Independent model evaluation

Compare model capability where real work happens.

Lens Frontier builds real benchmarks, runs models, and audits the evidence. We publish not only who leads, but why each score holds.

Explore all benchmarks

1 benchmark published

Benchmark model rankings

Choose one benchmark to see what it measures, who leads, and why the result is trustworthy.

All benchmarksCode-QA-Bench

Swipe through published evaluations; select one to switch

Model rankingHigher scores rank first4 models

Benchmark system

Every score resolves to a traceable Data Card.

The leaderboard reveals differences. The Data Card defines task scope, provenance, judgment, and status. They are two views of the same evaluation project.

01Released

Code-QA-Bench

A fully automated repository-level code understanding benchmark. Closed-book, code-only, and documented conditions separate genuine code reading from documentation recall and pretraining memorization. The paper reports 528 code-derivable and 100 doc-dependent tasks across ten pinned Python repositories.

Capability claim: Locate, trace, and explain implementation details in real repositories while separating code reading, memorization, and documentation use.
Data source: 10 pinned SWE-Bench Python repositories
Judge method: GPT-5.4 LLM judge · three-axis continuous score

Research & releases

Team papers, evaluation methods, benchmark releases, and model analysis.

View all research

Insight 5/21/2026

Building a Multi-Module Collaborative Evaluation Framework for T2I

An LLM-as-a-Judge evaluation system with OCR/VQA/CLIP/IQA/HP modules and cross-module calibration, breaking through single-metric benchmark limitations to capture the true capability boundaries of modern T2I models.

Evidence standard

Trusted scores require a complete evidence chain.

Task, protocol, run, and judgment must remain mutually traceable.

View research publications

01
Scope the task
Define repository, environment, tools, and success.
02
Freeze protocol
Fix resources, constraints, judge, and version.
03
Capture evidence
Retain outputs, artifacts, traces, and exceptions.
04
Review and publish
Separate verified, reviewed, and pending results.