bowen-upenn/llm_token_bias
[EMNLP 2024] A Peek into Token Bias: Large Language Models Are Not Yet Genuine Reasoners
This project offers a testing framework to determine if large language models (LLMs) genuinely understand reasoning tasks or merely rely on superficial patterns in their input. It takes various reasoning problems and systematically alters seemingly irrelevant words to see if the LLM's answers change. The output indicates if an LLM is susceptible to 'token bias,' suggesting a lack of true understanding. Researchers, evaluators, and practitioners working with LLMs would use this to assess model robustness beyond simple accuracy scores.
No commits in the last 6 months.
Use this if you need to rigorously evaluate whether an LLM's reasoning capabilities are genuine or if the model is just picking up on specific words and phrases.
Not ideal if you are looking for methods to improve an LLM's reasoning abilities directly or to benchmark its performance on standard, unaltered tasks.
Stars
26
Forks
2
Language
Python
License
MIT
Category
Last pushed
Dec 11, 2024
Commits (30d)
0
Get this data via API
curl "https://pt-edge.onrender.com/api/v1/quality/transformers/bowen-upenn/llm_token_bias"
Open to everyone — 100 requests/day, no key needed. Get a free key for 1,000/day.
Higher-rated alternatives
google-deepmind/long-form-factuality
Benchmarking long-form factuality in large language models. Original code for our paper...
gnai-creator/aletheion-llm-v2
Decoder-only LLM with integrated epistemic tomography. Knows what it doesn't know.
sandylaker/ib-edl
Calibrating LLMs with Information-Theoretic Evidential Deep Learning (ICLR 2025)
nightdessert/Retrieval_Head
open-source code for paper: Retrieval Head Mechanistically Explains Long-Context Factuality
MLD3/steerability
An open-source evaluation framework for measuring LLM steerability.