Ax Zheng Zhang, Cuong C. Nguyen, David Rosewarne, Kevin Wells, Gustavo Carneiro 4/7/2026

Fatigue-Aware Learning to Defer via Constrained Optimisation

arXiv paper on human-AI cooperation via fatigue-aware deferral systems. ML method modeling human fatigue to optimize when AI should defer to humans.

Ax Ken M. Nakanishi 4/7/2026

Screening Is Enough

arXiv paper introducing Multiscreen attention mechanism for language models. Alternative to softmax attention enabling absolute relevance scoring in transformers.

Ax Rafael Pardinas, Ehsan Kamalloo, David Vazquez, Alexandre Drouin 4/7/2026

Apriel-1.5-OpenReasoner: RL Post-Training for General-Purpose and Efficient Reasoning

arXiv paper on reinforcement learning post-training for reasoning models. Open-weight model training using verifiable rewards across diverse reasoning domains.

Ax Zun Li, Marc Lanctot, Kevin R. McKee, Luke Marris, Ian Gemp, Daniel Hennes, Paul Muller, Kate Larson, Yoram Bachrach, Michael P. Wellman 4/7/2026

Combining Tree-Search, Generative Models, and Nash Bargaining Concepts in Game-Theoretic Reinforcement Learning

arXiv paper on opponent modeling in game-theoretic reinforcement learning using tree-search and generative models. Research on scalable multi-agent RL methods.

Ax Dinesh Sharma, Ankit Shah, Chaitra Gopalappa 4/7/2026

A Multi-Agent Reinforcement Learning Framework for Public Health Decision Analysis

Multi-agent reinforcement learning framework for HIV prevention policy optimization across U.S. regions.

Ax Changdae Oh, Gyeongdeok Seo, Geunyoung Jung, Zhi-Qi Cheng, Hosik Choi, Jiyoung Jung, Kyungwoo Song 4/7/2026

Robust Adaptation of Foundation Models with Black-Box Visual Prompting

Black-box visual prompting method for parameter-efficient transfer learning of foundation models without full parameter access.

Ax Lechen Zhang, Tolga Ergen, Lajanugen Logeswaran, Moontae Lee, David Jurgens 4/7/2026

SPRIG: Improving Large Language Model Performance by System Prompt Optimization

SPRIG: Genetic algorithm for optimizing system prompts in LLMs to improve task performance.

Ax Ugur Dar, Mustafa Cavus 4/7/2026

From XAI to MLOps: Explainable Concept Drift Detection with Profile Drift Detection

XAI-based method combining explainability with concept drift detection for monitoring model performance degradation.

Ax Qi Wang, Zhipeng Zhang, Baao Xie, Xin Jin, Yunbo Wang, Shiyu Wang, Liaomo Zheng, Xiaokang Yang, Wenjun Zeng 4/7/2026

Disentangled World Models: Learning to Transfer Semantic Knowledge from Distracting Videos for Reinforcement Learning

World models using disentangled representations to transfer semantic knowledge from distracting videos for RL agents.

Ax Yifu Yuan, Haiqin Cui, Yibin Chen, Zibin Dong, Fei Ni, Longxin Kou, Jinyi Liu, Pengyi Li, Yan Zheng, Jianye Hao 4/7/2026

From Seeing to Doing: Bridging Reasoning and Decision for Robotic Manipulation

FSD framework combining vision-language models with robotic action models for zero-shot manipulation in novel scenarios.

Ax Edward L. Wang, Mohammad Sharifi Kiasari, Tianyu Wang, Hayden Helm, Avanti Athreya, Carey Priebe, Vince Lyzinski 4/7/2026

Gaussian mixture models as a proxy for interacting language models

Gaussian mixture models as computationally efficient proxy for LLM+RAG systems combining multiple models.

Ax Nikolay Blagoev, O\u{g}uzhan Ersoy, Lydia Yiyu Chen 4/7/2026

All is Not Lost: LLM Recovery without Checkpoints

Technique to recover LLM training on decentralized/spot nodes from partial model loss without full checkpoints.

Ax Yukun Huang, Sanxing Chen, Jian Pei, Manzil Zaheer, Bhuwan Dhingra 4/7/2026

Cite Pretrain: Retrieval-Free Knowledge Attribution for Large Language Models

Method for LLMs to reliably cite source documents seen during training without external retrievers at inference time.

Ax Tomer Krichli, Bhiksha Raj, Joseph Keshet 4/7/2026

WhisperRT -- Turning Whisper into a Causal Streaming Model

Modification of Whisper ASR model to enable low-latency streaming transcription through architectural and training changes.

Ax Yifu Yuan, Haiqin Cui, Yaoting Huang, Yibin Chen, Fei Ni, Zibin Dong, Pengyi Li, Yan Zheng, Hongyao Tang, Jianye Hao 4/7/2026

Embodied-R1: Reinforced Embodied Reasoning for General Robotic Manipulation

Vision-language model for robotic manipulation using embodiment-agnostic pointing representation to address generalization in embodied AI.

Ax Wangsong Yin, Daliang Xu, Mengwei Xu, Gang Huang, Xuanzhe Liu 4/7/2026

ShadowNPU: System and Algorithm Co-design for NPU-Centric On-Device LLM Inference

System co-design for efficient on-device LLM inference on NPU hardware, optimizing attention operations for privacy-preserving deployment.

Ax Vincent Nguyen, Guilherme Herzog, Jos\'e Cambronero, Marcus Revaj, Aditya Kini, Alexander Fr\"ommgen, Maxim Tabachnyk 4/7/2026

Smart Paste: Automatically Fixing Copy/Paste for Google Developers

Google developer tool using deep learning to automatically fix copy/paste code, predicting required edits from formatting to cross-language translation.

Ax Sicheng Lyu, Yu Gu, Xinyu Wang, Jerry Huang, Sitao Luan, Yufei Cui, Xiao-Wen Chang, Peng Lu 4/7/2026

EvoEdit: Evolving Null-space Alignment for Robust and Efficient Knowledge Editing

Knowledge editing method for LLMs enabling sequential updates through null-space alignment, improving robustness in continual model editing scenarios.

Ax Pankaj K Mishra, Sanni Laaksonen, Jochen Kamm, Anand Singh 4/7/2026

Three-dimensional inversion of gravity data using implicit neural representations and scientific machine learning

Scientific machine learning approach using implicit neural representations for 3D gravity inversion, modeling subsurface density as continuous field.

Ax Ilyass Moummad, Kawtar Zaher, Herv\'e Go\"eau, Alexis Joly 4/7/2026

Image Hashing via Cross-View Code Alignment in the Age of Foundation Models

Image hashing method using foundation models for efficient large-scale retrieval with compact binary codes instead of high-dimensional embeddings.

Ax Saeedeh Javadi, Sara Mirabi, Manan Gangar, Bahadorreza Ofoghi 4/7/2026

Contradictions in Context: Challenges for Retrieval-Augmented Generation in Healthcare

Study of RAG limitations in healthcare domain, showing how retrieval-augmented generation fails when source documents contain contradictory or outdated information.

Ax Md Tanvirul Alam, Saksham Aggarwal, Justin Yang Chae, Nidhi Rastogi 4/7/2026

SPHINX: A Synthetic Environment for Visual Perception and Reasoning

Sphinx synthetic environment for visual perception and reasoning with procedurally generated puzzles covering 25 task types, enabling precise evaluation and large-scale dataset construction.

Ax Ivan Yee Lee, Cheng Yang, Taylor Berg-Kirkpatrick 4/7/2026

Optical Context Compression Is Just (Bad) Autoencoding

Analysis showing optical context compression via vision tokens is functionally equivalent to lossy autoencoding, questioning effectiveness of DeepSeek-OCR's compression pipeline.

Ax Leijie Wang, Otilia Stretcu, Wei Qiao, Thomas Denby, Krishnamurthy Viswanathan, Enming Luo, Chun-Ta Lu, Tushar Dogra, Ranjay Krishna, Ariel Fuxman 4/7/2026

Agile Deliberation: Concept Deliberation for Subjective Visual Classification

Human-in-the-loop approach for visual classification through iterative concept deliberation, addressing subjective vision tasks in content moderation and curation.

Ax Rahul Baxi 4/7/2026

The Drill-Down and Fabricate Test (DDFT): A Protocol for Measuring Epistemic Robustness in Language Models

DDFT protocol measures epistemic robustness of language models under stress conditions, distinguishing knowledge gaps from verification mechanism failures beyond static benchmarks.

Ax Binxu Wang, Jingxuan Fan, Xu Pan 4/7/2026

Circuit Mechanisms for Spatial Relation Generation in Diffusion Transformers

Mechanistic interpretability study of how Diffusion Transformers generate correct spatial relations in text-to-image generation.

Ax Mayank Sharma, Roy Pea, Hari Subramonyam 4/7/2026

ConvoLearn: A Dataset for Fine-Tuning Dialogic AI Tutors

ConvoLearn dataset of 2,134 tutor-student dialogues for fine-tuning dialogue-based AI tutors grounded in knowledge-building theory.

Ax Ellen Xiaoqing Tan, Jack Lanchantin, Shehzaad Dhuliawala, Danwei Li, Thao Nguyen, Jing Xu, Ping Yu, Ilia Kulikov, Sainbayar Sukhbaatar, Jason Weston, Xian Li, Olga Golovneva 4/7/2026

Self-Improving Pretraining: using post-trained models to pretrain better models

Pretraining approach using post-trained models to incorporate reasoning and safety behaviors earlier in LLM development.

Ax Zixuan Xia, Hao Wang, Pengcheng Weng, Yanyu Qian, Yangxin Xu, William Dan, Fei Wang 4/7/2026

Improving Multimodal Learning with Dispersive and Anchoring Regularization

Regularization techniques for improving multimodal representation learning by addressing collapse and inconsistency issues.

Ax Henri A\"idasso, Francis Bordeleau, Ali Tizghadam 4/7/2026

Predicting Intermittent Job Failure Categories for Diagnosis Using Few-Shot Fine-Tuned Language Models

Few-shot fine-tuned language models for diagnosing intermittent CI pipeline failures in software development.

Ax Qian Wang, Xuandong Zhao, Zirui Zhang, Zhanzhi Lou, Nuo Chen, Dawn Song, Bingsheng He 4/7/2026

Making Bias Non-Predictive: Training Robust LLM Reasoning via Reinforcement Learning

Training LLMs to resist cognitive biases in reasoning via reinforcement learning rather than prompting.

Ax William Lugoloobi, Thomas Foster, William Bankes, Chris Russell 4/7/2026

LLMs Encode Their Failures: Predicting Success from Pre-Generation Activations

Research on predicting LLM success from internal pre-generation activations to optimize inference efficiency in reasoning tasks.

Ax Bowen Liu, Zhi Wu, Runquan Xie, Zhanhui Kang, Jia Li 4/7/2026

Scaling the Scaling Logic: Agentic Meta-Synthesis of Logic Reasoning

SSLogic agentic meta-synthesis framework where LLM agents iteratively generate and refine task specifications for logic reasoning.

Ax Camile Lendering, Erkut Akdag, Egor Bondarev 4/7/2026

SubspaceAD: Training-Free Few-Shot Anomaly Detection via Subspace Modeling

Training-free few-shot anomaly detection using subspace modeling of vision foundation model features.

Ax Abdul Rahman 4/7/2026

CSTS: A Canonical Security Telemetry Substrate for AI-Native Cyber Detection

Canonical Security Telemetry Substrate for standardizing cybersecurity data formats for AI-driven detection.

Ax Haiyue Song, Masao Utiyama 4/7/2026

OptiMer: Optimal Distribution Vector Merging Is Better than Data Mixing for Continual Pre-Training

OptiMer framework for optimizing data mixture ratios during continual LLM pre-training without manual tuning.

Ax Jack Young 4/7/2026

S0 Tuning: Zero-Overhead Adaptation of Hybrid Recurrent-Attention Models

S0 tuning method for efficient LLM adaptation via state matrix optimization, outperforming LoRA on code generation tasks.

HN gpi 4/7/2026

Claude Is Not Your Architect. Stop Letting It Pretend

Opinion piece on AI agent limitations in architecture decisions. Warns of overreliance on Claude for system design.

HN raviishgupta 4/7/2026

Agentic AI and Occupational Displacement: Multi-Regional Task Exposure Analysis

arXiv research on agentic AI and occupational displacement analyzing task exposure across regions.

HN ddalcu 4/7/2026

MLX-Serve a Native LLM Runtime for Apple Silicon

MLX-Serve: Native Zig inference server for Apple Silicon with OpenAI-compatible API. No Python. Includes 7 built-in tools, streaming, embeddings.

HN muthuishere 4/7/2026

Show HN: Portable offline LLM knowledge system that runs in browser

Browser-based offline LLM knowledge system. Bundles model and embeddings for reproducible, internet-free deployment.

HN opuslabs 4/7/2026

I'm building a local-first CLI "knowledge foundry" for LLM workflows

CLI knowledge system for local-first LLM workflows. Minimal details in title-only post.

HN skmadd 4/7/2026

Anthropic's refusal to drop AI safeguards for The Pentagon

Anthropic refused Pentagon demands to remove AI safety guardrails from Claude for military use, resulting in federal supply chain designation.

HN lostmsu 4/7/2026

Codexia: OSS Agent Workstation for Codex and Claude

Codexia: Tauri app combining Codex CLI and Claude Code with IDE editor, web server, and agent workflows. AGPL-3.0 licensed.

HN jnord 4/7/2026

From folding boxes to fixing vacuums, GEN-1 robotics model hits 99% reliability

Generalist's GEN-1 robotics model achieves 99% reliability on physical tasks and can improvise solutions not in training data.

HN menggg 4/7/2026

Show HN: I built my knowledge OS from ingesting to next step action

Personal knowledge OS powered by LLMs. Ingests knowledge, synthesizes patterns, plans actions. Works with Claude Code or any LLM CLI.

HN justnau1020 4/7/2026

Gave my Claude a subconscious memory system

Memory/knowledge system for Claude Code with hooks and persistent storage. Production-tested framework from hundreds of hours of use.

HN rguiscard 4/7/2026

Benchmarked 37 LLMs on MacBook Air M5 32GB

Benchmark of 37 LLMs on MacBook Air M5 with open-source tool. Tests performance across model families using llama-bench.

HN makeavish 4/7/2026

Claude Agent SDK vs. Messages API for a Production AI Assistant?

Developer discusses tradeoffs between Anthropic's Agent SDK and Messages API for production B2B assistant with tool calling and streaming.

HN jaafarskafi 4/7/2026

Show HN: MCP plugin that lets Claude autonomously pay for APIs via Lightning

BoltClaw MCP plugin enables Claude agents to autonomously pay for APIs via Lightning Network. Solves agent payment bottleneck.