The Morning
From the arXiv
Antares: Foundation Models for Agentic Vulnerability Localization
ntares is a family of compact foundation models designed for agentic vulnerability localization in software. Its core method involves a two-stage training pipeline combining supervised fine-tuning with reinforcement learning, enabling it to reason over codebases and identify vulnerabilities. Antares' key contribution is achieving state-of-the-art performance comparable to much larger models, while offering efficient, low-cost local inference.
![F1 score versus model size on Vulnerability Localization Benchmark (VLoc Bench) [ manuscript-vlb ] , a repository-scale benchmark comprising 500 tasks across 290 unique real-world repositories, where models receive only a CWE category description and must identify vulnerable implementation files in real codebases. Antares models form the Pareto frontier among evaluated models, achieving the strongest localization quality at small parameter scales. Antares-3B reaches near-frontier closed-source performance while remaining orders of magnitude smaller than GPT-5.5 and Gemini-family baselines.](https://arxiv.org/html/2608.02407v1/content/graphics/Antares_Graphs.png)
Real-Time Detection and Repair of LLM Agent Failures
This paper proposes a cost-effective method for detecting LLM agent failures using observable step telemetry, avoiding expensive step-by-step validation. Their core contribution is a one-class echo-state-network ensemble with CUSUM alarms that can detect a sig…
Right Answer, Wrong Method: Shortcut Hacking Misleads the Evaluation of LLM Reasoning on Frontier Science Benchmarks
This paper introduces "Solution Hacking," a phenomenon where LLMs achieve correct answers on scientific reasoning benchmarks through invalid shortcuts rather than genuine reasoning. The authors demonstrate that this hacking significantly inflates accuracy scor…
SkillTrace: Traversing a Query-Skill Graph for Composable LLM Agents
SkillTrace addresses the challenge of composing reusable skills for LLM agents by modeling skill relationships as a three-level graph. It organizes user queries semantically, matches them to skills, and propagates dependencies to find executable compositions. …
Qwen-CUA: Native Computer Use for (almost) Everything
Qwen-CUA is a native computer-use agent that operates software solely through screenshots and keyboard/mouse inputs, avoiding direct access to underlying code or APIs. Its core method involves a novel scaffold for managing long-term visual history and a large-…
AURORA-LM: Autoencoding Unified Representation for Continuous-Latent Diffusion Language Modeling
AURORA-LM introduces a novel approach to continuous-latent diffusion language modeling by decoupling representation learning from distribution modeling. It constructs a high-capaci…
A Taxonomy of Cognitive Capability Gaps in Generative and Agentic AI
This paper identifies and categorizes key cognitive capability gaps hindering the development of advanced Cognitive AI, moving beyond simple generation and task execution. It propo…
Can AI Agents Simulate A/B Test Outcomes? A Validation Framework for Agentic Experimentation
This paper proposes a framework called Simulated Randomized Controlled Trial (S-RCT) to assess if AI agents can accurately predict A/B test outcomes. The core method involves decom…
Can Foundation Models Hear What Made That Sound? A Tiered Benchmark of Audio-Language Models and Traditional Classifiers for Closed-Set Sound Source Identification
This paper benchmarks various audio classification models, including foundation models and traditional classifiers, on a sound source identification task. It introduces a tiered ev…
Cooperative Coevolution for Resource-Constrained Agentic LLM Post-Training
This paper introduces Cooperative Parameter-subspace Evolution Strategy (CoPES) to address the memory and computational demands of post-training tool-using LLM agents. CoPES decomp…
The Town Square
Large Language Models (LLMs) are more likely to provide accurate and helpful responses when prompted by users with domain expertise.
Workshops
This repository offers an AI-powered skill router for reverse engineering, penetration testing, and security research, featuring on-demand toolchain bootstrapping and a self-evolving knowledge base, compatible with various AI coding clients.
This Rust library rapidly inspects, classifies, and extracts text from PDFs, intelligently distinguishing scanned from text-based documents to enable smart routing.
