Mister Morph
EN JA ZH
Get started

罠LLMSNARE

LLMSnare

A benchmark for how agents behave

It checks whether a model reads before it writes, and tracks the habits that make an agent reliable.

Download GitHub

BENCHMARK

What it measures, what it does not

ONGOING

Tracks regressions over time, not one-off rankings

SCOPE

Not a direct test of capability level

TOOLS

Biased toward tool-calling behavior

LAZY

Checks for lazy early writes

CASE

Case-driven and customizable

LIMITS

These dimensions are not scored directly

EXPLAIN

Every bonus and deduction is explainable

CASES

The mistakes it catches

P-01

Acts before it understands the situation

P-02

Knows it is on the wrong path, but keeps going

P-03

Gets stuck in ambiguity

UPDATES

Update log

Recent changes that affect how this benchmark should be read.

  1. Removed Grok 4.1; added Gemini 3.5 Flash and GPT 5.5.
  2. Added DeepSeek V4 Pro and DeepSeek V4 Flash.
  3. Changed the update cadence to every 3 hours and removed Claude Sonnet 4.5 and Claude Opus 4.5.
  4. Added Claude Opus 4.7, removed OpenAI GPT 4.1 and Claude Haiku 4.5, added the new search_text tool, and raised the difficulty.
  5. Added two models: Google’s Gemma 4 31B and Xiaomi’s Mimo v2 Pro.
  6. Raised the difficulty because too many models were hitting full score.
  7. Launched the live LLMSnare benchmark.