llm-benchmark-analyst（LLM benchmark analyst 搜索）

LLM Benchmark Analyst

Overview

Use this skill to research benchmark evidence and write structured reports about:

1. a single model's strengths and weaknesses
2. best models in a capability domain
3. what a benchmark measures and how trustworthy it is
4. predecessor vs current-model progress

Default to the user's language. Never invent scores, ranks, dates, benchmark variants, or missing table values.

Core constraints

• Restrict the benchmark universe to references/benchmark-source.md. If a benchmark is not in that file, exclude it.
• Use references/core-dimensions.md to collapse scattered benchmarks into a small set of report dimensions.
• Follow references/search-playbook.md for routing, overlap expansion, evidence gathering, and comparison anchors.
• Follow references/report-template.md for output structure.
• Apply references/data-defect-warnings.md benchmark by benchmark, inline and again in the limitations section.
• Prefer official benchmark or benchmark-author pages. Use aggregators mainly to discover links and context.
• Record the evaluation mode exactly: benchmark version, split, difficulty, public/private, verified/original, with-tools/without-tools, pass@k, and any visible sub-score names.
• Keep score units exact. Do not average incompatible metrics into a fake composite.

Required workflow

1. Normalize the model identity before searching

- Resolve exact provider, family, generation, version suffix, and release label. - Put time and version first. Reject ambiguous aliases like claude, gemini pro, gpt latest, or qwen max until you have the exact currently relevant model string for the searched leaderboard rows. - Capture the evaluation time point or access date for every key score.

1. Route the request through core dimensions before web crawling

- Start with references/core-dimensions.md to select the primary dimension(s). - Then list candidate benchmarks inside those dimensions. - Only then start website-by-website retrieval. - Keep the first pass narrow and token-efficient: start from the best 3-6 benchmarks for the asked domain, then expand only if needed.

1. Expand beyond section labels

- Do not let the source document's headings blind you. - After selecting the primary dimension, inspect benchmark descriptions and overlap tags to find relevant benchmarks that live in other sections. - Example: a coding analysis may need coding benchmarks, agentic coding benchmarks, general benchmarks with coding components, and research/math benchmarks with strong code components. - Example: a multimodal analysis may need vision benchmarks, OCR, GUI/computer-use, multimodal deep-research, and omni/video/audio benchmarks.

1. Collect evidence in this order

- official leaderboard or benchmark site - benchmark paper or benchmark README - benchmark-author blog or release note - trusted aggregator - vendor blog only as secondary evidence, clearly labeled as vendor-reported if no independent leaderboard row exists

1. Use multimodal extraction when the leaderboard is not machine-readable

- If the page uses images, canvas, screenshots, or chart-only rendering and plain text extraction misses the table, inspect screenshots or page images. - Extract only values that are clearly visible. - Mark the provenance as image-extracted. - If the image is unreadable or partially occluded, say so instead of guessing.

1. Apply anchor comparisons

- For code or agentic coding, compare against the latest available Claude Opus, latest Claude Sonnet, and latest GPT family model. - For multimodal analysis, compare against the latest available Gemini model. Add the latest GPT multimodal model if relevant. - For intelligence or reasoning analysis, compare against the latest available GPT family model. - Never assume which model is currently latest. Search that first.

1. Apply predecessor comparison

- If data exists, compare the target model with its immediate predecessor or last broadly comparable prior generation from the same provider/family. - Only compare like-for-like benchmark variants. If the predecessor only appears under a different benchmark mode, say the comparison is not clean.

1. Attach defect warnings

- Any benchmark with a known quality or methodology issue must carry an inline warning from references/data-defect-warnings.md. - If the report's conclusion depends heavily on warned benchmarks, lower confidence and say so explicitly.

Decision rules

• When the user asks for best models in a domain, do not use only one benchmark. Use a cluster of relevant benchmarks and explain why each one matters.
• When the user asks for what is this model good or bad at, synthesize at the core-dimension level first, then support with benchmark evidence.
• When benchmark scores conflict, prefer freshness, exact version match, official source quality, and the number of agreeing benchmarks over one standout score.
• Treat very small gaps as non-decisive when the benchmark is noisy, image-extracted, or known to be unstable.
• Always include one short clause describing what each benchmark actually tests.

Minimum evidence to capture

For every benchmark you cite, capture:

• benchmark name
• what it tests in one short phrase
• exact model row name
• exact score and unit
• rank or relative placement if visible
• benchmark variant, split, or mode
• date or access time point
• source quality note if not official
• data warning if applicable

Output expectations

Use the matching template in references/report-template.md.

At minimum, every substantive report must include:

• a scope and identity section
• a short executive summary
• strengths
• weaknesses or gaps
• evidence table
• comparison section
• data-defect warnings and confidence
• methodology or exclusions

Resource map

• references/core-dimensions.md: benchmark routing and de-fragmentation map
• references/search-playbook.md: token-efficient search order, overlap expansion, and comparison rules
• references/data-defect-warnings.md: warning catalog and ready-to-use caution language
• references/report-template.md: output structures for single-model, domain-leader, and benchmark-explainer tasks
• references/benchmark-source.md: full allowed benchmark universe copied from the user's benchmark document

Example tasks

• analyze gpt-5's coding and agentic coding strengths and weaknesses, and compare it with the latest claude opus, claude sonnet, and gpt model
• find the best multimodal models right now using only the approved benchmark list and explain each benchmark briefly
• write a report on qwen's reasoning strengths, benchmark gaps, predecessor comparison, and all data-quality caveats
• tell me which models lead in deep research and search, with benchmark-specific warnings and freshness notes