Applied AI Digest

Agent Evaluation Benchmarks

• 🔥 One-Eval: One-Eval is an agentic system for automated and traceable LLM evaluation.
• 🔥 SWE-Skills-Bench:...

CV segmentation matures beyond proof-of-concept, blending text guidance and efficiency:

• Special issue advances image classification and...

Capy uses specialized captain (planning) and build (execution) agents to improve AI tool quality, as planning decides success. Most tools merge them, but this separation cuts iteration loops for higher outputs.

WorldCam uses camera pose as a unifying geometric representation for interactive autoregressive 3D gaming worlds. Paper out now.

Collinearity, a human visual phenomenon amplifying spatially aligned edges along straight lines, is the focus of new research transferring this perception to computer vision models.

• Industry defaults to fine-tuning for domain adaptation as it seems cheaper, ignoring inference costs.
• @datologyai's work: Mixing domain data...

InCoder-32B launches as a code foundation model tailored for industrial scenarios, with new paper out now.

TRUST-SQL advances Text-to-SQL with tool-integrated multi-turn reinforcement learning over unknown schemas. Key for robust agents in dynamic querying—join the discussion.

M^3 fuses dense matching with multi-view foundation models to enable monocular Gaussian Splatting SLAM. Join the discussion on this CV advance.

• MiroMind releases verification-based research agents MiroThinker-1.7 and MiroThinker-H1.
• Targets heavy-duty research through verification approaches.
• Emphasizes Effective Interaction Scaling to boost reasoning quality over quantity.

Emerging diagnostics reveal limitations in agent performance:

• AgentProcessBench targets step-level process quality in tool-using agents
  -...

• Core threat: Resource consumption in large language models
• Efficiency exploit defense: LightThinker Compressor reduces token footprint for KV cache in LLMs
• Overview: Table 9 summarizes representative defenses (Part 1)

New paper introduces online experiential learning for language models. Join the discussion on this paper page.

New paper proposes masked modeling for efficient image-only pre-training in UMM visual generation, targeting compute-efficient vision foundation models. Join the discussion.

New Agent Benchmarks

• 🔥 PokeAgent Challenge: PokeAgent Challenge is a large-scale benchmark for decision-making research built on Pokemon's...

Key advances in efficient LLM context compaction:

• Morph SLM: Dedicated model trained for compaction; ideal use case where SLMs fit and should...

Key highlights from the new multi-agent framework:

• Bridges natural language and GIS using LLMs in a multi-agent setup
• Promising results for LLM...

Key insights from "Beyond Language Modeling":

• Transfusion framework trains a unified autoregressive model from scratch, using next-token for...

Large language models (LLMs) integration into qualitative health research enhances efficiency and depth of data analysis. A game-changer for practical healthcare applications.

Practical offline LLM agent for business visualization:

• Converts natural or business process steps to HTML sequential diagrams.
• Runs fully...