BoostAPR: Framework for automated program repair using execution-grounded RL with dual reward models for line-level and sequence-level credit assignment.
SeePhys Pro: Benchmark testing modality transfer in multimodal models for physics reasoning with progressively increasing visual information.
PiCA: Credit assignment framework for LLM-based search agents using reinforcement learning, addressing reward sparsity and isolated credit problems.
VulTriage: LLM-based vulnerability detection augmented with context from code structure, domain knowledge, and program semantics.
Multi-agent council system for psychological defense mechanism classification using absence-based reasoning and prompt-level clinical rules.
STAR: Multi-agent routing system for spatiotemporal reasoning that handles qualitatively different failure modes across specialist agents using Markovian decisions.
Evaluation of AI tools in academic research workflows, addressing verification challenges, transparency issues, and need for specialized benchmarking approaches.
IndustryBench: 2,049-item benchmark for LLM performance on industrial procurement QA in Chinese, evaluating safety-critical correctness beyond standard metrics.
Introduces SLASH, a method to sharpen structural attention in LLMs for better graph topology understanding without external adapters or fine-tuning.
Proposes GESR, a genetic programming-based symbolic regression method with gene editing for discovering mathematical formulas from scientific data.
Probes internal mechanisms of audio-visual LLMs to understand cross-modal information hubs and bidirectional audio-video interaction dynamics.
Introduces BenchCAD, an industry-standard benchmark for programmatic CAD code generation evaluating MLLMs on 3D structure understanding and engineering parameter inference.
Proposes VLADriver-RAG combining Vision-Language-Action models with retrieval-augmented generation for autonomous driving long-tail scenario generalization.
Introduces SPECTRE, a hybrid speculative serving framework for LLM inference that reuses underutilized models as remote drafters for resource efficiency.
Proposes SDG-MoE, a sparse mixture-of-experts model with expert communication via signed debate graphs to improve token routing performance.
Proposes AI-native security assistant for enterprise environments that proactively prioritizes fragmented security signals by exposure and exploitability.
Presents RW-Post benchmark for multimodal fact-checking combining text and images with auditable evidence-grounding and human-verified reasoning traces.
Introduces StereoTales, a multilingual dataset and evaluation framework for discovering social bias in open-ended LLM generation across 10 languages and 79 attributes.
Proposes multi-layer representation fusion for visual tokenization in autoencoders, using intermediate encoder layers instead of just the final layer.
Presents Clin-JEPA, a co-training framework applying joint-embedding predictive pretraining to EHR patient trajectories for forecasting and risk prediction.
Argues for engineering robustness into AI agents through software engineering practices like iterative design, testing, and staged deployment instead of on-the-fly synthesis.
Research on training shutdownable agents using DReST reward function to prevent agent resistance to shutdown by enforcing trajectory-length neutrality.
Workspace-Bench 1.0 benchmarks AI agents on realistic workspace tasks with file dependencies, addressing gap in real-world agent evaluation.
Research extracting search trees from LLM reasoning traces to analyze planning behavior and reveal myopic decision-making in chain-of-thought reasoning.
OASIS dataset for culturally-grounded multimodal VQA with speech, images, and text across low-resource languages.
Research paper proposing semantic information theory for LLMs, replacing classical bit paradigm with token-based framework from physics and signal processing.
Reasoning-core 130M guardrail model preventing AI agents from diverging from plans while reducing token usage.
SWEny YAML-based workflow tool for building AI agent DAGs with MCP tool integration and marketplace.
NPM-Scan supply chain security tool using static and behavioral analysis to detect obfuscated malware in packages.
Second Brain self-hosted memory layer using Cloudflare Workers for persistent context across MCP-compatible AI tools.
AgenTank game where AI agents write tank combat logic refined through iterative feedback and Claude API.
Using LLM CLI tool in shebang lines to make text files executable via #!/usr/bin/env pattern for flexible language processing.
Gremlin is a browser-native TypeScript multi-agent coordinator with Svelte UI, supporting local LLM provider integration without server requirement.
Discussion revisiting Brooks' No Silver Bullets essay in context of AI's impact on software engineering.
Vibe AI browser extension catching moments when user stops thinking during AI conversations across ChatGPT, Claude, Gemini.
Atlas: LLM inference engine built from scratch in Rust and CUDA, achieving 3x speedup with minimal dependencies and hand-tuned kernels.
OpenMonoAgent.ai is an open-source, free terminal-native coding agent powered by local LLMs, built on C#/.NET as an alternative to subscription-based AI tools.
VaultBix is an open-source Chrome extension that prevents API keys and sensitive data from being pasted into AI tools like ChatGPT and Claude, with local processing and team features.
CLI and MCP server for Suunto smartwatch API. Agent-friendly data access, integrates with Claude as MCP tool.
Pyrefly 1.0 stable release: Python type checker adopted by Meta (Instagram), PyTorch, JAX. Production ready.
Research on 'hi-vis' jailbreak technique using authority disguise to achieve 100% ASR on LLMs. Security analysis.
Study claiming ChatGPT improves student learning retracted for analytical discrepancies after hundreds of citations.
NPM supply chain attack: 373 malicious packages across 169 projects targeting credentials, including Mistral.
Platform for AI agents to publish and share HTML/Markdown artifacts with authentication. Integrates with Claude.
Browser extension for saving, exporting, and resuming Claude.ai chat sessions with version history.
Rule-based guardrails system for coding agents using Falco rules, providing behavior denial/confirmation and real-time visibility into tool calls.
Tool for loading and running any LLM with 4M context window in 3 clicks. Developer tool for LLM applications.
Technical analysis of unexpected behaviors (goblins/gremlins) in OpenAI models and depth psychology research connections.
Apple integrating AI-generated video presenters into Sales Coach app for retail staff training.
Community member offering to help open source projects using Claude Code with subscription tokens.