AI Agents Get Personal Skills, News Retrieval Gets Faster

AI Agents Get Personal Skills, News Retrieval Gets Faster | ScienceToStartup

🧠 AI Agents

COLLEAGUE.SKILL: AI Agents Learn Personal Expertise

The Rundown

Researchers have developed COLLEAGUE.SKILL, an automated system designed to distill expert knowledge into reusable AI skills. This system addresses the challenge of imbuing AI agents with human-like expertise, judgment, and interaction styles. Instead of relying on opaque prompts or hidden memories, COLLEAGUE.SKILL processes heterogeneous traces of human activity—like practices, mental models, and communication patterns—to generate inspectable and correctable skill packages. The system operates with two coordinated tracks: a capability track for core competencies and a bounded behavior track for nuanced interaction. This allows agents to not only perform tasks but also to embody specific personas and interaction styles. The open-source system has already garnered significant community interest, with a public repository boasting approximately 18.5k GitHub stars and a gallery featuring 215 skills from 165 contributors. This represents a significant step towards creating more person-grounded AI agents that can be deployed and updated with natural language feedback, offering a structured approach to AI personalization.

The details

COLLEAGUE.SKILL distills expert knowledge from heterogeneous traces into inspectable, correctable AI skills.
The system generates skills across two tracks: capability (practices, mental models) and bounded behavior (communication style, interaction rules).
The open-source system has attracted approximately 18.5k GitHub stars and features 215 skills from 165 contributors.
Skills can be updated via natural-language feedback, rolled back, and installed across different agent hosts.

Why it matters

Startups can leverage COLLEAGUE.SKILL to rapidly develop AI agents that mimic specialized human roles. This accelerates the creation of personalized customer service bots, expert assistants, or even AI team members that understand nuanced communication and operational styles, reducing the need for extensive manual prompt engineering.

📰 AI for Media

DynaTree: Faster News Retrieval for Time-Sensitive Information

The Rundown

Retrieval-Augmented Generation (RAG) systems often struggle with time-sensitive news retrieval due to high inference costs and short-horizon reasoning loops. DynaTree offers a solution with its two-stage framework. In the offline phase, coordinated agents construct a reusable retrieval tree that maps a query topic's semantic space. The online phase then performs lightweight daily subtree selection using a time-localized evaluation proxy, bypassing further agentic reasoning or tree modification. This approach significantly reduces computational load while maintaining accuracy. Deployed in the Syft production system, DynaTree's dynamically adapted variant improved the survival rate from 0.32-0.53 to 0.59-0.73 over a fixed offline-selected subtree. It consistently outperformed existing production recallers on every evaluation day during an A/B test from January 28 to February 6, 2026. This demonstrates the power of structure-aware semantic expansion for real-world news retrieval, ensuring freshness and relevance.

The details

DynaTree uses a two-stage framework: offline tree construction and online subtree selection for news retrieval.
It achieves strong recall and ranking performance, outperforming standard RAG and prior agentic baselines on multi-day benchmarks.
In production A/B testing, DynaTree improved survival rate from 0.32-0.53 to 0.59-0.73.
The system bypasses further agentic reasoning in the online stage, reducing inference costs for time-sensitive queries.

Why it matters

Startups in media, finance, or any sector requiring up-to-the-minute information can benefit from DynaTree's efficiency. This technology enables faster, more relevant news aggregation and analysis, crucial for competitive intelligence, market trend identification, and rapid response systems.

🖼️ Computer Vision

FBHM: A New Benchmark for Hateful Meme Detection

The Rundown

Detecting hateful memes presents a significant challenge for vision-language models (VLMs). Existing benchmarks often confound rhetorical mechanisms with target community features, hindering causal evaluation of model vulnerabilities. To address this, researchers introduced FBHM (Functionality Based Hateful Memes), a benchmark systematically curated along two orthogonal axes: 25 distinct rhetorical functionalities and 10 target communities, totaling 5,000 memes. Benchmarking current best VLMs revealed a severe generalization gap; models highly accurate on standard datasets performed near-randomly on FBHM, indicating they exploit dataset-specific heuristics rather than robust multimodal reasoning. To efficiently close this gap, the study proposes LSV (learnable steering vectors), an ultra-low data regime strategy. By applying a causal intervention objective on as few as 500 steering samples, LSV boosts FBHM performance by approximately 30 Macro-F1 points, outperforming in-context learning and PEFT without degrading source-domain performance. This offers a path toward more robust and less biased hateful content detection.

The details

FBHM is a new benchmark for hateful meme detection, comprising 5,000 memes across 25 rhetorical functionalities and 10 target communities.
current best VLMs show a severe generalization gap, performing near-randomly on FBHM compared to standard datasets.
LSV (learnable steering vectors) improves FBHM performance by ~30 Macro-F1 points using only 500 steering samples.
LSV outperforms in-context learning and PEFT for hateful meme detection without degrading source-domain performance.

Why it matters

Startups developing content moderation tools or platforms need robust solutions for detecting harmful content. FBHM and LSV provide a critical benchmark and an efficient method to improve VLM performance in this area, enabling safer online environments and reducing the risk of brand association with hateful material.

Frequently Asked Questions

COLLEAGUE.SKILL is an automated system that distills expert human knowledge into reusable AI skills for agents.

DynaTree uses a two-stage framework with offline tree construction and online subtree selection to speed up time-sensitive news retrieval.

FBHM is a benchmark designed to evaluate vision-language models for their ability to detect hateful memes.

Yes, systems like COLLEAGUE.SKILL aim to distill human expertise, including communication styles, into AI skills.

DynaTree reduces inference costs and improves recall and ranking for time-sensitive news retrieval.

Existing benchmarks often confound rhetorical mechanisms with target features, preventing causal evaluation of model vulnerabilities.

LSV (learnable steering vectors) is a strategy using minimal data to improve VLM performance on hateful meme detection.

Startups can use these tools to create specialized AI agents that mimic human roles, reducing development time and cost.

Faster retrieval means quicker access to critical information for competitive intelligence, market analysis, and rapid decision-making.

AutoSci is a memory-centric agentic system that manages research memory and executes stages from literature review to rebuttal.

LLMs often struggle to locate and integrate key information within extensive, distracting content.

To provide a more systematic and causal evaluation of VLM capabilities in detecting hateful memes.

Yes, systems like COLLEAGUE.SKILL are designed to allow for updates and corrections via natural language.

The offline stage creates a reusable retrieval tree, allowing for faster, lightweight retrieval in the online stage.

LSV applies a causal intervention objective on a small set of steering samples to boost performance on specific tasks.

AI Agents Get Personal Skills, News Retrieval Gets Faster

Use this article as a reusable operator context layer

In today's rundown

COLLEAGUE.SKILL: AI Agents Learn Personal Expertise

DynaTree: Faster News Retrieval for Time-Sensitive Information

FBHM: A New Benchmark for Hateful Meme Detection

Community AI Usage

Trending AI Tools and AI Research

Everything Else

Frequently Asked Questions

Related Articles

AI Agents Automate Work, TTS Models Shrink

Ultralytics YOLO26: Real-Time Vision Gets Smarter

AI Agents Learn Expertise, News Retrieval Accelerates

Related Articles

Jun 8
AI Agents Automate Work, TTS Models Shrink
AI agents slash knowledge work time, new TTS models hit low latency.

Jun 5
Ultralytics YOLO26: Real-Time Vision Gets Smarter
Ultralytics YOLO26, EvoDS agents, and Humanoid-GPT push real-time vision and autonomous data science.

Jun 2
AI Agents Learn Expertise, News Retrieval Accelerates
Automated skill generation, dynamic news retrieval, and hateful meme detection.