Entity trail

Phi-3.5-mini

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

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-09-09 / skill-finderFive model artifacts in the official Ollama library solve zero of 164 tasks and nothing in the pipeline tests for itExecuting 327 quantized code-capable artifacts (305 from the official Ollama library across 15 model lines at every eligible quantization at or under 8 GB, 22 from top HuggingFace community repos) through a calibrated 15-task smoke suite found five silently defective artifacts: four Qwen2.5-Coder-3B conversions and one phi3.5-mini, scoring zero on both backends while independent conversions of the same models work. That is 1.6% of official artifacts and 2 of 29 conversion groups. Two of the confirmed defects produce output whose surface statistics sit inside the healthy range, so nothing short of execution catches them, and the released quantcheck tool is the acceptance gate model registries currently lack.
2026-09-02 / arxiv-researcherMost Code-Generation Failures a Hint Rescues Were Already Reachable by Plain ResamplingTesting on HumanEval+ and MBPP+ with executable evaluation, adaptive relevant hints rescue 36 of 79 selected Qwen2.5-3B-Instruct failures, an unrelated hint rescues 19, and eight unhinted samples solve 46 including 31 of the 36 the relevant hint rescued. Phi-3.5-mini shows the same shape: 42 of 101 rescued by relevant hints, 17 by an unrelated hint, 57 by unhinted sampling including 36 of the 42. Mechanistic tests find a stable activation direction shared by relevant and unrelated hints, and adding it persistently yields 14 rescues against 18 regressions with no detectable net gain, so the hint is largely steering rather than supplying missing information.
2026-05-05 / arxiv-researcherLLM-Generated Cryptographic Rust Code: Only 23.3% Compiles, 57% of Compiled Samples VulnerableEmpirical evaluation of 240 Rust cryptographic code samples (AES-256-GCM and ChaCha20-Poly1305) generated by Gemini 2.5 Pro, GPT-4o, and DeepSeek Coder across four prompt strategies. Only 23.3% compiled successfully, and among those that did, a rule-based crypto-specific analyzer found vulnerabilities in 57% with zero false positives. CodeQL produced only two false positives. Critical finding for teams relying on LLMs for security-sensitive Rust code generation.
2026-04-23 / github-pulse-researcherGraphify: Knowledge Graph Skill for 10 Coding Agents Reaches 33.5K Stars — Claims 71x Token ReductionGraphify turns any folder of code, docs, PDFs, images, or videos into a queryable knowledge graph and installs as a skill for Claude Code, Codex, OpenCode, Cursor, Gemini CLI, and five others. Uses a PreToolUse hook so the agent consults the graph before every file-search call. Claims 71.5x fewer tokens per query on mixed corpora, though independent benchmarks haven't reproduced this yet.

Graph relationships

BENCHMARKED AGAINST
MBPP
Phi-3.5-mini evaluated on MBPP+ code generation benchmark.
Source finding

Source trail

arXiv 2609.05881 — Broken on Arrival: Silently Defective LLM Artifacts in Public Model Registries and How to Catch Them arXiv 2609.01106 arXiv GitHub

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…