6/30/2026
Show HN: SlimSnap – mark a screenshot element, get JSON for your coding agent
SlimSnap: Convert screenshot annotations to JSON for coding agents (Claude, Aider, Codex). Enables agents to process UI visuals.
6/30/2026
TurboPrefill: 2.7× faster than llama.cpp Pipeline Parallel on Llama-3-70B
TurboPrefill: 2.7× faster LLM inference on Llama-3-70B via optimized pipeline parallelization and prefill scheduling mechanisms.
6/30/2026
Show HN: I built an agent that uses email as a file system
Supafax: Agent using email as file system for memory and configuration. Email-native interface for agent task execution and persistence.
6/30/2026
Ask HN: ASICs made GPU crypto mining obsolete. Will ASICs for AI be a thing?
Discussion on whether ASICs will emerge for AI training/inference similar to crypto mining.
6/30/2026
A GitHub-compatible Git service built for AI agents
Self-hosted GitHub-compatible Git service designed as agent-first with durable identities, scoped tokens, and AI agent as first-class citizens.
6/30/2026
AI Tools Accelerates Coding, but Not Overall Software Delivery
GitLab research shows AI coding tools improve speed but don't accelerate overall delivery due to testing/review bottlenecks and governance challenges.
6/30/2026
AI agents are not your "coworkers"
Analysis of risks in marketing AI agents as coworkers, focusing on human error detection and accountability concerns.
6/30/2026
Analysis of Adam Algorithms for Stochastic Dynamic Systems
Theoretical analysis of Adam optimizer for time-varying and nonstationary optimization problems.
6/30/2026
An Integrated Machine Learning and Hierarchical Variance Decomposition Pipeline for Student Performance Prediction and Metacognitive Calibration on Multi-Signal Telemetry
ML pipeline (UBP-CAP) for student performance prediction and metacognitive calibration in intelligent tutoring systems.
6/30/2026
Multi-Agent Routing as Set-Valued Prediction: A WildChat Benchmark and Cost-Aware Evaluation
Introduces WildChat benchmark for multi-agent routing as set-valued prediction problem with 3000 prompts and cost-aware evaluation.
6/30/2026
DLR: Zero-Inference-Cost Latent Residuals for Low-Rank Pre-Training
Proposes DLR, a training-only plug-in for low-rank LLM pre-training that reduces parameters and FLOPs without inference cost.
6/30/2026
ReGuide: From Test-Time Guidance to Self-Improving Diffusion Policies
Introduces ReGuide for improving behavior-cloned diffusion policies through test-time guidance and self-improvement mechanisms.
6/30/2026
Modification-Considering Value Learning for Reward Hacking Mitigation in RL
Proposes Modification-Considering Value Learning to mitigate reward hacking in RL by detecting task violations during policy learning.
6/30/2026
When Can Conformal Risk Control Certify LLM Outputs? Bounds, Impossibility, and Adaptation for Structured Generation
Characterizes when conformal risk control can certify LLM outputs for structured generation tasks and proposes adaptation methods.
6/30/2026
Statistically Indistinguishable, Operationally Distinct: A Formal Barrier for Tabular Foundation Models
Shows tabular foundation models cannot reason about system-generated data without access to underlying rules, introducing Operational Turing Test.
6/30/2026
DiLaServe: High SLO Attainment Serving for Diffusion Language Models
Proposes DiLaServe system for efficient serving of diffusion language models while meeting latency SLOs through parallel token generation.
6/30/2026
How Token Influence Decays with Distance: A Green-Function View of Trained Language Models
Studies how token influence decays with distance in transformer LMs using green-function analysis from differential equations, inspired by operator learning.
6/30/2026
On the Nonlinearity of Learning Rate Scaling for LLM Training
Empirical study examining non-linear learning rate scaling laws for LLM training across model sizes 22M-707M parameters.
6/30/2026
Invariant Reasoning Directions in Latent Trajectories of Language Models
Study of latent reasoning trajectories in language models, showing low-rank structure in contrastive refinement signals between reasoning paths.
6/30/2026
Symbolic Mechanistic Data Attribution: Tracing Training Influence to Learned Behavioral Policies
Framework for attributing training data to interpretable symbolic policies in language models using Ridge regression over mechanistic circuits.
6/30/2026
BaRA: Bayesian Adaptive Rank Allocation for Parameter-Efficient Fine-Tuning
BaRA: Bayesian adaptive rank allocation for LoRA fine-tuning; dynamically optimizes rank per layer for uncertainty calibration.
6/30/2026
Representational Depth of Evaluation Awareness Shifts With Scale in Open-Weight Language Models
Analysis of evaluation-awareness in language models across scales; shows systematic shift in layer depth of test-detection representation.
6/30/2026
Multi-Block Diffusion Language Models
Multi-Block Diffusion extends block diffusion language models with concurrent multi-block decoding and inter-block parallelism.
6/30/2026
Depth Exploration for LLM Decoding
Depth Exploration for LLM decoding: exploits varying token predictability depths to accelerate inference beyond single-exit methods.
6/30/2026
On the Policy Gradient Foundations of Group Relative Policy Optimization: Credit Assignment, Gradient Sparsity, and Rank Collapse
Analysis of Group Relative Policy Optimization revealing credit assignment failure and gradient sparsity from policy gradient first principles.
6/30/2026
Towards Evaluating Data Priors for Tabular Foundation Models
Methodological framework for evaluating data-generating priors as independent components in tabular foundation models.
6/30/2026
KrishokChat: A Citation-Grounded Dataset and Benchmark for Bengali Agricultural Advisory
KrishokChat: citation-grounded Bengali agricultural instruction-tuning dataset with 290 knowledge nodes for crop advisory in low-resource settings.
6/30/2026
Adaptive Block Diffusion: Resolving Training-Inference Mismatch in Diffusion Language Models
Adaptive Block Diffusion resolves training-inference mismatch in diffusion language models by optimizing over variable context configurations.
6/30/2026
Deterministic Decisions for High-Stakes AI. A Zero-Egress Pipeline with the Deployability of RAG and the Accuracy of Machine Learning
Study of intervention bias in LLM educational advisory agents using RAG pipeline; demonstrates models recommend action when inaction optimal.
6/30/2026
SP-CACW: Convergence-Aware Client Weighting for Selfish Personalized Learning
Federated learning method (SP-CACW) for personalized client objectives with convergence-aware weighting to avoid negative transfer.
6/30/2026
Sample Complexity of Scientific Discovery: PAC Learnability of Compositional Function Trees
PAC learning analysis of sample complexity for symbolic regression via compositional function trees.
6/30/2026
Randomized neural operator for parametric PDEs with fast training and conformal uncertainty quantification
Randomized neural operator combining PCA and fixed random features for fast training of parametric PDE solvers.
6/30/2026
Interpretable Inverse Design of Metal-Organic Frameworks with Large Language Model Agents
LLM4MOF framework where language model agents autonomously design metal-organic frameworks through closed-loop reasoning and simulation.
6/30/2026
Structured Proper Loss Geometries for Multiclass Classification: Theory and Controlled Empirical Evaluation
Theoretical and empirical analysis of proper loss geometries and curvature effects in multiclass classification objectives.
6/30/2026
CRAFT: Counterfactual Credit Assignment from Free Sibling Rollouts for Self-Distilled Agentic Reinforcement Learning
Self-distilled agentic RL method using counterfactual credit assignment from sibling rollouts to improve token-level learning.
6/30/2026
Reported Confidence in LLMs Tracks Commitment More Than Correctness
Study showing LLM confidence reports correlate more with commitment than correctness, questioning their use as uncertainty measures.
6/30/2026
Reinforcement Learning in Super Mario Bros: Curriculum, Pedagogy, and Optimal Level Design in World 1-1
Empirical evaluation comparing Q-Learning, SARSA, Monte Carlo, and DQN algorithms on Super Mario Bros World 1-1.
6/30/2026
Anti-Collapse Dynamics and the Emergence of Multi-Time-Scale Learning in Recurrent Neural Networks
Theoretical analysis of why recurrent networks struggle with long-range dependencies and how multi-time-scale learning emerges.
6/30/2026
Not All Objectives Are Born Equal: Priority-Constrained Descent for Hierarchical Multi-Objective Optimization
Priority-Constrained Descent algorithm for multi-objective optimization with inherent hierarchies in deep learning.
6/30/2026
Do Models Read What They Write? Causal Registers in Scratchpad Reasoning
Experimental analysis testing whether model scratchpad outputs causally influence later computation in chain-of-thought reasoning.
6/30/2026
The Mirage of Optimizing Training Policies: Monotonic Inference Policies as the Real Objective for LLM Reinforcement Learning
Study of training-inference mismatch in LLM reinforcement learning and proposal for monotonic inference policies to improve stability.
6/30/2026
Optimizer Memory Makes Shuffle Order a First-Order Source of Fine-Tuning Noise
Analysis showing optimizer memory makes shuffle order a significant source of fine-tuning noise in model training.
6/30/2026
Speculative Pre-Positioning: Decoding Stateful Sessions to the Next Decision Point Off the Critical Path
Optimization technique for inference servers to decode stateful sessions off the critical path without draft models.
6/30/2026
The Joint Effect of Quantization and Sampling Temperature on LLM Safety Alignment: A Factorial Analysis
Factorial analysis of how quantization and sampling temperature jointly affect LLM safety alignment during deployment.
6/30/2026
Bilevel Optimization for Neural Architecture Search
Bilevel optimization framework for Neural Architecture Search: hierarchical approach for NAS, hyperparameter tuning, meta-learning with arXiv research.
6/30/2026
STEMGym: Benchmarking Sequential Decision-Making under Dose Budgets in Autonomous Electron Microscopy
STEMGym: open-source benchmark for sequential decision-making in autonomous electron microscopy under dose constraints using RL/Bayesian approaches.
6/30/2026
How AI settled the complexity of the oldest SGD algorithm
AI models helped determine worst-case complexity of the 1937 Kaczmarz algorithm, foundational to modern SGD used in training ChatGPT and Gemini.
6/30/2026
Mechanistically Eliciting Latent Behaviors in Language Models
Causal Perturbative Elicitation discovers interpretable low-rank adapters (LoRAs) that surface latent behavioral modes in LLMs for risk evaluation and behavior reshaping.
6/30/2026
Does Role Specialization Matter for Explanation Faithfulness in Mixture-of-Experts?
Investigates whether role-based expert specialization in Mixture-of-Experts architectures preserves explanation faithfulness and interpretability.
6/30/2026
Fuzzing Large Language Models to Elicit Hidden Behaviours
Fuzzing LLM weights and activations with Gaussian noise to systematically elicit hidden backdoored behaviors in 7B-13B models without knowing triggers.