))
Discovering Latent Knowledge in Language Models Without Supervision
Manage episode 424087972 series 3498845
Contenu fourni par BlueDot Impact. Tout le contenu du podcast, y compris les épisodes, les graphiques et les descriptions de podcast, est téléchargé et fourni directement par BlueDot Impact ou son partenaire de plateforme de podcast. Si vous pensez que quelqu'un utilise votre œuvre protégée sans votre autorisation, vous pouvez suivre le processus décrit ici https://fr.player.fm/legal.
Abstract:
Existing techniques for training language models can be misaligned with the truth: if we train models with imitation learning, they may reproduce errors that humans make; if we train them to generate text that humans rate highly, they may output errors that human evaluators can't detect. We propose circumventing this issue by directly finding latent knowledge inside the internal activations of a language model in a purely unsupervised way. Specifically, we introduce a method for accurately answering yes-no questions given only unlabeled model activations. It works by finding a direction in activation space that satisfies logical consistency properties, such as that a statement and its negation have opposite truth values. We show that despite using no supervision and no model outputs, our method can recover diverse knowledge represented in large language models: across 6 models and 10 question-answering datasets, it outperforms zero-shot accuracy by 4\\% on average. We also find that it cuts prompt sensitivity in half and continues to maintain high accuracy even when models are prompted to generate incorrect answers. Our results provide an initial step toward discovering what language models know, distinct from what they say, even when we don't have access to explicit ground truth labels.
Original text:
https://arxiv.org/abs/2212.03827
Narrated for AI Safety Fundamentals by Perrin Walker of TYPE III AUDIO.
---
A podcast by BlueDot Impact.
Learn more on the AI Safety Fundamentals website.
Chapitres
1. Discovering Latent Knowledge in Language Models Without Supervision (00:00:00)
2. ABSTRACT (00:00:12)
3. 1 INTRODUCTION (00:01:29)
4. 2 PROBLEM STATEMENT AND FRAMEWORK (00:06:32)
5. 2.1 PROBLEM: DISCOVERING LATENT KNOWLEDGE (00:07:04)
6. 2.2 METHOD: CONTRAST-CONSISTENT SEARCH (00:08:31)
7. Constructing contrast pairs. (00:10:16)
8. Feature extraction and normalization. (00:11:43)
9. Inference. (00:15:58)
10. 3 RESULTS (00:17:04)
11. 3.1 EXPERIMENTAL SETUP (00:17:07)
12. 3.2 EVALUATING CCS (00:23:41)
13. 3.2.1 CCS OUTPERFORMS ZERO-SHOT (00:23:44)
14. 3.2.2 CCS IS ROBUST TO MISLEADING PROMPTS (00:25:17)
15. 3.3 ANALYZING CCS (00:26:41)
16. 3.3.1 CCS FINDS A TASK-AGNOSTIC REPRESENTATION OF TRUTH (00:27:12)
17. 3.3.2 CCS DOES NOT JUST RECOVER MODEL OUTPUTS (00:30:00)
18. 3.3.3 TRUTH IS A SALIENT FEATURE (00:31:45)
85 episodes
Partager
