Applied AI Digest

New Agent Benchmarks

• 🔥 AgentProcessBench: AgentProcessBench diagnoses step-level process quality in tool-using agents.
• One-Eval: One-Eval...

Key PIM event ahead: 5th Workshop on Memory-Centric Computing Systems, 23 March 2026.

• Focus: Latest PIM advances in hardware/software, tools,...

InCoder-32B launches as a code foundation model for industrial scenarios. New paper available—essential read for enterprise coding advances.

• V-Co method scrutinizes visual representation alignment using co-denoising for VLMs.
• Power-aware analysis evaluates vision-language models on...

Baidu Qianfan Team's Qianfan-OCR, a 4B-parameter end-to-end document intelligence model on ModelScope, unifies document parsing, layout analysis, and more—pushing multimodal OCR for real-world AI.

• Verification for research: MiroThinker-1.7 advances heavy-duty research agents through verification techniques.
• Planning-execution split: Capy's...

Mistral AI unveils Forge, a platform enabling companies to train their own large language models based on its open-weight models. Key tooling for enterprise customization.

Emerging frameworks expose gaps in agentic LLMs across domains:

• FinToolBench evaluates real-world financial tool use
• PokéAgent tests competitive...

Emerging papers signal a trend in perception-to-simulation fidelity for embodied AI and gaming:

• Kinema4D enables kinematic 4D world modeling for...

New paper introduces masked modeling as a strategy for efficient image-only pre-training in UMM visual generation.

TRUST-SQL pioneers tool-integrated multi-turn reinforcement learning to enable robust Text-to-SQL over unknown schemas, advancing SQL agents for real-world databases.

SocialOmni benchmarks audio-visual social interactivity in omni models, targeting gaps in multimodal social cues for researchers.

Nvidia pivots to full-stack AI with a $26B investment over five years for open-weight models, challenging OpenAI and Anthropic.

Key highlights:
-...

New Open-Source Releases

• 🔥 Mistral Small 4 and Leanstral: Mistral released Leanstral and Small 4 models under Apache 2.0, with Small 4...

PokeAgent Challenge is a large-scale benchmark for decision-making research, built on Pokemon's multi-agent battle system and expansive mechanics—ideal for probing competitive agent coordination.

Emerging ecosystem of cost-sensitive models accelerates consumer-grade AI:

• OpenAI GPT-5.4 mini matches flagship perf within 5% on SWE-Bench...

Mixture-of-Depths Attention (MoDA) enables each attention head to attend to sequence KV pairs at the current layer and depth KV pairs, promising dynamic depth mixing for long-context efficiency.

• New benchmark: InterveneBench assesses LLMs' intervention reasoning in realistic social settings
• Targets state-of-the-art LLMs with experimental results
• Key for evaluating causal intervention in agentic scenarios

Mistral AI's Mistral Small 4 packs 119 billion parameters into a compact MoE with 128 experts, excelling in fast text responses, logical reasoning, and image processing.

Key breakthrough for applied vision:

• Stereo foundation models deliver strong zero-shot generalization
• But remain computationally prohibitive for real-time use
• Fast-FoundationStereo unlocks real-time zero-shot stereo matching via GitHub release