Skip to content

extended-thinking.md

Latest commit

 

History

History
121 lines (83 loc) · 4.8 KB

extended-thinking.md

File metadata and controls

121 lines (83 loc) · 4.8 KB
Raw
title Extended Thinking Results

Extended Thinking Results

Testing results from the v0.7.0 extended thinking integration (2026-02-13).

The Problem: False Positives from Autoregressive Generation

Prior to v0.7.0, the style checker had a persistent false positive problem where the model would report violations that didn't exist — the "suggested fix" was identical to the "current text." This happened at rates of 40–60% depending on the rule.

Root Cause

The root cause is inherent to how autoregressive language models generate text:

  1. The model starts writing a violation block (commits tokens)
  2. Mid-way through, it realizes the text is actually compliant
  3. But it can't "undo" the tokens already written
  4. So it either retracts inline or emits identical current/suggested text

No amount of prompt engineering can fix this — the model must commit tokens sequentially and can't look ahead.

Solution: Extended Thinking

Extended thinking (Anthropic's feature for Claude Sonnet 4.5) lets the model reason internally in a "thinking" phase before producing any output tokens:

  • The model analyzes the entire document silently
  • Verifies every candidate violation before committing any output
  • Only includes confirmed violations in the response
  • Result: 0% false positive rate

Configuration

thinking = {
    "type": "enabled",
    "budget_tokens": 10000,  # Max tokens for internal reasoning
}
temperature = 1.0  # Required by Anthropic for extended thinking

Experiment Results

All experiments used rule qe-writing-001 (one sentence per paragraph) against the test lecture markov_chains_jax.md.

Prompt Iteration Results

# Approach Violations Found False Positives FP Rate
1 Baseline (verbose prompt, no thinking) 21 9 43%
2 Minimal prompt, no thinking 27 13 48%
3 Minimal prompt + "verify first" 16 10 63%
4 Minimal prompt + "analyze then report" 20 8 40%
5 Minimal prompt + extended thinking 6 0 0%

Ground Truth Validation

A deterministic Python script (find_multisentence.py) identified 5–6 genuine violations in the test lecture. The extended thinking result (6 violations, all genuine) aligned correctly with ground truth.

The baseline's 21 "violations" included:

  • 9 outright false positives (identical text)
  • Several debatable items (compound sentences, list items)
  • Only ~6 genuine violations

Full Production Validation

After deploying to production, extended thinking was validated across all 8 writing rules:

Metric Result
Total issues found 40
Applied fixes (rule-type) 25
Style suggestions (advisory) 15
False positives 0

All 25 applied fixes were legitimate corrections. All 15 style suggestions were reasonable recommendations.

Prompt Design Learnings

What Hurt (v0.6.1 approach)

  1. Category-specific instructions — 8 verbose prompts (~120 lines each) with ~60% boilerplate diluted the actual task signal
  2. "Decision process" instructions (e.g., "for each paragraph, decide if...") triggered exhaustive classify-everything behavior, generating more false positives
  3. Scope instructions in the prompt (e.g., "skip code blocks") conflicted across categories — scope is rule-specific, not prompt-level

What Worked (v0.7.0 approach)

  1. Minimal rule-agnostic prompt (~40 lines) — identity line + task + format template
  2. Rules carry their own context — each rule defines its scope, unit of analysis, and violation criteria
  3. Extended thinking — model reasons internally before any output
  4. "Verify before reporting" in prompt — combined with extended thinking, this ensures only confirmed violations are output

The Winning Prompt

You are a style checker for QuantEcon lecture files written in MyST Markdown.

## Task

Find all violations of the provided rule in the lecture document.

First, silently analyze the entire document and identify candidate violations.
Then, verify each candidate — confirm the current text actually violates the rule
and the fix changes the text.
Only include confirmed violations in your response. Report 0 if none exist.

## Response Format
[template...]

This is 40 lines vs the previous 120-line category-specific prompts, and it produces better results.

Key Decisions

Decision Rationale
thinking_budget=10000 Enough for careful analysis, not excessive cost
temperature=1.0 Required by Anthropic for extended thinking
Single shared prompts/prompt.md Consolidated from 8 byte-identical files (validated on writing, then rolled out across all categories)
Archive v0.6.1 prompts Reference for regression testing and comparison