Page 34 - Paper Library - AiPaper

InterPLM: Discovering Interpretable Features in Protein Language Models via Sparse Autoencoders

Published:11/13/2024

Feature Extraction in Protein Language ModelsApplication of Sparse AutoencodersInterpretable Protein BiologyHuman-Interpretable Latent FeaturesAnalysis of ESM-2 Model

This paper introduces a method using sparse autoencoders to extract interpretable features from protein language models, revealing up to 2,548 features correlated with 143 biological concepts, and demonstrating applications in database annotation and targeted protein generation.

010

SCALING LARGE LANGUAGE MODELS FOR NEXT-GENERATION SINGLE-CELL ANALYSIS

Published:4/17/2025

Large Language Model Fine-TuningSingle-Cell RNA SequencingCell Text ModelingBiological Information SynthesisMulticellular Context Reasoning

This study introduces a novel approach using the Cell2Sentence framework to convert singlecell RNA sequencing data into textual 'cell sentences,' training large language models on over a billion tokens. Scaling to 27 billion parameters resulted in enhanced performance in multice

RaC: Robot Learning for Long-Horizon Tasks by Scaling Recovery and Correction

Published:9/10/2025

Robot Learning for Long-Horizon TasksHuman-in-the-Loop Recovery and CorrectionRobotic Policy Fine-TuningBimanual Control TasksEfficiency and Robustness Enhancement

The paper introduces , a method that enhances robot learning for longhorizon tasks by scaling recovery and correction behaviors, finetuning robotic policies through human interventions, thus improving efficiency and robustness in complex tasks.

ShiDianNao: Shifting Vision Processing Closer to the Sensor

Image Application-Specific Neural Network AcceleratorOptimization of Convolutional Neural NetworksNear-Sensor Architecture DesignHigh-Efficiency Neural Network Accelerator65nm Layout Design

The ShiDianNao CNN accelerator is placed next to CMOS or CCD sensors to eliminate DRAM accesses, achieving 60x energy efficiency improvement and 30x faster performance than highend GPUs, with a compact design of 4.86mm² area and 320mW power consumption.

INT v.s. FP: A Comprehensive Study of Fine-Grained Low-bit Quantization Formats

Published:10/29/2025

Low-bit Quantization FormatsFine-grained Quantization ComparisonFloating-point and Integer QuantizationAI Hardware OptimizationFine-grained INT Training

This study offers a systematic comparison between floatingpoint (FP) and integer (INT) quantization formats, revealing that MXINT8 outperforms FP in 8bit finegrained formats. For 4bit formats, FP often excels, but NVINT4 can surpass it with outliermitigation techniques. A ne

A genome-to-proteome map reveals how natural variants drive proteome diversity and shape fitness

Published:10/9/2025

Genome-to-Proteome MappingNatural Variation and Proteome DiversityImpact of Natural Genetic VariantsYeast Strain Genetic StudyGenotype-Phenotype Relationship Study

This study presents a nucleotideresolution genometoproteome map, revealing how natural genetic variants drive proteome diversity and enhance fitness. Analyzing meiotic progeny from yeast strains highlights the interplay of small variations and selection mechanisms, underlining

DaDianNao: A Machine-Learning Supercomputer

Published:12/1/2014

Machine-Learning SupercomputerConvolutional Neural Network AcceleratorDeep Neural Network ArchitectureMulti-Chip System DesignHigh-Performance Computing Optimization

DaDianNao is a machinelearning supercomputer optimized for CNNs and DNNs, demonstrating a 450.65x speedup and 150.31x energy reduction compared to GPUs, effectively addressing the high computational and memory demands of machine learning.

Physics-Based Dexterous Manipulations with Estimated Hand Poses and Residual Reinforcement Learning

Published:8/7/2020

Residual Reinforcement Learning and Imitation Learning3D Hand Pose Estimation and Physics SimulationDexterous Manipulation in Virtual EnvironmentsGesture-Based Object InteractionPhysics-Guided Target Pose Remapping

This study presents a residual reinforcement learning approach that enables agents to perform dexterous manipulation in virtual environments by mapping estimated hand poses to target poses, effectively overcoming the challenges of absent physical feedback.

DreamControl: Human-Inspired Whole-Body Humanoid Control for Scene Interaction via Guided Diffusion

Published:9/18/2025

Diffusion Model-Based Robot ControlWhole-Body Humanoid Action LearningHuman Motion Data-Guided Reinforcement LearningSim-to-Real Motion TransferTask Execution on Unitree G1 Robot

DreamControl introduces a novel method for learning wholebody skills in humanoid robots by combining diffusion models and reinforcement learning, enabling complex tasks and facilitating simtoreal transfer.

Omnigrasp: Grasping Diverse Objects with Simulated Humanoids

Published:7/16/2024

Grasp Control with Simulated HumanoidsGrasping and Moving Diverse ObjectsHumanoid Motion Representation LearningTraining without Paired DatasetObject Trajectory Following Task

Omnigrasp is a method for controlling simulated humanoids to grasp and manipulate over 1200 diverse objects along predefined trajectories. It enhances control accuracy through humanoid motion representation, requiring no paired training data and demonstrating excellent scalabilit

Stable-Predictive Optimistic Counterfactual Regret Minimization

Published:2/14/2019

Counterfactual Regret MinimizationStable-Predictive Regret MinimizationLarge-Scale Game SolvingConvergence Rate OptimizationStability in Decision Trees

This paper introduces a new CFR variant achieving

O(T^{3/4})

convergence rate for largescale extensiveform games. By combining advances in predictive and stable regret minimization, the concept of 'stablepredictivity' enhances algorithm performance beyond traditional CFR.

Learning Smooth Humanoid Locomotion through Lipschitz-Constrained Policies

Published:10/16/2024

Lipschitz-Constrained PoliciesSmooth Locomotion Control for Legged RobotsReinforcement Learning and Sim-to-Real TransferDevelopment of Smooth Behaviors for RobotsLow-Pass Filtering and Smoothness Rewards

This paper introduces LipschitzConstrained Policies (LCP) to enhance humanoid robot locomotion control. LCP enforces smooth behaviors in a reinforcement learning framework, replacing traditional smoothing rewards, and integrates easily with automatic differentiation. Experiments

Spatial Forcing: Implicit Spatial Representation Alignment for Vision-language-action Model

Published:10/14/2025

Vision-Language-Action ModelEnhanced Spatial Understanding CapabilitiesImplicit Spatial Representation AlignmentAlignment with 3D Foundation ModelsPrecise Execution of Robotic Tasks

This paper introduces 'Spatial Forcing' (SF), an implicit alignment method enhancing spatial understanding in VisionLanguageAction (VLA) models. By aligning visual embeddings with 3D foundation models, SF improves robotics' operational precision in 3D environments without relyi

$π_\texttt{RL}$: Online RL Fine-tuning for Flow-based Vision-Language-Action Models

Published:10/30/2025

Flow-based Vision-Language-Action ModelsOnline Reinforcement Learning Fine-TuningLIBERO BenchmarkMultitask Reinforcement LearningDenoising Modeling in Environment Interaction

The paper introduces the

π ext{RL}

framework, using online reinforcement learning to finetune flowbased VisionLanguageAction models, addressing challenges in action loglikelihoods. It demonstrates significant performance improvements on LIBERO and ManiSkill benchmarks.

ManiSkill-HAB: A Benchmark for Low-Level Manipulation in Home Rearrangement Tasks

Published:12/9/2024

Low-Level Manipulation BenchmarkIn-Home Object Rearrangement TasksReinforcement Learning and Imitation Learning BaselinesGPU-Accelerated Home Assistant BenchmarkData Generation and Demonstration Filtering

ManiSkillHAB introduces a benchmark for lowlevel manipulation in home rearrangement tasks, addressing the need for faster simulations and complex environments. It features GPU acceleration, enhanced speed, extensive reinforcement and imitation learning baselines, and a rulebas

DisCO: Reinforcing Large Reasoning Models with Discriminative Constrained Optimization

Published:5/18/2025

RL Training for Large Language ModelsGroup Relative Policy OptimizationDiscriminative Constrained Optimization FrameworkLarge Reasoning ModelsEnhancement of Mathematical Reasoning Capabilities

DisCO is a new framework for Large Reasoning Models, addressing limitations of Group Relative Policy Optimization. By using a discriminative objective and nonclipping scoring functions, it eliminates difficulty bias and achieves stable longterm training, enhancing mathematical

Why Low-Precision Transformer Training Fails: An Analysis on Flash Attention

Published:10/5/2025

Analysis of Low-Precision Transformer Training FailuresFlash Attention MechanismTraining Dynamics StabilityLow-Rank Representations and Bias ErrorsError Accumulation in Model Training

This paper explains the loss explosion in lowprecision transformer training as a result of lowrank representations and biased rounding errors, leading to a vicious cycle of error accumulation. A simple modification to Flash Attention stabilizes the training process.

014

Transformations in exposure to debris flows in post-earthquake Sichuan, China

Post-Earthquake Debris Flow ExposureDebris Flow Simulation and Assessment in SichuanImpact of Drainage Structures on Debris FlowsHigh-Resolution Satellite Imagery AnalysisUrban Development and Natural Disaster Interactions

This study examines how catchment interventions in three gullies in Sichuan, postearthquake, affect debris flow exposure. Findings indicate urban development increased the risk of a 2019 debris flow, and check dams effectively manage low and high flow events, but fail against ex

Distributed LLM Serving on Consumer-Grade GPUs by Reconciling Computation and Communication

Published:1/1/2025

Distributed LLM Serving on Consumer-Grade GPUsMoLink Serving SystemPrefill Request Transmission Scheduling AlgorithmCommunication Efficiency Optimization for LLMsDistributed Inference Computing Architecture

This paper presents MoLink, an efficient distributed LLM serving system that reduces costs using consumergrade GPUs. It splits the prefill request data into smaller chunks and optimizes transmission scheduling, achieving up to 46% reductions in firsttoken generation time, pert

Order-agnostic Identifier for Large Language Model-based Generative Recommendation

Published:2/15/2025

LLM-based Generative Recommendation SystemsOrder-Agnostic Identifier DesignIntegration of Collaborative Filtering and Semantic InformationSETRec FrameworkSparse Attention Mechanism

This paper presents an orderagnostic identifier design for LLMbased generative recommendations, addressing efficiency and performance issues. By integrating CF and semantic information using the SETRec framework, it significantly enhances recommendation effectiveness and genera

661 - 680 / 991

Papers