29d ago
KernelSight-LM: A Kernel-Level LLM Inference Simulator
arXiv paper presenting KernelSight-LM, a simulator for evaluating LLM inference performance across hardware and serving parameters.
29d ago
Four Types of LLM Reliance and Their Predictors Among Undergraduate Writers: A Mixed-Methods Study at a Minority-Serving R1 University
arXiv paper studying undergraduate reliance patterns on LLMs for academic writing. Identifies four types and predictors of LLM use.
29d ago
Translating Natural Language to Strategic Temporal Specifications via LLMs
arXiv paper on using LLMs to translate natural language into formal temporal specifications for multi-agent systems verification.
29d ago
Resolving superposition in AI for interpretability and cross-modal alignment in patient-neuronal images
arXiv paper on resolving superposition in neural networks for biological data interpretability and cross-modal alignment in patient imaging.
29d ago
WorldOdysseyBench: An Open-World Benchmark for Long-Horizon Stability of Interactive World Models
arXiv paper introducing WorldOdysseyBench benchmark for evaluating long-horizon stability of interactive world models across four dimensions.
29d ago
Look But Don't Touch with Sparse Autoencoders for Unlearning in Diffusion Models
arXiv paper evaluating sparse autoencoders for concept manipulation and unlearning in diffusion models, focusing on object erasure.
29d ago
GR2 Technical Report
arXiv paper on using LLMs for re-ranking in industrial recommendation systems. Addresses gaps in adopting LLMs for multi-stage ranking.
29d ago
BaRA: Budget-constrained and Reliable Web Data Collection Agent
arXiv paper proposing BaRA, a budget-constrained LLM web agent for multimodal data collection with fixed interaction budgets.
29d ago
Learning User-Aware Recall: Personalized Retrieval in Long-Term Conversational Memory
arXiv paper on personalized retrieval for long-term conversational agents using profile-guided memory recall in LLM-based systems.
29d ago
From World Models to World Action Models: A Concise Tutorial for Robotics
Tutorial on world models as action-conditioned predictive models for embodied AI, comparing observation-space vs state-space approaches with trade-offs.
29d ago
MemSyco-Bench: Benchmarking Sycophancy in Agent Memory
MemSyco-Bench evaluates sycophancy in LLM-agent memory systems where retrieved memories cause over-alignment with users at cost of factual accuracy.
29d ago
Staleness-Learning Rate Scaling Laws for Asynchronous RLHF
Study of stale rollout effects in asynchronous GRPO for high-throughput RLHF, analyzing learning-rate scaling laws for decoupled policy optimization.
29d ago
Fixed-Set Robustness in Programming by Example: Example Corruption and Semantic Partition Recovery
Study of adversarial robustness in programming-by-example systems when examples are corrupted by adversaries aware of the synthesizer.
29d ago
Scaling Laws for Grid-Based Approximate Nearest Neighbor Search in High Dimensions
Scaling analysis of grid-based approximate nearest neighbor search revealing d-scaling crossover behavior on embeddings as dimensionality increases.
29d ago
From Approximation to Emergence: A Theory of Deep Learning
Comprehensive theoretical book/paper unifying deep learning theory from approximation foundations through overparameterization, transformers, in-context learning, scaling laws, and emergence.
29d ago
Black-Box Inference of LLM Architectural Properties with Restrictive API Access
Methods to infer LLM architectural properties (hidden dimensions, feed-forward layers) via black-box API access with restricted logits, studying commercial provider protections.
29d ago
The Rollout Infrastructure Tax in Coding-Agent Reinforcement Learning
Analysis of execution infrastructure overhead in coding-agent RL systems, measuring efficiency gains from different container/sandbox substrates for interactive rollouts.
29d ago
Conditional Inference Trees and Forests for Feature Selection
Study of conditional inference trees/forests for feature selection using permutation tests, comparing computational efficiency vs ranking accuracy on benchmarks.
29d ago
On the Utility and Factual Reliability of Pruned Mixture-of-Experts Models in the Biomedical Domain
Evaluates factual reliability of pruned mixture-of-experts models in biomedical domain, examining trade-offs between inference speedup and accuracy.
29d ago
Token Geometry
Ember lightweight optimizer exploits gradient geometry of embedding tables and LM-heads, improving Pareto frontier for finetuning, RL, and pretraining.
29d ago
Class-Grouped Normalized Momentum and Faster Hyperparameter Exploration to Tackle Class Imbalance in Federated Learning
FedCGNM optimizer for federated learning addresses class imbalance via class-grouped momentum and faster hyperparameter exploration.
29d ago
How to Allocate Your Tokens? Scaling Laws with Training Steps and Batch Size
Three-term scaling law for LLM training explicitly modeling batch size and training steps, enabling robust fitting with fewer training runs.
29d ago
Don't Let Gains FADE: Breaking Down Policy Gradient Weights in RL
Framework decomposing advantage functions for RL post-training in LLMs, unifying diverse advantage formulations to address training instability.
29d ago
Unveiling the Non-Monotonic Effect of Privacy on Generalization under Byzantine Robustness
Shows privacy-generalization relationship in distributed learning depends on noise regime, contradicting prior Byzantine robustness trilemma results.
29d ago
Towards Learning Representations of Policies in Two-Player Zero-Sum Imperfect-Information Games
Methods for creating policy datasets and learning policy embeddings in two-player zero-sum imperfect-information games with evaluation tasks.
29d ago
The risk of KV cache compression
Theoretical analysis of KV cache compression in transformer inference showing when compression is impossible and deriving fundamental limits.
29d ago
Multi-Head Recurrent Memory Agents
Multi-Head Recurrent Memory Agents diagnose reliability degradation in long-context LLMs, attributing failures to memory retention rather than capture.
29d ago
Evolutionary Feature Engineering for Structured Data
EFE framework uses LLM-based evolutionary optimization to discover preprocessing transformations for structured data as Python programs.
29d ago
X-LogSMask: Expand Transformer for Graph-Structured Data
X-LogSMask modifies transformer architecture with explainable multi-head attention for improved performance on sparse, structured graph data.
29d ago
Geometric Signatures of Reasoning: A Spectral Perspective on Task Hardness
Studies geometric properties of chain-of-thought reasoning trajectories in transformer hidden states to understand task difficulty and reasoning mechanisms.
29d ago
BOUNDARY_SYNC: Measuring Communication-Induced Representational Coupling in Multi-Agent LLM Systems
BOUNDARY_SYNC measures representational coupling in multi-agent LLM systems, quantifying how inter-agent communication causes convergence or divergence.
29d ago
SINA: A Fully Automated Circuit Schematic Image to Netlist Generator Using Artificial Intelligence
SINA uses AI to convert circuit schematic images to machine-readable netlists for electronic design automation tasks.
29d ago
DeadPool: Resilient LLM Training with Hot-Swapping via Zero-Overhead Checkpoint
DeadPool enables resilient LLM training at scale by implementing hot-swapping with zero-overhead checkpointing for GPU failure recovery.
29d ago
SCAPE: Accurate and Efficient LLM Training with Extreme Sparse Communication
Communication-efficient LLM training via extreme sparse gradient synchronization with stable sparse Adam optimizer.
29d ago
WARP: Weight-Space Analysis for Recovering Training Data Portfolios
Method for inferring training data mixture weights and source distributions from released foundation models.
29d ago
Model Merging as Probabilistic Inference in Fine-Tuning Parameter Space
Probabilistic inference framework for merging task-specific fine-tuned models into multi-task solutions.
29d ago
A Mathematical Introduction to Diffusion Models
Mathematical introduction to diffusion models covering sampling dynamics, error analysis, and inference-time control.
29d ago
Frequency Shift Physics-Informed Extreme Learning Machine for Solving High-Frequency Partial Differential Equations
Physics-informed extreme learning machine addressing spectral bias for solving high-frequency PDEs.
29d ago
Predicting Closed-Loop Performance of Latent World Models: Offline Checkpoint Selection for MPC and Model-Based RL Under Non-Markovian Rewards in LunarLander
Validation diagnostics for selecting optimal checkpoints of latent world models in model predictive control and model-based RL.
29d ago
Role-Aware Neural Convex Divergence Heads for Asymmetric Representation Learning
Neural divergence heads for asymmetric representation learning in directed relational tasks.
29d ago
Denser $\neq$ Better: Limits of On-Policy Self-Distillation for Continual Post-Training
Analysis of self-distillation for continual post-training showing trade-offs between in-domain specialization and knowledge preservation.
29d ago
Set Diffusion: Interpolating Token Orderings Between Autoregression and Diffusion for Fast and Flexible Decoding
Discrete diffusion model for language generation combining autoregressive and diffusion decoding with flexible token ordering.
29d ago
EPnG: Adaptive Expert Prune-and-Grow for Parameter-Efficient MoE Fine-tuning
Adaptive prune-and-grow framework for parameter-efficient fine-tuning of Mixture-of-Experts models using LoRA.
29d ago
Expander Sparse Autoencoders: Parameter-Efficient Dictionaries for Mechanistic Interpretability
Parameter-efficient sparse autoencoders for interpreting neural network activations using expander graphs.
29d ago
Do LLMs Truly Generalize in the Molecular Domain? A Perturbation-Based Analysis
Research on whether LLMs generalize in molecular discovery tasks beyond local neighborhoods of sequence representations.
29d ago
Many Voices, One Reward: Multi-Role Rubric Generation for LLM Judging and Reward Modeling
Multi-role rubric generation for LLM evaluation addressing dimensional blind spots in preference-based reward modeling.
29d ago
Decomposer: Learning to Decompile Symbolic Music to Programs
Decomposer: LLM post-training framework for symbolic music decompilation recovering executable music programs from MIDI.
29d ago
Rank-Then-Act: Reward-Free Control from Frame-Order Progress
Rank-Then-Act: Reward-free policy learning from expert videos using vision-language models and ordinal scoring.
29d ago
SABER: A Semantic-Aligned Brain Network Analysis Framework via Multi-scale Hypergraphs
SABER: Brain network analysis framework integrating LLM semantics with multi-scale hypergraphs for disease diagnosis.
29d ago
Conditional Co-Ablation: Recovering Self-Repair Backups in Transformer Circuits
Mechanistic interpretability study of transformer self-repair mechanisms during ablation using conditional co-ablation analysis.