4/21/2026
Lil: Less is Less When Applying Post-Training Sparse-Attention Algorithms in Long-Decode Stage
Lil: Study showing sparse-attention algorithms underperform in LLM decode stage despite improvements in prefill stage.
4/21/2026
Do LLMs Encode Functional Importance of Reasoning Tokens?
Analysis of whether LLMs internally encode token-level functional importance in reasoning chains for compact reasoning generation.
4/21/2026
MMErroR: A Benchmark for Erroneous Reasoning in Vision-Language Models
MMErroR: 1997-sample benchmark evaluating whether Vision-Language Models detect erroneous reasoning across 24 subdomains.
4/21/2026
From Domains to Instances: Dual-Granularity Data Synthesis for LLM Unlearning
BiForget framework for automated synthesis of high-quality forget sets addressing domain and instance-level unlearning granularities in LLMs.
4/21/2026
Tape: A Cellular Automata Benchmark for Evaluating Rule-Shift Generalization in Reinforcement Learning
Tape: Cellular automata benchmark isolating rule-shift generalization in RL with controlled protocol and continuous metrics.
4/21/2026
Compositional Steering of Large Language Models with Steering Tokens
Compositional steering tokens enable multi-behavior control of LLMs simultaneously, addressing underexplored problem of steering toward multiple objectives.
4/21/2026
GanitLLM: Difficulty-Aware Bengali Mathematical Reasoning through Curriculum-GRPO
GanitLLM: Bengali mathematical reasoning model with difficulty-aware corpus and curriculum-based GRPO training pipeline for multi-step math problems.
4/21/2026
Attention-space Contrastive Guidance for Efficient Hallucination Mitigation in LVLMs
Training-free method to mitigate hallucinations in vision-language models by using attention-space contrastive guidance for visually grounded generation.
4/21/2026
Who Gets Which Message? Auditing Demographic Bias in LLM-Generated Targeted Text
Systematic analysis of demographic bias in LLM-generated targeted messaging across GPT-4o, Llama-3.3, and Mistral models with evaluation framework.
4/21/2026
CaseFacts: A Benchmark for Legal Fact-Checking and Precedent Retrieval
Benchmark for legal fact-checking: evaluates systems on verifying layperson legal claims against U.S. Supreme Court precedents.
4/21/2026
GeoRC: A Benchmark for Geolocation Reasoning Chains
Benchmark for evaluating vision-language model reasoning: assesses whether VLMs can explain geolocation predictions with supporting image evidence.
4/21/2026
SIGMA: A Semantic-Grounded Instruction-Driven Generative Multi-Task Recommender at AliExpress
SIGMA multi-task recommender system using LLMs for semantic-grounded instruction-driven generative recommendations in e-commerce.
4/21/2026
ReflexiCoder: Teaching Large Language Models to Self-Reflect on Generated Code and Self-Correct It via Reinforcement Learning
ReflexiCoder teaches LLMs to self-reflect and self-correct generated code via reinforcement learning without external oracles.
4/21/2026
Decidable By Construction: Design-Time Verification for Trustworthy AI
Design-time verification framework for trustworthy AI enabling pre-training verification of model correctness and stability.
4/21/2026
LogicDiff: Logic-Guided Denoising Improves Zero-Shot Reasoning in Masked Diffusion Language Models
LogicDiff improves zero-shot reasoning in masked diffusion language models via logic-guided token unmasking at inference time.
4/21/2026
SkillX: Automatically Constructing Skill Knowledge Bases for Agents
SkillX automatically constructs plug-and-play skill knowledge bases for LLM agents to improve learning efficiency and generalization.
4/21/2026
ConsistRM: Improving Generative Reward Models via Consistency-Aware Self-Training
ConsistRM improves generative reward models via consistency-aware self-training to address human preference alignment at scale.
4/21/2026
Reasoning-Based Refinement of Unsupervised Text Clusters with LLMs
LLM-based framework for validating and restructuring unsupervised text clusters using semantic reasoning and refinement.
4/21/2026
SEARL: Joint Optimization of Policy and Tool Graph Memory for Self-Evolving Agents
Self-evolving agents framework jointly optimizing policy and tool graph memory for autonomous learning and tool synthesis.
4/21/2026
From Answers to Arguments: Toward Trustworthy Clinical Diagnostic Reasoning with Toulmin-Guided Curriculum Goal-Conditioned Learning
LLMs for clinical decision support with transparent reasoning using Toulmin argumentation and curriculum learning to improve trustworthiness.
4/21/2026
ProbeLogits: Kernel-Level LLM Inference Primitives for AI-Native Operating Systems
ProbeLogits presents kernel-level LLM inference operations for AI-native OS, enabling logit-based classification of agent actions without learned parameters.
4/21/2026
VoxSafeBench: Not Just What Is Said, but Who, How, and Where
VoxSafeBench introduces a benchmark for evaluating safety in speech language models within multi-user environments, considering speaker identity and environmental context.
4/21/2026
VeriGraphi: A Multi-Agent Framework of Hierarchical RTL Generation for Large Hardware Designs
VeriGraphi proposes a multi-agent LLM framework for generating hierarchical Verilog hardware designs, addressing context loss and hallucination issues in RTL generation.
4/21/2026
Differentially Private Conformal Prediction
Differentially private conformal prediction method for uncertainty quantification with statistical efficiency under privacy constraints.
4/21/2026
World-Value-Action Model: Implicit Planning for Vision-Language-Action Systems
World-Value-Action model for vision-language-action agents that enables implicit planning and long-horizon reasoning beyond direct action prediction.
4/21/2026
Scaling Codex to enterprises worldwide
OpenAI expands Codex to enterprises through partnerships with GSIs, reaching 4M weekly developers. Codex deployment across software development lifecycle workflows.
4/20/2026
Show HN: Sprag – full-stack Python framework with no JavaScript Required
Full-stack Python framework compiling Python to JavaScript for browser execution with unified state management.
4/20/2026
Show HN: Auto-generated titles and colors for parallel Claude Code sessions
Plugin for Claude Code that auto-generates session titles and colors to distinguish parallel LLM sessions.
4/20/2026
Even 'uncensored' models can't say what they want
Research on safety-filtered LLM pretrains showing probabilistic censoring gaps across models from multiple labs.
4/20/2026
Show HN: Dynamic Hybrid Search That Beats Pure Dense and Fixed Hybrid
Managed vector index service with dynamic hybrid search combining dense embeddings and BM25, achieving 12× compression and 10× faster queries with per-query fusion.
4/20/2026
Agentic framework that self-improves its stock portfolio strategy (GitHub).)
Open-source agentic framework for quantitative finance with self-improving portfolio strategies and backtesting.
4/20/2026
AI and the Neurodivergent Unlock
Study examining how AI tools like Cursor enable improved workflow and focus for neurodivergent developers.
4/20/2026
DevArch 2.0
DevArch 2.0 provides directives and agents for Claude Code enforcing engineering discipline with automated quality gates and testing.
4/20/2026
Show HN: MCPfinder – An MCP server that finds and installs other MCP servers
MCPfinder aggregates and helps discover/install Model Context Protocol servers from multiple registries for agent configuration.
4/20/2026
What we learned using AI agents to refactor a monolith
1Password shares experience using AI agents to refactor a multi-million-line Go monolith, covering successes and failures.
4/20/2026
AI quota inflation is no token effort. It's baked in
Opinion on token inflation in LLM pricing and feature bloat driven by AI capabilities.
4/20/2026
Show HN: I Built SwiftUI but for macOS MDM
Astrolabe declarative macOS configuration framework inspired by SwiftUI for MDM state convergence.
4/20/2026
LLM reasoning makes multi-provider systems significantly harder to operate
Multi-provider LLM systems become harder to operate when adding reasoning capabilities due to infrastructure and abstraction challenges.
4/20/2026
Verus is a tool for verifying the correctness of code written in Rust
Verus tool for static verification of Rust code using theorem proving without runtime overhead.
4/20/2026
Cursor CLI Agent gets Debug Mode and /btw support
Cursor CLI agent adds Debug Mode for hypothesis-driven bug fixing and /btw support for side questions.
4/20/2026
Show HN: GeoFastMapAPI – open-source Fast vector and raster server for mapmakers
Open-source geospatial vector/raster server using Python, PostGIS, and OGC APIs. Built with AI assistance.
4/20/2026
Openheim – open-source LLM agent in Rust (CLI, REPL, or HTTP server)
Openheim is an open-source LLM agent written in Rust that operates as CLI, REPL, or HTTP server.
4/20/2026
Show HN: Explain The Law – Simplified legislation and executive orders using AI
Web tool using Gemini 2.5 Pro to simplify and explain legislation and executive orders.
4/20/2026
OpenAI's Chronicles is basically what I open-sourced last week, to continue?
Developer compares open-source AgentHandover project with OpenAI's Chronicles feature for agent screen monitoring.
4/20/2026
LeWorldModel: Stable End-to-End JEPA from Pixels
LeWorldModel: stable end-to-end JEPA for learning world models from raw pixels using minimal loss terms. Original ML research on self-supervised learning.
4/20/2026
Soul Player C64 – A real transformer running on a 1 MHz Commodore 64
2-layer transformer model implemented in 6502 assembly running on Commodore 64. Demonstrates transformer architecture feasibility with ~25K parameters.
4/20/2026
Show HN: I built Comrade – the security-focused AI agent
Comrade: open-source AI workspace for security teams with transparency and extensibility focus. GitHub project for agent-based workflows.
4/20/2026
Prism v11.0 – $O(1)$ Zero-Search Memory for AI Agents Using HRR and Act-R
Prism v11.0 cognitive architecture for AI agents with memory and causal reasoning. Product claims unclear with limited technical details.
4/20/2026
Agents Aren't Coworkers, Embed Them in Your Software
Discussion of embedding AI agents directly in software applications rather than standalone systems, with practical examples from Feldera's infrastructure use.
4/20/2026
Argos–AI infrastructure agent that self-deploys VMs and self-heals (open source)
ARGOS is an open-source autonomous AI infrastructure agent for monitoring and managing server fleets via natural language, built for self-hosted use without cloud dependency.