2/27/2026
Simplex-to-Euclidean Bijections for Categorical Flow Matching
Method for learning probability distributions on simplexes via smooth bijections and Aitchison geometry.
2/27/2026
UniQL: Unified Quantization and Low-rank Compression for Adaptive Edge LLMs
UniQL framework combining quantization and low-rank compression with adaptive on-device pruning for edge LLM deployment.
2/27/2026
Sparse Attention Post-Training for Mechanistic Interpretability
Post-training method achieving 99.6% attention sparsity without performance loss for mechanistic interpretability research.
2/27/2026
WebGym: Scaling Training Environments for Visual Web Agents with Realistic Tasks
WebGym: largest open-source environment with 300k tasks for training visual web agents on realistic websites.
2/27/2026
A Confidence-Variance Theory for Pseudo-Label Selection in Semi-Supervised Learning
Confidence-Variance theory framework for improved pseudo-label selection in semi-supervised learning beyond fixed thresholds.
2/27/2026
Rethinking Cross-Modal Fine-Tuning: Optimizing the Interaction between Feature Alignment and Target Fitting
Method for optimizing interaction between feature alignment and target fitting in cross-modal model fine-tuning.
2/27/2026
Learning Hamiltonian Flow Maps: Mean Flow Consistency for Large-Timestep Molecular Dynamics
Framework for learning Hamiltonian flow maps to enable stable large-timestep molecular dynamics simulations.
2/27/2026
Beyond Fixed Rounds: Data-Free Early Stopping for Practical Federated Learning
Data-free early stopping framework for federated learning using task vector growth rate monitoring.
2/27/2026
Agentic Framework for Epidemiological Modeling
EPIAGENT agentic framework that automatically synthesizes and calibrates epidemiological simulators via iterative program synthesis.
2/27/2026
A Minimum Variance Path Principle for Accurate and Stable Score-Based Density Ratio Estimation
Framework for score-based density ratio estimation addressing path-variance issues in practical training objectives.
2/27/2026
Phase Transitions for Feature Learning in Neural Networks
Theoretical analysis of phase transitions in neural network feature learning on multi-index models.
2/27/2026
Detecting Misbehaviors of Large Vision-Language Models by Evidential Uncertainty Quantification
Technique to detect misbehaviors and hallucinations in large vision-language models using evidential uncertainty quantification.
2/27/2026
Learning Credal Ensembles via Distributionally Robust Optimization
Method for training ensemble models that quantify epistemic uncertainty via distributionally robust optimization.
2/27/2026
When Less is More: The LLM Scaling Paradox in Context Compression
Study of LLM scaling paradox showing larger compressor models can reduce context reconstruction faithfulness despite lower training loss.
2/27/2026
Versor: A Geometric Sequence Architecture
Novel sequence architecture using Conformal Geometric Algebra instead of linear operations for improved generalization and interpretability.
2/27/2026
Learning beyond Teacher: Generalized On-Policy Distillation with Reward Extrapolation
On-policy distillation method that aligns student models with teacher logit distributions, theoretically framed as KL-constrained RL with reward extrapolation.
2/27/2026
\"UberWeb: Insights from Multilingual Curation for a 20-Trillion-Token Dataset
Analysis of multilingual data curation across 13 languages for 20-trillion-token dataset, addressing multilinguality challenges in foundation models.
2/27/2026
Soft Sequence Policy Optimization
Introduces Soft Sequence Policy Optimization, advancing LLM alignment methods beyond GRPO with sequence-level importance sampling.
2/27/2026
On the Equivalence of Random Network Distillation, Deep Ensembles, and Bayesian Inference
Theoretical analysis connecting random network distillation, deep ensembles, and Bayesian inference for uncertainty quantification in deep learning.
2/27/2026
Why Pass@k Optimization Can Degrade Pass@1: Prompt Interference in LLM Post-training
Shows Pass@k optimization can degrade Pass@1 due to prompt interference, revealing trade-offs in LLM fine-tuning for code generation.
2/27/2026
AngelSlim: A more accessible, comprehensive, and efficient toolkit for large model compression
AngelSlim comprehensive toolkit for large model compression combining quantization, pruning, distillation, and speculative decoding.
2/27/2026
Muon+: Towards Better Muon via One Additional Normalization Step
Muon+ optimizer improves LLM pre-training by adding normalization step after gradient orthogonalization.
2/27/2026
Knowledge Fusion of Large Language Models Via Modular SkillPacks
Knowledge Fusion via SkillPacks enables efficient cross-capability transfer between LLMs for multi-task integration and model compression.
2/27/2026
On the Interpolation Error of Nonlinear Attention versus Linear Regression
Theoretical analysis of nonlinear attention mechanisms compared to linear regression for understanding interpolation error.
2/27/2026
Adaptive Hybrid Caching for Efficient Text-to-Video Diffusion Model Acceleration
Adaptive hybrid caching optimization for efficient inference in video diffusion transformer models to reduce computational cost.
2/27/2026
Generative Value Conflicts Reveal LLM Priorities
ConflictScope pipeline automatically evaluates how LLMs prioritize different values when facing conflicting objectives.
2/27/2026
Supervised Reinforcement Learning: From Expert Trajectories to Step-wise Reasoning
Supervised Reinforcement Learning method for training small open-source LLMs on multi-step reasoning tasks, combining SFT and RLVR approaches.
2/27/2026
Inference-Cost-Aware Dynamic Tree Construction for Efficient Inference in Large Language Models
Method for efficient LLM inference via speculative decoding with dynamic tree construction accounting for system variables.
2/27/2026
Temporal Sparse Autoencoders: Leveraging the Sequential Nature of Language for Interpretability
Temporal Sparse Autoencoders using sequential language structure for discovering interpretable features in LLM representations.
2/27/2026
Intelligence per Watt: Measuring Intelligence Efficiency of Local AI
Analysis of energy efficiency of small LLMs on local hardware accelerators versus cloud inference for practical deployment.
2/27/2026
VLM-Pruner: Buffering for Spatial Sparsity in an Efficient VLM Centrifugal Token Pruning Paradigm
VLM-Pruner: Token pruning method for vision-language models addressing spatial sparsity and inter-token redundancy.
2/27/2026
One-Step Diffusion Samplers via Self-Distillation and Deterministic Flow
One-step diffusion samplers using self-distillation and deterministic flow for efficient sampling from unnormalized distributions.
2/27/2026
Imitation Game: Reproducing Deep Learning Bugs Leveraging an Intelligent Agent
Intelligent agent system for automatically reproducing deep learning bugs by leveraging nondeterminism and environment coupling.
2/27/2026
LeanCat: A Benchmark Suite for Formal Category Theory in Lean (Part I: 1-Categories)
LeanCat: Benchmark of 100 formalized category theory tasks in Lean evaluating LLMs on library-grounded abstraction and theorem proving.
2/27/2026
Controlling Exploration-Exploitation in GFlowNets via Markov Chain Perspectives
Framework connecting GFlowNets to Markov chain reversibility to control exploration-exploitation trade-off during training.
2/27/2026
The Wisdom of Many Queries: Complexity-Diversity Principle for Dense Retriever Training
Study of multi-query synthesis for dense retriever training showing quality-diversity trade-off benefits out-of-domain and multi-hop retrieval.
2/27/2026
GPT-4o Lacks Core Features of Theory of Mind
Evaluation showing GPT-4o lacks causal models of mental states required for true Theory of Mind despite benchmark performance.
2/27/2026
Large-scale online deanonymization with LLMs
Agent with internet access performs at-scale deanonymization of Hacker News and interview participants using LLMs.
2/27/2026
LLM4Cov: Execution-Aware Agentic Learning for High-coverage Testbench Generation
LLM4Cov: Offline agent-learning framework using execution feedback from hardware simulators for test generation with high coverage.
2/27/2026
Scaling Laws for Precision in High-Dimensional Linear Regression
Theoretical analysis of how low-precision quantization affects model and data capacities in high-dimensional linear regression.
2/27/2026
Not Just How Much, But Where: Decomposing Epistemic Uncertainty into Per-Class Contributions
Method to decompose epistemic uncertainty in Bayesian deep learning by per-class contributions for asymmetric cost classification tasks.
2/27/2026
Under the Influence: Quantifying Persuasion and Vigilance in Large Language Models
Research on LLM capacity to be persuaded and detect manipulation, testing vigilance and persuasion in high-stakes decision-making contexts.
2/27/2026
Multilingual Safety Alignment Via Sparse Weight Editing
Research paper on sparse weight editing for multilingual LLM safety alignment across low-resource languages without expensive retraining.
2/27/2026
Show HN: Agentplace, the tool we built to become a 20x company
Agentplace tool for building and automating AI agents at scale. Addresses challenges in agent development and scaffolding.
2/27/2026
Microsoft's Copilot Tasks AI uses its own computer to get things done
Microsoft Copilot Tasks uses AI agent to autonomously complete user tasks like converting emails to slideshows.
2/27/2026
Show HN: Ambit-OpenCode – Cloud IDE with Seamless Mobile <> Desktop Handoff
Cloud IDE with agent-driven features, mobile-desktop handoff, built-in shell. Open-source, $5/month after free trial.
2/27/2026
Show HN: Agent Bazaar – Billing and metering for MCP tool servers
Billing platform for MCP tool servers enabling developers to monetize AI agent tools. Stripe integration, TypeScript SDK, live with 6 tools.
2/27/2026
Would you use LLM:// connection URLs?
Proposal for llm:// URI scheme to standardize LLM connection URLs. Led to draft IETF RFC.
2/27/2026
Sharesight MCP
Model Context Protocol server integrating Sharesight portfolio platform with Claude AI assistants.
2/27/2026
HeadElf – C-Suite Executive Intelligence System
Claude Code extension providing autonomous C-suite executive decision-making across multiple executive functions.