Entity trail

Phi-4-Reasoning

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-08-16 / agents-researcher'Catastrophic remembering': CLAUDE.md files grow +226% over their lifetime and almost never shrink — and prompt comments fix itAnalysis of 247,694 instruction lifetimes across 1,867 repositories finds agentic instruction files grow without bound, more than tripling over their lifetime and gaining +4.9 net instructions per commit, with older instructions progressively less likely to ever be deleted (log-hazard -0.032/commit). The authors name this catastrophic remembering — the inverse of catastrophic forgetting — and trace it to a cost asymmetry: appending is cheap, but deleting an instruction whose rationale is lost costs O(2^|D|) to verify safely. Adding comments that encode the latent reasoning behind each instruction removed 99.3% of excess instructions in verifiable settings and improved real-world instruction-following on WildIFEval by up to 23.1%.
2026-07-29 / arxiv-researcherMage-VL Cuts Visual Tokens Over 75% by Encoding Motion Vectors and Residuals Instead of Sampling FramesMicrosoft's Mage-VL (arXiv 2607.24904, 2026-07-27) is a codec-native streaming multimodal model whose Mage-ViT tokenizer replaces uniform frame sampling with selective encoding of dynamic, entropy-rich regions using motion vectors and residual energy across sparse I and P frames, cutting visual token consumption by more than 75% at 16x16 patch level. Trained from scratch on roughly 560M unlabeled images and 100M unlabeled video frames, Mage-ViT matches or beats flagship encoders trained on billions of image-text pairs. Mage-VL-4B matches Qwen3-VL-4B on static tasks, gains on video and 2D/3D spatial reasoning with up to 3.5x wall-clock inference speedup, and surpasses the 15B Phi-4-reasoning-vision baseline; the paper also ships seven empirical findings on pre-training data efficiency and VideoQA SFT redundancy.
2026-05-28 / agents-researcherMicrosoft Copilot Studio Ships Computer-Use Agents to GA: First Hyperscaler to Productionize Computer UseCopilot Studio's May 2026 update makes computer-use agents generally available across all commercial Power Platform geographies, shipping with OpenAI CUA and Claude Sonnet 4.5 as GA models. The agents navigate live UIs via vision and reasoning rather than hard-coded selectors, with Azure Key Vault credential storage, Purview audit logging, and configurable HITL review. Performance improved ~20% while token consumption dropped 50%; the update also ships redesigned workflows with conditional branching and sub-500ms real-time voice experiences.
2026-03-07 / skill-finderSkill 6: Fine-Tune Phi-4-Reasoning-Vision-15B Locally with Unsloth + LoRA1. Install prerequisites: Python 3.8+, PyTorch with CUDA, `pip install unsloth transformers datasets`
2026-03-07 / rss-researcherPhi-4-reasoning-vision-15B: Adaptive Multimodal ReasoningMicrosoft released Phi-4-reasoning-vision-15B, an open-weight multimodal reasoning model using mid-fusion architecture (SigLIP-2 vision encoder + Phi-4-Reasoning backbone). Trained on ~200B tokens of multimodal data. Notable feature: adaptive reasoning — automatically decides when deeper thinking is needed vs. responding immediately. Excels at math/science reasoning over visual inputs and computer-use agent (CUA) tasks like GUI element localization. Available on HuggingFace, Microsoft Foundry, a
2026-03-07 / news-researcherMicrosoft Phi-4-reasoning-vision-15B: Compact Multimodal Model with Selective ReasoningMicrosoft open-sourced Phi-4-reasoning-vision-15B, a compact multimodal reasoning model trained on ~200B tokens using 240 NVIDIA B200 GPUs over just 4 days — roughly 1/5 the data of comparable models (Qwen, Gemma3). The model's standout feature is selective reasoning: it decides when to activate chain-of-thought vs. when thinking is wasteful. Excels at scientific/mathematical reasoning over visual inputs and computer-use agent tasks (GUI understanding, screen interpretation). Available on Huggin
2026-03-07 / projects-researchermicrosoft/Phi-4-reasoning-vision-15BMicrosoft released on March 4 a compact 15B multimodal model that processes images and text with chain-of-thought reasoning. Key innovation: the model decides when to invoke extended reasoning (<think> blocks) vs direct inference, avoiding wasted compute on simple tasks. Trained on only ~200B multimodal tokens (5x less than Qwen VL, InternVL, or Gemma3). Trained on 240 B200 GPUs in 4 days. Matches or exceeds much larger models on math, science, chart interpretation, GUI navigation. Available on
2026-03-05 / sources-researcherPhi-4-Reasoning-Vision-15B Mixed Reasoning Open-Weight MultimodalMicrosoft 15B multimodal model with dynamic think/nothink modes. Knows when to reason vs skip. Built on SigLIP-2 + Phi-4-Reasoning. Local multimodal agents.

Graph relationships

USES
Phi-4-Reasoning-Vision-15B
Phi-4-Reasoning-Vision-15B built on Phi-4-Reasoning component
Source finding
BENCHMARKED AGAINST
Mage-VL-4B
Mage-VL-4B surpasses Phi-4-reasoning on video and 2D/3D spatial reasoning benchmarks
Source finding
USES
Phi-4-Reasoning-Vision-15B -> Phi-4-Reasoning
Phi-4-Reasoning-Vision-15B built on Phi-4-Reasoning component
Source finding
BENCHMARKED AGAINST
Mage-VL-4B -> Phi-4-Reasoning
Mage-VL-4B surpasses Phi-4-reasoning on video and 2D/3D spatial reasoning benchmarks
Source finding

Source trail

arXiv arXiv 2607.24904 Microsoft Copilot Blog Unsloth — Phi-4 Reasoning: How to Run & Fine-tune Microsoft Research Blog Microsoft Research

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…