All papers

This paper introduces AdaRSI, an on-demand recursive self-improvement framework for embodied agents that isolates and fine-tunes only persistent sub-step failures.

Proposes LinMorph, a learnable weight transformation method to initialize linear attention models from pretrained video diffusion transformers efficiently.

Introduces a dataset and benchmark to model multimodal emotion understanding through cognitive appraisal reasoning rather than superficial affective cues.

This paper constructs an empirical graph of 62 private transformer inference systems to evaluate comparability, cryptographic progress, and scaling behavior.

This paper measures semantic drift and stability across AI agent communities, finding higher drift and lower temporal continuity than human forums.

Introduces DriveWLA, a mixture-of-transformers architecture unifying scene semantics, future dynamic prediction, and vehicle trajectory planning.

Surveys benchmarks, metrics, and experimental practices used to evaluate automated research systems across ideation, coding, writing, and peer review.

Analyzes why feature-based test-time adaptation fails to recover accuracy on corrupted inputs in generative multimodal language models.

This paper investigates whether gossip discourse naturally emerges in autonomous AI agent online communities compared to human discussion boards.

The paper introduces a training-free token pruning and merging framework with new robustness metrics for vision transformers.

Proposes a three-dimensional framework evaluating decision-making artificial intelligence methods across sequential formalisms, autonomy levels, and computational cost.

This paper analyzes the mathematical limits of statistical safety certification for embodied and agentic AI facing adaptive adversaries.

The paper introduces a framework that decomposes user feedback into atomic rules to retain only reusable preferences for personalizing language models.

The authors propose a visual token compression framework that reduces computational costs by retaining only visual tokens with high reconstruction residuals.

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

Studies statistical calibration techniques to bound risk when fast language models make automated decisions under uncertainty.

The paper proposes an agentic framework enabling multimodal models to identify information gaps and retrieve in-context demonstrations iteratively.

The paper introduces an information extraction benchmark for extracting structured regulatory fields and source provenance from dense environmental policy documents.

Proposes MedEvo, an iterative training framework that repairs multimodal medical reasoning trajectories using minimal hints and taxonomy-guided feedback.

The authors evaluate how developer scaffolding and general prompt instructions affect the performance of environment-agnostic coding agents across diverse benchmarks.

The authors propose a learned steering agent that monitors execution context and intervenes to prevent goal drift in tool-using language model systems.

Proposes a world-centric rotary positional embedding using 3D Gaussian primitives to improve cross-view attention in multi-view vision tasks.

The paper benchmarks twenty-three self-supervised 3D medical vision foundation models across segmentation, classification, and regression tasks.

Introduces a benchmark evaluating whether language model agents can independently direct and validate goal-specified physics research investigations.

The authors present a random projection method to protect sensitive datastore embeddings against extraction attacks in retrieval-augmented generation systems.

This paper introduces a benchmark of expert-annotated question-answer pairs to evaluate multimodal comprehension of long historical Chinese documents.

This paper introduces Unravel, a training-free framework that converts a single image into an editable, multi-object 3D scene.

The paper introduces an authorization system that reuses validated consequence evidence across agent execution branches while maintaining separate live authority.

The paper builds state graphs across multiple agent rollouts to evaluate behavioral consistency and predict execution failures.

This paper introduces a visual associative test benchmark to evaluate the lateral reasoning and creative association capabilities of vision-language models.

The paper proposes an image-to-video framework combining three coordinate spaces to enable camera-controllable video generation.

Introduces a reinforcement learning framework that evaluates memory operations in agents using a shared state and process rewards.

The paper separates task generation from model update inheritance to avoid compounding errors in self-evolving language model loops.

This paper presents a psychometric framework to test whether automated sentiment classification remains equivalent across different social media platforms.

The paper proposes a reasoning framework that discretizes visual states into tokens to jointly optimize visual computation and chain-of-thought steps.

Proposes a decentralized coordination mechanism using commitment tracking to preserve clear accountability across multi-agent delegation chains.

The authors evaluate how easily item identities and user profiles can be recovered from language embeddings in recommender systems.

This paper measures whether ecological alignment in large language models survives under pressure using a calibrated pretraining ladder.

The paper presents a dataset and framework mapping sketches and text to executable parametric computer-aided design code.

Presents an autonomous supervisor agent that iteratively tunes an agent's code, harness, and prompt on few examples.

The authors introduce a framework to benchmark language model agent robustness under interaction perturbations across diverse tasks.

Uses language model agents and deterministic translation to convert TLA+ formal specifications into machine-checkable Verus proofs.

This study analyzes a decade of crowdsourced disaster photos to evaluate vision-language models against human geographic reasoning for location estimation.

The paper proposes OpCredit, a policy learning framework that assigns individual operation credits for agentic knowledge graph question answering.

Investigates how partial interventions in concept-based models can unexpectedly reduce downstream task performance due to information leakage.

The paper introduces a standardized 25-task benchmark and evaluation protocol for docking-based molecular optimization in drug design.

This paper proposes a curriculum-guided reinforcement learning framework and an agentic reward model for spatial text-to-image generation.

Presents a benchmark with open-ended tasks and rubrics to assess social awareness and response generation in Chinese.

The paper analyzes multi-reward reinforcement learning failures where auxiliary objectives degrade primary performance and introduces a trust-anchored gating method.

The authors introduce 3DCodeVerse, a dataset and multi-agent framework enabling coding agents to generate, reason over, and verify 3D executable code.

The paper introduces a hybrid kinematic graph neural network and elastic transformer framework for efficient Indian sign language recognition from video.

The paper proposes a test-time aggregation method using bias control and injection to resolve disagreements among reasoning language model agents.

The paper introduces an agentic reasoning controller framework and an evaluation benchmark for optimizing open radio access networks.

The paper introduces a multi-head prediction method that injects future long-range supervision to improve autoregressive video generation consistency.

Proposes a rolling-denoising framework for vision-language-action models to improve robot control reactivity without sacrificing long-horizon execution consistency.

Estimates language model ensemble member reliability from unlabeled agreement patterns using an expectation-maximization jury formulation.

This paper introduces a multimodal unlearning method using counterfactual inputs and on-policy self-distillation to remove sensitive data without damaging general knowledge.

The paper studies how demonstration quality weighting, filtering, and conditioning affect vision-language-action model pretraining and downstream performance.

Introduces a controlled Minecraft benchmark to evaluate whether embodied agents acquire persistent, transferrable skills across environment resets.

Proposes an adversarial meta-prompt evolution method to improve model-generated evaluation rubrics for open-ended tasks.