Improving LLM Accuracy with Targeted Critiques

# Illustrative sketch (not from the paper)
import torch
from typing import List, Dict

# Placeholder: load a small LLM that will generate critiques (CritBank construction)
small_model = torch.hub.load('huggingface/pytorch-transformers', 'model', 'qwen2.5-1.5B-Instruct')
# Placeholder: load the strong target model
target_model = torch.hub.load('huggingface/pytorch-transformers', 'model', 'Qwen2.5-72B-Instruct')

# ------------------------------------------------------------------
# 1. Failure‑mode profiling (static example – aggregate distribution)
# In practice this would be pre‑computed; here we mock it as a dict.
static_profile = {"arithmetic": 0.4, "algebra": 0.35, "geometry": 0.25}

# ------------------------------------------------------------------
# 2. Retrieve relevant critique examples from CritBank

def retrieve_critiques(predicted_modes: Dict[str, float], critbank: List[Dict], k: int = 5) -> List[str]:
    """Select up to *k* critiques whose labeled failure mode overlaps most with *predicted_modes*.
    critbank entries are assumed to have keys: 'mode' and 'critique'."""
    # Simple overlap scoring: sum of probabilities for matching mode
    scored = []
    for entry in critbank:
        mode = entry["mode"]
        score = predicted_modes.get(mode, 0.0)
        scored.append((score, entry["critique"]))
    # Sort descending and take top‑k non‑zero scores
    scored.sort(reverse=True, key=lambda x: x[0])
    return [c for s, c in scored[:k] if s > 0]

# ------------------------------------------------------------------
# 3. Targeted inference with retrieved critiques as in‑context examples

def generate_with_critiques(question: str, critiques: List[str]) -> str:
    # Build prompt: prepend each critique as a separate example
    prompt = "\n".join(critiques) + "\nQuestion: " + question + "\nAnswer:"
    # Placeholder generation call – in reality use target_model.generate(...)
    output = target_model.generate(prompt)
    return output

# ------------------------------------------------------------------
# Mock CritBank (normally built offline from weak model CoT failures)
critbank = [
    {"mode": "arithmetic", "critique": "Avoid assuming addition is commutative when signs differ."},
    {"mode": "algebra", "critique": "Check that you distribute multiplication over addition correctly."},
    {"mode": "geometry", "critique": "Remember that the sum of interior angles of a triangle is 180°."},
]

# Example usage
question = "What is the result of -3 + 5?"
# Use static profile as the predicted failure‑mode distribution for this input
critiques = retrieve_critiques(static_profile, critbank, k=5)
answer = generate_with_critiques(question, critiques)
print("Answer:", answer)

// Illustrative sketch (not from the paper)
const { pipeline } = require('@xenova/transformers'); // placeholder library

// Load a weak model for CritBank construction (mocked here)
const weakModel = await pipeline('text-generation', { model: 'qwen2.5-1.5b-instruct' });
// Load the strong target model
const targetModel = await pipeline('text-generation', { model: 'Qwen2.5-72B-Instruct' });

// ------------------------------------------------------------------
// 1. Static failure‑mode profile (pre‑computed distribution)
const staticProfile = { arithmetic: 0.4, algebra: 0.35, geometry: 0.25 };

// ------------------------------------------------------------------
// 2. Retrieve relevant critiques from CritBank
function retrieveCritiques(predictedModes, critBank, k = 5) {
  // Score each critique by the probability of its failure mode
  const scored = critBank.map(entry => {
    const score = predictedModes[entry.mode] || 0;
    return { score, critique: entry.critique };
  });
  // Sort descending and keep top‑k with non‑zero score
  scored.sort((a, b) => b.score - a.score);
  return scored.filter(item => item.score > 0).slice(0, k).map(item => item.critique);
}

// ------------------------------------------------------------------
// 3. Targeted inference – prepend critiques as in‑context examples
async function generateWithCritiques(question, critiques) {
  const prompt = critiques.join('\n') + `\nQuestion: ${question}\nAnswer:`;
  const result = await targetModel(prompt);
  return result[0].generated_text;
}

// ------------------------------------------------------------------
// Mock CritBank (normally built from weak model CoT failures)
const critBank = [
  { mode: 'arithmetic', critique: 'Avoid assuming addition is commutative when signs differ.' },
  { mode: 'algebra', critique: 'Check that you distribute multiplication over addition correctly.' },
  { mode: 'geometry', critique: 'Remember that the sum of interior angles of a triangle is 180°.' },
];

// Example usage
(async () => {
  const question = 'What is the result of -3 + 5?';
  const critiques = retrieveCritiques(staticProfile, critBank, 5);
  const answer = await generateWithCritiques(question, critiques);
  console.log('Answer:', answer);
})();