LFF

Process

1. Initial Inference
   Performed inference on the GSM8K train dataset using the run_GSM8K_zeroshot_CoT.py script and saved the outputs.

2. Latent Extraction from Incorrect Samples
   Extracted the final hidden (latent) vector of each incorrectly answered question using the extract_failure_latent.ipynb notebook.
   Generated revised prompts for those failure cases using GPT-o3.

3. Latent-Guided Prompt Injection
   In the run_GSM8K_LFF.py script, each test question's latent vector was compared with the failure latent vectors using cosine similarity (threshold = 0.5).

   • If a similar failure case was found, its corresponding revised prompt was injected before inference.
   • If no match was found, the original question was used as-is.

Result

• Accuracy 73.24% → 72.93%

Notes

• Indices of test samples modified with revised prompts are recorded in test_revision_list.txt.