All papers

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.

Proposes Syn-Omni, a framework combining modality-expert LoRA paths and progressive routing to balance shared and modality-specific omnimodal embeddings.

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

Introduces AAArena, a benchmark evaluating how AI agents revise executable code policies across 12 adversarial games under fixed budgets.

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

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

Proposes a streaming optimal transport method to monitor LLM agent trajectories in real time against reference runs.

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.

Introduces a benchmark evaluating language models on eliciting geospatial data requirements and constraints from natural language requests.

Trains an unsupervised vision-language reward model using intermediate layer representations and caption perturbations without human preference labels.

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

Introduces an evolutionary optimization method using multi-turn reflective mutators to improve production data agents and pipelines.

Develops a framework for evolving deterministic formal policies to serve as runtime guardrails for large language model agents.

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

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

Introduces a multi-turn conversational benchmark to evaluate sales agents against resistant, psychologically grounded buyer personas.

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

Presents a simulation benchmark evaluating how embodied agents continually acquire basic skills and recombine them for complex 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.

Introduces a framework allowing agents to rewind long-horizon environments to earlier states while retaining learned exploratory knowledge.

Proposes JANUS, using a co-evolving world model to predict, verify, and guide exploration paths in autonomous scientific research.

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

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

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

Introduces ProgramGym, a framework providing verifiable sandbox environments and black-box test feedback for training long-horizon program reproduction agents.

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

Presents a multi-agent LLM framework that automatically tunes specific internal algorithm components while preserving overall neural network structures.

Introduces UPTURN, a framework that trains a small classifier to inject redirection prompts as new user turns during reasoning failures.

Combines load-adaptive hierarchical key-value caching with speculative decoding to accelerate long-context language model serving.

Examines error localization and credit assignment failures in critique-conditioned on-policy self-distillation for mathematical reasoning.

Introduces a Ukrainian meme dataset and evaluates vision-language models on detecting persuasive and propaganda techniques.

Applies vision-language models and structured annotation to parse and analyze historical comic books and manga at scale.

Combines dual-encoder semantic retrieval with a contextual bandit to adaptively select agent tools using execution reliability feedback.

Investigates how changing tasks redistributes attention head influence and alters internal activations in open-weight large language models.

Tests language models on graph-based information retrieval and navigation using the Wikipedia Game across multiple languages and citation datasets.