All papers

The paper proposes a rank-local data-loading policy that rebalances variable-cost training batches without cross-node coordination to speed up distributed training.

The authors propose JSB, a query-efficient jailbreak framework that uses multidimensional judge feedback and bandit strategies to guide prompt optimization.

The paper introduces a benchmark to evaluate text-to-video models across controlled emotional expression levels and specific audiovisual affective cues.

The authors analyze limitations in direct model editing benchmarks and compare them against observed generalization in practical chat settings.

Presents speculative recursive self-improvement to draft successor agents before full evaluation and optimize verification budgets.

The authors propose a theoretical and empirical framework to route token-level teacher selection during on-policy student distillation.

The paper introduces a training-free sampling framework that uses internal model activations to efficiently estimate rare failure probabilities in language models.

This paper presents a spiking mixture-of-experts architecture to improve the resilience of multimodal 3D occupancy perception against missing sensor inputs.

The paper re-evaluates coding-agent benchmark results to measure sensitivity to test execution environments and grading configurations.

Presents a framework allowing language model agents to evolve environment dynamics beliefs through falsifiable hypothesis testing.

This paper presents a pathwise conformal prediction method that guarantees valid coverage across adaptive early-exit stopping policies without multi-exit score inflation.

The paper identifies evaluation pitfalls in multimodal machine unlearning and introduces a benchmark grounded in real-world knowledge.

This paper presents a unified model for dense visual correspondence across 25 pairs of multi-modal sensor inputs.

The paper presents SciCurateBench to evaluate language models on verifying complex experimental conditions from unprocessed scientific literature.

Introduces AgenticBBO-Bench, a benchmark across multiple domains to evaluate LLM agents on black-box optimization under unified finite budgets.

The paper introduces a multi-agent framework that reframes harmful requests using Moral Foundations Theory to test model jailbreak vulnerabilities.

The paper presents OpenMobile-2, combining open mobile environments, trajectory datasets, and hybrid GUI-tool actions for mobile agents.

This paper introduces a pairwise evaluation benchmark and protocol designed to measure subjective aesthetic taste while controlling for shortcut artifacts.

Presents a distributed system architecture aggregating idle Apple Silicon devices into a virtual GPU cloud for inference.

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

The paper analyzes post-training in audio-language models to separate genuine audio reasoning gains from question-format memorization.

This paper introduces a method for multimodal models to reason over images using token-level latent regions instead of bounding-box coordinates.

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

Analyzes Mamba-3's input-dependent state rotations as oscillatory dynamics to explain their utility for tracking states and long-context language modeling.

Presents a framework that uses language models offline to generate executable feature-engineering code for tabular risk models.

The authors introduce MMA-SafetyBench to evaluate multimodal agent vulnerability against deceptive interface elements like fake system alerts.

This paper evaluates category name matching versus pairwise geometry metrics for candidate retrieval in context-constrained 3D vision-language models.

The authors present a framework where autonomous agents iteratively refine data engineering strategies to improve model training outcomes.

Introduces Sauron, an evidence-first agentic verifier that checks both environment state and execution traces against original sources to validate task rubrics.

Develops a category-theoretic framework to analyze and compare parameter-efficient fine-tuning methods as smooth maps in weight space.

This paper proposes a semantic reconstruction game to accurately evaluate jailbreak success by filtering out harmless repetitions and refusals.

The paper introduces a framework that distills iterative prompt search into rule modules for efficient text-to-video generation.

The paper audits historical evaluation tasks to see whether an agent's internal post-cutoff memory skews its point-in-time evidence gathering.

This paper redesigns 4-bit AdamW optimizer-state quantization by performing stochastic rounding directly within preconditioner space.

Benchmarks combinations of language models, communication protocols, and agent frameworks on web interaction tasks.

Introduces a benchmark to evaluate how language model agents learn and self-evolve reusable skills when exposed to non-adversarial noise.

The paper presents an architecture that decouples video and action prediction models using latent bridge queries.

The paper presents a framework and benchmark to evaluate whether self-evolving robotic agents genuinely learn reusable skills across tasks.

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

Proposes SWE-Critix, a post-training pipeline designed to train verifiers that evaluate and explain long coding-agent execution trajectories.

Presents a dual-kernel attention mechanism that uses 3D geometric positions to improve spatial reasoning in vision-language models.

This paper introduces a reinforcement learning approach that trains language models to execute safe fallbacks instead of unreliable guesses.

The paper presents a reinforcement learning framework that trains language models to solve complex logical reasoning tasks using interactive symbolic solvers.

Introduces Double-Take to mitigate capacity and cross-position binding failures in long-context test-time training architectures.

Evaluates how pretraining and test-time training computational budgets affect ARC benchmark performance across two distinct reasoning model architectures.

Models reward overoptimization in reinforcement learning with human feedback as economic rent-seeking driven by proxy reward signals.

This paper updates a long-horizon terminal benchmark to measure sustained agent progress across complex engineering workflows using dense process rewards.

The paper introduces an evaluation framework to measure the trade-off between task quality and computation cost when replacing iterative neural modules.

AffectWorld benchmarks world models on predicting emotion-driven human actions and subsequent environment changes across annotated video samples.

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

The paper introduces a framework to detect spurious shortcuts in safety benchmarks used to evaluate tool-using AI agents.

This paper frames Earth observation model development as an automated loop where coding agents build and iteratively refine models using verifiable harnesses.

The authors present a framework using level-wise multimodal memory to reconstruct coherent heading hierarchies across long, multi-page documents.

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

Presents RESCUE, a counterfactual simulation framework to evaluate whether runtime failure monitor alarms enable beneficial robot interventions.

The paper evaluates how choices in rating scale design systematically alter text evaluation scores produced by large language models.

The paper presents a runtime framework that tracks evolving repository state and dependencies to assist long-horizon coding agents.

The paper introduces a training-free technique and benchmark to help vision-language models correctly process and reject negated concepts in text prompts.

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

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