6/9/2026
Does a token buy you more or less now than it did a few months ago?
First-of-its-kind analysis of thousands of conversations measuring token value changes and invisible failures in coding assistants over recent months.
6/9/2026
Productivity Effects Across Generations of AI Coding Tools
Research paper analyzing macroeconomic productivity effects of LLM coding tools across generations, showing gaps between token generation claims and actual shipping velocity.
6/9/2026
Do Better Research with NotebookLM
Google upgrades NotebookLM with agentic reasoning capabilities, code execution, and multiformat output generation for research projects.
6/9/2026
Stop Asking Claude to Agree with You
Best practices guide on using Claude effectively by rejecting auto-agreement, using plan mode critically, and ensuring conceptual soundness before execution.
6/9/2026
SRT: Super-Resolution for Time Series via Disentangled Rectified Flow
Disentangled rectified flow approach for time series super-resolution to reconstruct high-resolution signals from low-resolution inputs.
6/9/2026
LEAF: Growing Trees Without Branching for Speech-Aware Large Language Model Post-Training
LEAF reinforcement learning method for speech-aware LLM post-training using tree-based token credit assignment instead of coarse rewards.
6/9/2026
Structured Neuron Pruning in Deep Neural Networks Using Multi-Armed Bandits
Multi-armed bandit algorithm for structured neuron pruning in deep neural networks to improve efficiency and parameter reduction.
6/9/2026
Item Response Scaling Laws: A Measurement Theory Approach for Efficient and Generalizable Neural Scaling Estimation
Item Response Theory framework for estimating LLM scaling laws efficiently without expensive checkpoint evaluations across thousands of models.
6/9/2026
Query Lens: Interpreting Sparse Key-Value Features with Indirect Effects
Query Lens extends Logit Lens to interpret sparse autoencoder features by analyzing encoder and decoder-side key-value pairs to understand feature activation and output promotion.
6/9/2026
ScaleSweep: Accurate NVFP4 Post-Training Quantization of LLMs via Block Scale Initialization
ScaleSweep improves NVFP4 4-bit quantization of LLMs through efficient block scale optimization, reducing gap to optimal solutions.
6/9/2026
HASA: Subnet Allocation for Compute-Constrained Model-Heterogeneous Federated Learning
HASA allocates subnets in federated learning to handle heterogeneous device resources and data distributions while optimizing statistical performance.
6/9/2026
Finite Certificates for In-Context Determinacy and a Threshold Theory of Emergence in Language Models
Model-theoretic framework with finite semantic certificates for verifying context-conditioned LLM behavior and understanding emergence through row-space criteria.
6/9/2026
Sequential statistical inference for Large Language Models: Representation, validity, and monitoring
Sequential statistical inference framework for LLM trustworthiness, modeling dependent stochastic processes in deployment with behavioral monitoring.
6/9/2026
Large Language Models Should Learn Personalized Rather Than Aggregated Human Preferences
Position paper arguing LLMs should optimize for personalized individual preferences rather than aggregated preferences that represent no real user well.
6/9/2026
Active Learning with Foundation Model Priors: Efficient Learning under Class Imbalance
Active learning framework using foundation model priors to mitigate class imbalance and reduce annotation costs on imbalanced datasets.
6/9/2026
Trait-space Monitoring for Emergent Misalignment During Supervised Finetuning
Trait-space monitoring detects emergent misalignment during LLM supervised finetuning using linear directions in activation space without behavioral evaluation.
6/9/2026
Evaluation of ML Resource Utilization Requires Model Life Cycle Assessment
Framework for evaluating ML resource utilization and environmental impact across full model life cycle from development through deployment.
6/9/2026
KITE: A Tri-Modal Transformer Integrating Text, Images, and Knowledge Graphs for Fake News Detection
KITE integrates text, images, and knowledge graphs in a tri-modal transformer for fake news detection with improved semantic consistency analysis.
6/9/2026
DOG-DPO:Dynamic Optimization in Geometry for Safety Alignment
DOG-DPO optimizes safety alignment for LLMs through dynamic data selection that preserves directional preference information across multi-dataset settings.
6/9/2026
Semantic Cache Distillation: Efficient State Transfer via Reuse and Selective Patching
Semantic Cache Distillation optimizes LLM inference by reducing KV cache communication overhead and enabling cache reuse across model variants.
6/9/2026
Test-Time Adaptive Composition for Machine Learning as a Service (MLaaS) in IoT Environments
Test-time adaptive composition framework for ML-as-a-service in IoT environments handling heterogeneous client resources and data distributions.
6/9/2026
HARP: Efficient Data Selection for Finetuning Large Language Models
HARP presents efficient data selection for LLM finetuning, balancing scalability and downstream utility through train-free and train-based selection methods.
6/9/2026
Adversarial Robustness of Activation Steering in Large Language Models
First systematic evaluation of activation steering robustness in LLMs under adversarial text perturbations, testing four extraction methods and three attack strategies.
6/9/2026
EvoCSFL: Surrogate-Assisted Evolutionary Client Selection for Efficient and Robust Federated Learning
Proposes surrogate-assisted evolutionary algorithm for client selection in federated learning to improve convergence and robustness.
6/9/2026
How Much Dense Attention is Necessary? Oracle-Guided Sparse Prefill for Full/GQA Layers in Hybrid Long-Context Models
Research on optimizing dense attention in long-context LLMs using oracle-guided sparse prefill to reduce computational costs while preserving task performance.
6/9/2026
FunctionEvolve: Structure-Guided Symbolic Regression with LLMs
FunctionEvolve method combining LLM guidance with symbolic structure for symbolic regression, discovering scientific laws from data with domain-informed search.
6/9/2026
SAW: Stage-Aware Dynamic Weighting for Multi-Objective Reinforcement Learning in Large Language Models
SAW stage-aware dynamic weighting technique for multi-objective RL in LLMs, addressing asynchronous reward learning across objectives.
6/9/2026
WhiFlash: Accelerating Speculative Decoding with Token-Level Cross-Paradigm Routing
WhiFlash accelerates speculative decoding via token-level cross-paradigm routing, switching between autoregressive and diffusion drafting based on token type.
6/9/2026
Rosetta Memory: Adaptive Memory for Cross-LLM Agents
Rosetta Memory system providing adaptive memory for multi-LLM agents, enabling persistence and experience accumulation across different LLM backends.
6/9/2026
Cutting LLM Evaluation Costs with SySRs: A Bandit Algorithm that Provably Exploits Model Similarity
SySRs bandit algorithm reducing LLM evaluation costs by exploiting model similarity and adaptively allocating evaluation budget.
6/9/2026
Contrast encodes inductive bias: separating slow noise from dynamics in predictive representation learning
Study showing contrastive learning methods confuse slow noise with dynamics, proposing fixes for predictive representation learning frameworks like JEPA.
6/9/2026
A Framework for Evaluating and Benchmarking Concept Drift Detection Methods
Framework and benchmarking suite for evaluating concept drift detection methods in data stream mining with standardized metrics.
6/9/2026
Byzantine Cheap Talk: Adversarial Resilience and Topology Effects in LLM Coordination Games
Study of Byzantine adversarial resilience in multi-agent LLM coordination games, examining robustness of communication protocols under attacks.
6/9/2026
Teacher-Free Self-Training Amplifies but Does Not Compound: A Pass@$K$ Crossover on a Free-Verifier Domain
Teacher-free self-training study examining whether language models acquire new capabilities or improve expression of existing ones using exact verifiers.
6/9/2026
Instrumented data for causal scientific machine learning
Instrumented data approach for scientific machine learning combining observational data with mechanistic models and uncertainty quantification.
6/9/2026
Still: Amortized KV Cache Compaction in a Single Forward Pass
STILL method for KV cache compaction in language models during inference, addressing memory bottleneck for long-horizon deployment.
6/9/2026
Breaking the Bubble: Asynchronous Pipeline Parallel Training with Bounded Weight Inconsistency
Pipeline parallelism optimization technique (PACI) for training large neural networks with bounded weight inconsistency, reducing memory bubbles.
6/9/2026
Strained Coherence: A Pre-Failure Signal in Coding Agent Execution Trajectories
Research on strained coherence failure mode in LLM-based coding agents where agents acknowledge problems but proceed anyway, related to reward hacking.
6/9/2026
Partially Performative Prediction
Framework studying feedback loops when ML models deployed in real-world settings, examining endogenous distribution shifts.
6/9/2026
CAAL: Contextual Bandits based Online Hand-Craft Active Learning Strategy Selection
Contextual bandits approach for selecting active learning strategies dynamically without relying solely on labeled data feedback.
6/9/2026
The Easy, the Hard, and the Learnable: Confidence and Difficulty-Adaptive Policy Optimization for LLM Reasoning
RL method for improving LLM reasoning via verifiable rewards, analyzing confidence inflation and inefficient compute allocation in standard GRPO training.
6/9/2026
Minibatch Selection via Partition Matroid Constrained Gradient Matching
Cross-domain minibatch selection method using gradient matching and partition matroid constraints for training LLMs on heterogeneous data.
6/9/2026
Overcoming the Limits of Finite Difference Method; Physics-Informed Neural Network for Noisy High-Dimensional Heat Diffusion
Physics-informed neural network framework for solving high-dimensional heat diffusion under noisy boundary conditions.
6/9/2026
Enhancing AI Interpretability and Safety through Localised Architectures
Discussion of localized architectures for improving interpretability and safety of large language models and reasoning models.
6/9/2026
Evaluating the Impact of Task Granularity on Catastrophic Forgetting in Continual Learning
Study investigating task ordering effects on catastrophic forgetting in continual learning, comparing general-to-specific versus mixed learning orders.
6/9/2026
Semantic Quorum Assurance: Collective Certification for Non-Deterministic AI Infrastructure
Semantic reliability certification framework for multi-agent LLM systems in autonomous cloud operations, addressing operationally unsafe mutations.
6/9/2026
CausShield: Sample Reconstruction-Resilient Vertical FL via Causal Representation Learning
Privacy-preserving vertical federated learning using causal representation learning to defend against sample reconstruction attacks.
6/9/2026
SafeECGMatch: Calibration-Aware Joint Frequency and Time Space Semi-Supervised Learning for Open-Set ECG Classification
Semi-supervised learning approach for ECG classification handling label scarcity and out-of-distribution anomalies in clinical settings.
6/9/2026
When Behavioral Safety Evaluation Fails: A Representation-Level Perspective
Analysis of LLM safety evaluation gaps, formalizing audit gap between behavioral safety and representation-level robustness under intervention.
6/9/2026
Beyond Homophily: Towards Generalized Graph Reconstruction Attack and Defense
Study of graph reconstruction attacks on GNNs and defenses against model inversion attacks that reconstruct training adjacency from trained models.