Notableresearch
Neuron Populations Exhibit Divergent Selectivity with Scale
This paper explores how neuron populations in neural networks evolve with scale, revealing a sublinear growth pattern and increased selectivity in Rosetta Neurons.
AI Research Papers — Page 46
Archive, newest first. Filter by topic, impact and company →
Notableresearch
PixVOD: Pixel-Distributed Direct Visual Odometry and Depth Estimation
This paper introduces PixVOD, a pixel-distributed approach for visual odometry and depth estimation using Gaussian Belief Propagation for enhanced efficiency.
Notableresearch
Imaginative Perception Tokens Enhance Spatial Reasoning in Multimodal Language Models
This paper introduces Imaginative Perception Tokens (IPT) to enhance spatial reasoning in vision-language models, addressing limitations in unobservable contexts.
Notableresearch
NewtPhys: Do Foundation Models Understand Newtonian Physics?
This paper introduces NewtPhys, a dataset for evaluating foundation models' understanding of Newtonian physics through realistic visual scenarios.
Majorresearch
Humanoid-GPT: Scaling Data and Structure for Zero-Shot Motion Tracking
This paper presents Humanoid-GPT, a generative Transformer model that achieves zero-shot motion tracking through extensive pre-training on a large motion corpus.
Notableresearch
Language Models Compare Quantities Using Number-specific and Unit-specific Heuristics
This paper investigates how language models compare quantities with measurement units, revealing systematic errors and heuristic-based decision-making.
Notableresearch
Skill-RM: Unifying Heterogeneous Evaluation Criteria via Agent Skill
Skill-RM introduces a unified framework for reward modeling in LLMs, enhancing evaluation consistency and performance across diverse tasks.
Notableresearch
Language Models Need Sleep: Learning to Self-Modify and Consolidate Memories
This paper introduces a "Sleep" paradigm for LLMs, enabling continual learning and memory consolidation through a novel distillation process.
Formalizing the Binding Problem
This paper formalizes the binding problem in visual recognition, introducing a method to measure binding information in Vision Transformers.
Notableresearch
AAD-1: Asymmetric Adversarial Distillation for One-Step Autoregressive Video Generation
AAD-1 introduces an Asymmetric Adversarial Distillation framework that enhances one-step autoregressive video generation by addressing motion collapse and training instability.
Notableresearch
Video-Mirai: Autoregressive Video Diffusion Models Need Foresight
This paper introduces Video-Mirai, a method that enhances autoregressive video diffusion models by incorporating foresight to improve future frame consistency.
Majorresearch
Quantifying Faithful Confidence Expression in Large Reasoning Models
This paper introduces a framework for quantifying faithful confidence expression in large reasoning models, addressing a critical gap in uncertainty communication.
Notableresearch
QUBRIC: Co-Designing Queries and Rubrics for RL Beyond Verifiable Rewards
This paper introduces QUBRIC, a framework that co-designs queries and rubrics to enhance reinforcement learning beyond verifiable rewards.
AlignAtt4LLM: Fast AlignAtt for Decoder-Only LLMs at IWSLT 2026 Simultaneous Speech Translation Task
This paper introduces AlignAtt4LLM, a novel approach for simultaneous speech translation using decoder-only LLMs, achieving significant performance improvements.
Notableresearch
Agentic Chain-of-Thought Steering for Efficient and Controllable LLM Reasoning
This paper introduces Agentic Chain-of-Thought Steering (ACTS), a method for efficient and controllable reasoning in large language models using reinforcement learning.
Majorresearch
Self-Refining Agentic Reinforcement Learning for Vision-Conditioned UAV Navigation
This paper introduces AgenticRL, a self-refining framework for enhancing UAV navigation through autonomous reward design and policy refinement.
Notableresearch
Using Reward Uncertainty to Induce Diverse Behaviour in Reinforcement Learning
This paper introduces a novel reinforcement learning framework that incorporates reward uncertainty to promote diverse agent behavior without sacrificing performance.
Notableresearch
Efficient ASR Training with Conversations that Never Happened
This paper presents a novel augmentation pipeline for generating synthetic conversational data to enhance ASR training in low-resource languages.
Notableresearch
VLESA: Vision-Language Embodied Safety Agent for Human Activity Monitoring
This paper presents VLESA, a framework for real-time safety monitoring of human activities using egocentric video and goal-conditioned safety assessments.
Majorresearch
Demo2Tutorial: From Human Experience to Multimodal Software Tutorials
This paper presents Demo2Tutorial, a framework that converts human interaction data into structured multimodal software tutorials, enhancing learning for both humans and agents.
Notableresearch
A Pocket Offline Model for Simultaneous Speech Translation as CUNI Submission to IWSLT 2026
This paper presents a compact offline model for simultaneous speech translation, achieving high quality and low computational requirements across multiple languages.
Notableresearch
MLSkip: Data Skipping for ML Filters via Lightweight Metadata
This paper introduces MLSkip, a novel data skipping technique for ML filters using lightweight metadata, enhancing query efficiency in data management.
Notableresearch
SEAOTTER: Sensor Embedded Autoencoding with One-Time Transcode for Efficient Reconstruction
This paper presents SEAOTTER, a novel compression framework that integrates learned autoencoding with standard JPEG for efficient visual data reconstruction in robotics.
Notableresearch
FlashbackCL: Mitigating Temporal Forgetting in Federated Learning
This paper introduces FlashbackCL, a method to mitigate temporal forgetting in Federated Learning, enhancing model performance under data distribution shifts.
q0: Primitives for Hyper-Epoch Pretraining
This paper introduces hyper-epoch pretraining (q0), a novel approach that enhances model diversity and efficiency in multi-epoch training.
Notableresearch
Entropy Is Not Enough: Unlocking Effective Reinforcement Learning for Visual Reasoning via Vision-Anchored Token Selection
This paper introduces VEPO, a novel reinforcement learning framework that enhances visual reasoning by integrating visual sensitivity with token entropy.
Correcting Neural Operator Spectral Bias via Diffusion Posterior Sampling with Sparse Observations
This paper introduces FreqNO-DPS, a method that mitigates spectral bias in neural operator surrogates using diffusion posterior sampling with sparse observations.
Notableresearch
Quadratic integrate-and-fire neurons exhibit less fragmented loss landscapes and outperform leaky integrate-and-fire neurons in spike-based gradient descent
This paper demonstrates that quadratic integrate-and-fire neurons provide smoother loss landscapes and superior performance over leaky integrate-and-fire neurons in spiking neural networks.
Notableresearch
Value-Aware Stochastic KV Cache Eviction for Reasoning Models
This paper introduces Value-aware Stochastic KV Cache Eviction (VaSE), enhancing reasoning model efficiency while maintaining accuracy through improved cache management.
Notableresearch
FFR: Forward-Forward Learning for Regression
This paper introduces FFR, a novel framework extending Forward-Forward learning to regression tasks, achieving competitive accuracy with reduced resource consumption.