How GPT learns layer by layer

• URL: http://arxiv.org/abs/2501.07108v1
• Date: Mon, 13 Jan 2025 07:42:55 GMT
• Title: How GPT learns layer by layer
• Authors: Jason Du, Kelly Hong, Alishba Imran, Erfan Jahanparast, Mehdi Khfifi, Kaichun Qiao,
• Abstract summary: We analyze OthelloGPT, a GPT-based model trained on Othello gameplay, as a testbed for studying representation learning.<n>We compare Sparse Autoencoders (SAEs) with linear probes, finding that SAEs offer more robust, disentangled insights into compositional features.<n>We use SAEs to decode features related to tile color and tile stability, a previously unexamined feature that reflects complex gameplay concepts.
• Score: 0.28926166547031595
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Large Language Models (LLMs) excel at tasks like language processing, strategy games, and reasoning but struggle to build generalizable internal representations essential for adaptive decision-making in agents. For agents to effectively navigate complex environments, they must construct reliable world models. While LLMs perform well on specific benchmarks, they often fail to generalize, leading to brittle representations that limit their real-world effectiveness. Understanding how LLMs build internal world models is key to developing agents capable of consistent, adaptive behavior across tasks. We analyze OthelloGPT, a GPT-based model trained on Othello gameplay, as a controlled testbed for studying representation learning. Despite being trained solely on next-token prediction with random valid moves, OthelloGPT shows meaningful layer-wise progression in understanding board state and gameplay. Early layers capture static attributes like board edges, while deeper layers reflect dynamic tile changes. To interpret these representations, we compare Sparse Autoencoders (SAEs) with linear probes, finding that SAEs offer more robust, disentangled insights into compositional features, whereas linear probes mainly detect features useful for classification. We use SAEs to decode features related to tile color and tile stability, a previously unexamined feature that reflects complex gameplay concepts like board control and long-term planning. We study the progression of linear probe accuracy and tile color using both SAE's and linear probes to compare their effectiveness at capturing what the model is learning. Although we begin with a smaller language model, OthelloGPT, this study establishes a framework for understanding the internal representations learned by GPT models, transformers, and LLMs more broadly. Our code is publicly available: https://github.com/ALT-JS/OthelloSAE.

Related papers

• If an LLM Were a Character, Would It Know Its Own Story? Evaluating Lifelong Learning in LLMs [55.8331366739144]
  We introduce LIFESTATE-BENCH, a benchmark designed to assess lifelong learning in large language models (LLMs) Our fact checking evaluation probes models' self-awareness, episodic memory retrieval, and relationship tracking, across both parametric and non-parametric approaches.
  arXiv  Detail & Related papers  (2025-03-30T16:50:57Z)
• LLM Knows Geometry Better than Algebra: Numerical Understanding of LLM-Based Agents in A Trading Arena [44.13836431002364]
  We design a virtual numerical game simulating complex economic systems through zero-sum games, where agents invest in stock portfolios. Our experiments reveal that LLMs, including GPT-4o, struggle with algebraic reasoning when dealing with plain-text stock data, often focusing on local details rather than global trends. In contrast, LLMs perform significantly better with geometric reasoning when presented with visual data, such as scatter plots or K-line charts.
  arXiv  Detail & Related papers  (2025-02-25T08:41:01Z)
• Predicting the Performance of Black-box LLMs through Self-Queries [60.87193950962585]
  Large language models (LLMs) are increasingly relied on in AI systems, predicting when they make mistakes is crucial.<n>In this paper, we extract features of LLMs in a black-box manner by using follow-up prompts and taking the probabilities of different responses as representations.<n>We demonstrate that training a linear model on these low-dimensional representations produces reliable predictors of model performance at the instance level.
  arXiv  Detail & Related papers  (2025-01-02T22:26:54Z)
• SPaR: Self-Play with Tree-Search Refinement to Improve Instruction-Following in Large Language Models [88.29990536278167]
  We introduce SPaR, a self-play framework integrating tree-search self-refinement to yield valid and comparable preference pairs free from distractions.<n>Our experiments show that a LLaMA3-8B model, trained over three iterations guided by SPaR, surpasses GPT-4-Turbo on the IFEval benchmark without losing general capabilities.
  arXiv  Detail & Related papers  (2024-12-16T09:47:43Z)
• ST-LLM: Large Language Models Are Effective Temporal Learners [58.79456373423189]
  Large Language Models (LLMs) have showcased impressive capabilities in text comprehension and generation. How to effectively encode and understand videos in video-based dialogue systems remains to be solved. We propose ST-LLM, an effective video-LLM baseline with spatial-temporal sequence modeling inside LLM.
  arXiv  Detail & Related papers  (2024-03-30T10:11:26Z)
• RecExplainer: Aligning Large Language Models for Explaining Recommendation Models [50.74181089742969]
  Large language models (LLMs) have demonstrated remarkable intelligence in understanding, reasoning, and instruction following. This paper presents the initial exploration of using LLMs as surrogate models to explain black-box recommender models. To facilitate an effective alignment, we introduce three methods: behavior alignment, intention alignment, and hybrid alignment.
  arXiv  Detail & Related papers  (2023-11-18T03:05:43Z)
• Implicit meta-learning may lead language models to trust more reliable sources [9.073765860925395]
  We introduce random strings ("tags") as indicators of usefulness in a synthetic fine-tuning dataset. Fine-tuning on this dataset leads to implicit meta-learning (IML) We reflect on what our results might imply about capabilities, risks, and controllability of future AI systems.
  arXiv  Detail & Related papers  (2023-10-23T15:50:08Z)
• LLM-augmented Preference Learning from Natural Language [19.700169351688768]
  Large Language Models (LLMs) are equipped to deal with larger context lengths. LLMs can consistently outperform the SotA when the target text is large. Few-shot learning yields better performance than zero-shot learning.
  arXiv  Detail & Related papers  (2023-10-12T17:17:27Z)
• From Language Modeling to Instruction Following: Understanding the Behavior Shift in LLMs after Instruction Tuning [63.63840740526497]
  We investigate how instruction tuning adjusts pre-trained models with a focus on intrinsic changes. The impact of instruction tuning is then studied by comparing the explanations derived from the pre-trained and instruction-tuned models. Our findings reveal three significant impacts of instruction tuning.
  arXiv  Detail & Related papers  (2023-09-30T21:16:05Z)
• Evaluating and Explaining Large Language Models for Code Using Syntactic Structures [74.93762031957883]
  This paper introduces ASTxplainer, an explainability method specific to Large Language Models for code. At its core, ASTxplainer provides an automated method for aligning token predictions with AST nodes. We perform an empirical evaluation on 12 popular LLMs for code using a curated dataset of the most popular GitHub projects.
  arXiv  Detail & Related papers  (2023-08-07T18:50:57Z)
• Don't Judge a Language Model by Its Last Layer: Contrastive Learning with Layer-Wise Attention Pooling [6.501126898523172]
  Recent pre-trained language models (PLMs) achieved great success on many natural language processing tasks through learning linguistic features and contextualized sentence representation. This paper introduces the attention-based pooling strategy, which enables the model to preserve layer-wise signals captured in each layer and learn digested linguistic features for downstream tasks.
  arXiv  Detail & Related papers  (2022-09-13T13:09:49Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.