexts directly into low-rank adapter weights for large language models.

Pairs attack success rate with operative understanding rate to assess whether models recognize harmful intent during safety evaluations.

Proposes an interpretability framework using sparse autoencoders to explain mixture of experts routing based on superposed fine-grained features.

Proposes an external memory framework allowing frozen language model agents to learn and iteratively refine natural-language world models compiled into executable code.

Presents a rollout selection strategy that drops high-probability positive-advantage trajectories to prevent entropy collapse during language model reinforcement learning.

Proposes a framework modeling research ideation as modular thinking flows sampled dynamically from a probabilistic supernet.

Introduces a benchmark evaluating multimodal language models on extracting data, tracking narrative flows, and answering questions from animated data videos.

Presents a bound-propagation verification framework to bound worst-case reconstruction errors for deep joint source-channel coding decoders under noise.

Proposes a policy mechanism enabling language models to decide when, how, and which test-time training procedures to execute.

Proposes a benchmark and probing methodology based on attribution theory to examine how language models assign responsibility and blame.

Presents a simulated participant environment to evaluate multi-turn AI interviewers across market research and social science scenarios.

Introduces an immutable, provenance-aware state representation framework with deterministic replay and audit logging for robotic systems.

Proposes a training method to adaptively skip turn-by-turn agent reasoning when previous thoughts remain sufficient for future actions.

Audits the reliability of LLM judges across multiple prompt formats, temperatures, and orderings, introducing a trustworthy verdict metric.

Presents a parameter-free distillation framework transferring knowledge between language models via an iteratively updated natural language guide.

Evaluates whether LLM compliance systems adjust verdicts when governing rules are perturbed, finding many verdicts are insensitive to rule changes.

Analyzes long-tail concept learning during supervised fine-tuning using a theoretical prior barrier metric to guide data needs.

Proposes an evaluation framework to measure how LLM agents handle uncertainty and resolve conflicting knowledge during execution.

Audits legal chain-of-thought faithfulness by substituting cited authorities and observing whether model verdicts causally change.

Evaluates whether internal model representations can predict how fine-tuning on one value generalizes across other held-out values.

Introduces a generative model that allows flexible post-hoc text editing through learned insertion, substitution, and deletion operations.

Presents a benchmark evaluating agentic large language models on long-horizon strategic reasoning in the game Slay the Spire.

Investigates self-preservation in language models by using task banks to steer synthetic replay during continual instruction tuning.

Demonstrates a backdoor attack where malicious reasoning degradation is triggered incrementally using invisible prompt tokens and stage tracking.

Proposes a recurrent harness framework that uses streaming feedback to revise textual memory for continual agent learning.