5/29/2026
BioArc: Discovering Optimal Neural Architectures for Biological Foundation Models
Neural architecture search for biological foundation models considering physicochemical properties unique to biology.
5/29/2026
E-valuator: Reliable Agent Verifiers with Sequential Hypothesis Testing
Sequential hypothesis testing framework for reliable verification of AI agent trajectories and action quality scoring.
5/29/2026
When the Same Coefficients Reach Different Places: Asymmetric Realizability in Transplanting Tokenizers across Large Language Models
Analyzes asymmetric realizability problem when transplanting tokenizers across LLMs, studying coefficient reconstruction over shared lexical anchors.
5/29/2026
Bridging the Semantic Gap for Categorical Data Clustering via Large Language Models
Uses LLMs to recover semantic relationships in categorical data clustering by overcoming limitations of co-occurrence statistics.
5/29/2026
Mechanism Shift During Post-training from Autoregressive to Masked Diffusion Language Models
Circuit analysis comparing computational mechanisms of autoregressive models versus masked diffusion language models post-trained from the same base.
5/29/2026
NCSAM Noise-Compensated Sharpness-Aware Minimization for Noisy Label Learning
Noise-compensated sharpness-aware minimization optimizer for learning from noisy labels by connecting label noise to SAM flatness-seeking behavior.
5/29/2026
HE-SNR: Uncovering Latent Logic via Entropy for Guiding Mid-Training on SWE-bench
Proposes entropy-based metric (HE-SNR) for guiding mid-training of LLMs on software engineering tasks, addressing limitations of perplexity metrics.
5/29/2026
Representation Unlearning: Forgetting through Information Compression
Introduces Representation Unlearning framework for machine unlearning via information compression in representation space rather than parameter modification.
5/29/2026
Pushing the Limits of Block Rotations in Post-Training Quantization
Analyzes block Hadamard rotations in post-training quantization for neural networks, providing systematic non-asymptotic analysis of outlier suppression.
5/29/2026
Learn from A Rationalist: Distilling Intermediate Interpretable Rationales
Framework for extracting interpretable rationales from DNNs via knowledge distillation using select-predict architecture with remote supervision.
5/29/2026
Good SFT Optimizes for SFT, Better SFT Prepares for Reinforcement Learning
Shows that SFT optimization for downstream RL performance differs from isolated SFT performance, revealing checkpoint selection mismatch in reasoning LLM post-training.
5/29/2026
Rectified LpJEPA: Joint-Embedding Predictive Architectures with Sparse and Maximum-Entropy Representations
Proposes Rectified Distribution Matching Regularization for Joint-Embedding Predictive Architectures to learn sparse representations while preventing collapse.
5/29/2026
How Far Ahead Do LLMs Plan? Uncovering the Latent Horizon in Chain-of-Thought Reasoning
Research on how LLMs encode future reasoning in hidden states versus explicit chain-of-thought steps, exploring the latent planning horizon in multi-step reasoning.
5/29/2026
Achieving Linear Speedup for Composite Federated Learning
FedNMap method for composite federated learning with nonsmooth regularizers, achieving linear speedup across heterogeneous distributed systems.
5/29/2026
Optimization and Generation in Aerodynamics Inverse Design
Learning-based approach to aerodynamic inverse design balancing performance optimization with design language preservation.
5/29/2026
Discrete diffusion samplers and bridges: Off-policy algorithms and applications in latent spaces
Discrete diffusion sampling algorithms for off-policy sampling from unnormalized densities, with applications in latent space modeling.
5/29/2026
Rare Event Analysis of Large Language Models
Framework for systematic analysis of rare events and edge-case behaviors in large language models during deployment, addressing gaps between development and production.
5/29/2026
Gradient Preconditioning for Efficient and Reliable Reward-Guided Generation
Gradient preconditioning method for efficient reward-guided generation with one-step models, addressing reward hacking and speed issues in test-time optimization.
5/29/2026
Ask HN: About Claude Code's New Feature: Dynamic Workflows
Claude Code Dynamic Workflows feature enables resumable multi-day parallel tasks with preserved progress, potentially replacing manual context management.
5/29/2026
Gitea CVE-2026-27771 exposed private container images without authentication
Gitea security vulnerability CVE-2026-27771 allowed unauthenticated private container image access affecting 30k+ deployments for 4 years.
5/29/2026
The mysterious Hy3 LLM is topping OpenRouter Model Rankings by a large margin
Unknown LLM model 'Hy3' ranking high on OpenRouter leaderboard.
5/29/2026
A shared playbook for trustworthy third party evaluations
Recommendations for designing independent third-party evaluations of frontier AI model capabilities and safety mitigations.
5/28/2026
Building Trustworthy LLM Judges
Research on building reliable LLM-based judges for evaluation tasks.
5/28/2026
A disk-first C++ vector engine
Brinicle: C++ vector database engine with Python wrapper achieving sub-millisecond latency on 1.2M items, supporting hybrid semantic/lexical search.
5/28/2026
Show HN: Mira – Search for files semantically – no exact filenames required
Local-first desktop app using semantic search to find files by plain English queries instead of exact filenames, built with ML.
5/28/2026
Symposium – Replayable, multi-agent LLM deliberation as a protocol
Protocol for multi-agent LLM deliberation systems with replay capabilities for agent coordination.
5/28/2026
code2prompt: Convert any folder into a single LLM prompt
Tool to convert code folders into single LLM prompts for easier AI processing and context.
5/28/2026
Ask HN: Is Claude Opus 4.8 broken?
User reports Claude Opus 3.5 performance degradation with file reading errors and hallucinated commands in typical development workflows.
5/28/2026
Ask HN: Does Claude Code remove the need for so many front-end frameworks?
User discussion on whether Claude Code reduces need for frontend frameworks like React in building complex web applications.
5/28/2026
Show HN: Grove, an open-source MCP server to search/write to your Obsidian Vault
Open-source MCP server connecting AI clients to Obsidian vaults for semantic search and writing.
5/28/2026
Show HN: Rig – Local-first code graph for coding agents, in one npx command
Local-first semantic code graph for AI coding agents via MCP protocol. Runs locally without external APIs.
5/28/2026
Show HN: Open-source browser agent that runs 24/7
Open-source AI browser agent that automates repetitive web tasks 24/7 by learning from user actions.
5/28/2026
Bipolar Tracker – LLM journal that detects bipolar escalation from natural text
LLM-based journal application that analyzes natural language entries to detect bipolar mood escalation patterns.
5/28/2026
Various LLM Smells
Analysis of LLM-generated text patterns showing identical sentence structures spreading across internet, identifying systemic AI writing artifacts.
5/28/2026
Show HN: Bootstrap a team of coding agents from a template, OSS
Open-source tool (aw team bootstrap) for structuring multi-agent teams with global IDs, communication, roles, and automation from templates.
5/28/2026
Show HN: Enough, an Agent Harness for Writers
Enough is extensible personal language system with local models and OpenRouter support for planning, writing, and translation tasks.
5/28/2026
Show HN: Ktx – Open-source executable context layer for data agents
Ktx is open-source executable context layer for reliable data agents, addressing accuracy issues in SQL generation and warehouse interactions.
5/28/2026
Show HN: My custom agent skill for commit messages
Custom Claude Code agent skill for enforcing conventional commit message format with validation hooks.
5/28/2026
Giving LLMs a Formal Reasoning Engine for Code Analysis
Research on enhancing LLM code analysis with formal reasoning engine for transitive code questions and dead code detection.
5/28/2026
Why Google's AI can't spell Google (or anything else)
Analysis of failures in Google's AI Overview demonstrating character counting and spelling errors in simple tasks.
5/28/2026
The enterprise security risks of autonomous AI agents, and how to manage them
Analysis of enterprise security risks from autonomous AI agents with 66% benchmark accuracy increase, discusses control mechanisms.
5/28/2026
Show HN: Genesis Architect – scans 20 GitHub repos before scaffolding
Genesis Architect scans 15-20 GitHub repos to extract architecture patterns, bug issues, and security patches, then generates project scaffolds based on production data rather than templates.
5/28/2026
Creating got easier. Meaning got harder
Opinion piece on how AI-generated content proliferation made creation easier but quality assessment harder across media formats.
5/28/2026
Malware dev tries to steal Claude users secrets NPM slop, leaks own GitHub token
Analysis of AI-generated malware npm package targeting Claude users that leaked its own GitHub token via security misconfiguration.
5/28/2026
Show HN: Open-source alternative to Codex Chronicle, using Apple's local OCR
Familiar: open-source screen/clipboard capture tool using Apple's local OCR, providing context for AI agents via markdown output.
5/28/2026
Never Pay for Claude
rotom: local Rust gateway enabling Codex and Grok OAuth with OpenAI/Anthropic-compatible APIs, allowing use of alternative providers.
5/28/2026
Getting Agent-Ready with Symfony
Guide for making Symfony applications agent-ready with markdown negotiation, OpenAPI docs, and API exposure for AI agent integration.
5/28/2026
Show HN: Code Block Selector – select code by structure, not by dragging
VS Code extension for selecting code blocks by syntax structure rather than manual dragging, with AST-aware navigation.
5/28/2026
Claude Code Skills set for SEO research
SEO automation toolkit using NodeHub API and optional LLM for SERP analysis and cluster naming. Promotional package by Senuto.
5/28/2026
How Endava builds an agentic organization with Codex
Endava uses Codex to accelerate requirements analysis and software delivery, reducing analysis time from weeks to hours.