Distributional Associations vs In-Context Reasoning: A Study of Feed-forward and Attention Layers

• URL: http://arxiv.org/abs/2406.03068v2
• Date: Thu, 06 Mar 2025 23:55:51 GMT
• Title: Distributional Associations vs In-Context Reasoning: A Study of Feed-forward and Attention Layers
• Authors: Lei Chen, Joan Bruna, Alberto Bietti,
• Abstract summary: We study the distinction between feed-forward and attention layers in large language models.<n>We find that feed-forward layers tend to learn simple distributional associations such as bigrams, while attention layers focus on in-context reasoning.
• Score: 49.80959223722325
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Large language models have been successful at tasks involving basic forms of in-context reasoning, such as generating coherent language, as well as storing vast amounts of knowledge. At the core of the Transformer architecture behind such models are feed-forward and attention layers, which are often associated to knowledge and reasoning, respectively. In this paper, we study this distinction empirically and theoretically in a controlled synthetic setting where certain next-token predictions involve both distributional and in-context information. We find that feed-forward layers tend to learn simple distributional associations such as bigrams, while attention layers focus on in-context reasoning. Our theoretical analysis identifies the noise in the gradients as a key factor behind this discrepancy. Finally, we illustrate how similar disparities emerge in pre-trained models through ablations on the Pythia model family on simple reasoning tasks.

Related papers

• Toward Understanding In-context vs. In-weight Learning [50.24035812301655]
  We identify simplified distributional properties that give rise to the emergence and disappearance of in-context learning. We then extend the study to a full large language model, showing how fine-tuning on various collections of natural language prompts can elicit similar in-context and in-weight learning behaviour.
  arXiv  Detail & Related papers  (2024-10-30T14:09:00Z)
• The Foundations of Tokenization: Statistical and Computational Concerns [51.370165245628975]
  Tokenization is a critical step in the NLP pipeline. Despite its recognized importance as a standard representation method in NLP, the theoretical underpinnings of tokenization are not yet fully understood. The present paper contributes to addressing this theoretical gap by proposing a unified formal framework for representing and analyzing tokenizer models.
  arXiv  Detail & Related papers  (2024-07-16T11:12:28Z)
• Fine-Grained Causal Dynamics Learning with Quantization for Improving Robustness in Reinforcement Learning [26.34622544479565]
  Causal dynamics learning is a promising approach to enhancing robustness in reinforcement learning. We propose a novel model that infers fine-grained causal structures and employs them for prediction.
  arXiv  Detail & Related papers  (2024-06-05T13:13:58Z)
• Mitigating Reversal Curse in Large Language Models via Semantic-aware Permutation Training [57.771940716189114]
  We show that large language models (LLMs) suffer from the "reversal curse" The root cause of the reversal curse lies in the different word order between the training and inference stage. We propose Semantic-aware Permutation Training (SPT) to address this issue.
  arXiv  Detail & Related papers  (2024-03-01T18:55:20Z)
• Predictive Churn with the Set of Good Models [61.00058053669447]
  This paper explores connections between two seemingly unrelated concepts of predictive inconsistency. The first, known as predictive multiplicity, occurs when models that perform similarly produce conflicting predictions for individual samples. The second concept, predictive churn, examines the differences in individual predictions before and after model updates.
  arXiv  Detail & Related papers  (2024-02-12T16:15:25Z)
• The mechanistic basis of data dependence and abrupt learning in an in-context classification task [0.3626013617212666]
  We show that specific distributional properties inherent in language control the trade-off or simultaneous appearance of two forms of learning. In-context learning is driven by the abrupt emergence of an induction head, which subsequently competes with in-weights learning. We propose that the sharp transitions in attention-based networks arise due to a specific chain of multi-layer operations necessary to achieve ICL.
  arXiv  Detail & Related papers  (2023-12-03T20:53:41Z)
• Modeling Hierarchical Reasoning Chains by Linking Discourse Units and Key Phrases for Reading Comprehension [80.99865844249106]
  We propose a holistic graph network (HGN) which deals with context at both discourse level and word level, as the basis for logical reasoning. Specifically, node-level and type-level relations, which can be interpreted as bridges in the reasoning process, are modeled by a hierarchical interaction mechanism.
  arXiv  Detail & Related papers  (2023-06-21T07:34:27Z)
• Token-wise Decomposition of Autoregressive Language Model Hidden States for Analyzing Model Predictions [9.909170013118775]
  This work presents a linear decomposition of final hidden states from autoregressive language models based on each initial input token. Using the change in next-word probability as a measure of importance, this work first examines which context words make the biggest contribution to language model predictions.
  arXiv  Detail & Related papers  (2023-05-17T23:55:32Z)
• A Theory of Emergent In-Context Learning as Implicit Structure Induction [8.17811111226145]
  Scaling large language models leads to an emergent capacity to learn in-context from example demonstrations. We argue that in-context learning relies on recombination of compositional operations found in natural language data. We show how in-context learning is supported by a representation of the input's compositional structure.
  arXiv  Detail & Related papers  (2023-03-14T15:24:05Z)
• APOLLO: A Simple Approach for Adaptive Pretraining of Language Models for Logical Reasoning [73.3035118224719]
  We propose APOLLO, an adaptively pretrained language model that has improved logical reasoning abilities. APOLLO performs comparably on ReClor and outperforms baselines on LogiQA.
  arXiv  Detail & Related papers  (2022-12-19T07:40:02Z)
• ALERT: Adapting Language Models to Reasoning Tasks [43.8679673685468]
  ALERT is a benchmark and suite of analyses for assessing language models' reasoning ability. ALERT provides a test bed to asses any language model on fine-grained reasoning skills. We find that language models learn more reasoning skills during finetuning stage compared to pretraining state.
  arXiv  Detail & Related papers  (2022-12-16T05:15:41Z)
• Disentangling Reasoning Capabilities from Language Models with Compositional Reasoning Transformers [72.04044221898059]
  ReasonFormer is a unified reasoning framework for mirroring the modular and compositional reasoning process of humans. The representation module (automatic thinking) and reasoning modules (controlled thinking) are disentangled to capture different levels of cognition. The unified reasoning framework solves multiple tasks with a single model,and is trained and inferred in an end-to-end manner.
  arXiv  Detail & Related papers  (2022-10-20T13:39:55Z)
• Testing Pre-trained Language Models' Understanding of Distributivity via Causal Mediation Analysis [13.07356367140208]
  We introduce DistNLI, a new diagnostic dataset for natural language inference. We find that the extent of models' understanding is associated with model size and vocabulary size.
  arXiv  Detail & Related papers  (2022-09-11T00:33:28Z)
• Turning Tables: Generating Examples from Semi-structured Tables for Endowing Language Models with Reasoning Skills [32.55545292360155]
  We propose to leverage semi-structured tables, and automatically generate at scale question-paragraph pairs. We add a pre-training step over this synthetic data, which includes examples that require 16 different reasoning skills. We show that our model, PReasM, substantially outperforms T5, a popular pre-trained encoder-decoder model.
  arXiv  Detail & Related papers  (2021-07-15T11:37:14Z)
• On the Interplay Between Fine-tuning and Composition in Transformers [7.513100214864645]
  We investigate the impact of fine-tuning on the capacity of contextualized embeddings to capture phrase meaning information. Specifically, we fine-tune models on an adversarial paraphrase classification task with high lexical overlap, and on a sentiment classification task. We find that fine-tuning largely fails to benefit compositionality in these representations, though training on sentiment yields a small, localized benefit for certain models.
  arXiv  Detail & Related papers  (2021-05-31T01:49:56Z)
• Grounded Compositional Outputs for Adaptive Language Modeling [59.02706635250856]
  A language model's vocabulary$-$typically selected before training and permanently fixed later$-$affects its size. We propose a fully compositional output embedding layer for language models. To our knowledge, the result is the first word-level language model with a size that does not depend on the training vocabulary.
  arXiv  Detail & Related papers  (2020-09-24T07:21:14Z)
• Leap-Of-Thought: Teaching Pre-Trained Models to Systematically Reason Over Implicit Knowledge [96.92252296244233]
  Large pre-trained language models (LMs) acquire some reasoning capacity, but this ability is difficult to control. We show that LMs can be trained to reliably perform systematic reasoning combining both implicit, pre-trained knowledge and explicit natural language statements. Our work paves a path towards open-domain systems that constantly improve by interacting with users who can instantly correct a model by adding simple natural language statements.
  arXiv  Detail & Related papers  (2020-06-11T17:02:20Z)
• Pairwise Supervision Can Provably Elicit a Decision Boundary [84.58020117487898]
  Similarity learning is a problem to elicit useful representations by predicting the relationship between a pair of patterns. We show that similarity learning is capable of solving binary classification by directly eliciting a decision boundary.
  arXiv  Detail & Related papers  (2020-06-11T05:35:16Z)

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.