arXiv

Models Struggle to Invert Charts into Values

The article explains that multimodal models can describe charts well but often fail to accurately recover numeric values because reading a chart requires inverting visual encodings (length, angle, position, color) into numbers. Error modes depend on the encoding (e.g., truncated y-axes, log scales, overlapping colours, legends far from marks). The author reviews benchmarks (ChartQA, PlotQA, CharXiv), recommends separating label-reading from arithmetic in evaluations, and provides practical advice: attach raw data instead of images, request extracted values before calculations, increase image resolution, and explicitly state axis properties. An example Python snippet shows how to generate grounded evaluation charts from known data.

Microsoft SkillOpt: Agents Self-Evolve via Skill Documents

Microsoft Research published SkillOpt, a research system and open-source toolkit that optimizes AI agent behavior by treating agent 'skill documents' (markdown files) as a trainable state. Instead of fine-tuning model weights, SkillOpt uses a separate optimizer model to propose bounded edits to skill files, then validates changes on held-out benchmarks. The paper reports best-or-tied results in 52/52 test cells across 7 target models (including GPT-5.5 and Claude Opus 4.8), 6 benchmarks and 3 harnesses, with large score improvements (e.g., +23.5 points on GPT-5.5 direct chat). Version v0.2.0 (2026-07-02) adds SkillOpt-Sleep, a nightly offline self-evolution engine. Code is available on GitHub (microsoft/SkillOpt) and the package can be installed from PyPI; the research paper is on arXiv (2605.23904).

Anthropic's Model Context Protocol (MCP) Explained

Model Context Protocol (MCP) is an open standard introduced by Anthropic that standardizes how applications provide external tool context to large language models (LLMs). MCP defines three core components — host, client, and server — and lets tool providers implement MCP-compatible servers so any MCP-speaking host can access tools without custom integration code. The protocol reduces the maintenance burden that arises when many different tools and provider APIs must be integrated, because updates are handled by the tool provider's MCP server rather than each host. The article outlines the MCP request/response flow and gives a minimal example (a weather MCP server exposing get_alerts and get_forecast) to demonstrate how hosts like Cursor can call MCP servers via an MCP client and return grounded results to an LLM.