REVIEW 8 cited by
Pretraining Data Mixtures Enable Narrow Model Selection Capabilities in Transformer Models
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
abstract
Transformer models, notably large language models (LLMs), have the remarkable ability to perform in-context learning (ICL) -- to perform new tasks when prompted with unseen input-output examples without any explicit model training. In this work, we study how effectively transformers can bridge between their pretraining data mixture, comprised of multiple distinct task families, to identify and learn new tasks in-context which are both inside and outside the pretraining distribution. Building on previous work, we investigate this question in a controlled setting, where we study transformer models trained on sequences of $(x, f(x))$ pairs rather than natural language. Our empirical results show transformers demonstrate near-optimal unsupervised model selection capabilities, in their ability to first in-context identify different task families and in-context learn within them when the task families are well-represented in their pretraining data. However when presented with tasks or functions which are out-of-domain of their pretraining data, we demonstrate various failure modes of transformers and degradation of their generalization for even simple extrapolation tasks. Together our results highlight that the impressive ICL abilities of high-capacity sequence models may be more closely tied to the coverage of their pretraining data mixtures than inductive biases that create fundamental generalization capabilities.
Forward citations
Cited by 8 Pith papers
-
Induction Heads Interpolate N-Grams
Induction-head circuits implement soft context-matching (Jelinek–Mercer-style interpolation over partial matches) plus BOS-induced Dirichlet pseudo-counts, and trained transformers recover both mechanisms.
-
How Context Attribution Handles What the Model Already Knows
Context attribution methods cannot disentangle in-context from in-weight knowledge and assign unfaithful scores under overlap; new metrics and WMDP-Cyber++ quantify the failure.
-
Dissecting Multimodal In-Context Learning: Modality Asymmetries and Circuit Dynamics in modern Transformers
In a two-modality transformer, a primary-modality pretraining stage installs an induction circuit, so the secondary modality needs only low class diversity to learn in-context from examples.
-
How Does the Pretraining Distribution Shape In-Context Learning? A Fundamental Trade-Off
Heavy-tailed pretraining distributions improve in-context task selection under distribution shift but worsen ICL generalization, especially in low-data regimes.
-
Train Once, Reuse Everywhere: Generalizable Implicit In-Context Learning by Routing Attention
ICR extracts shared attention directions from in-context learning and routes them at inference time, enabling zero-shot reuse across tasks.
-
Selective Induction Heads: How Transformers Select Causal Structures In Context
Transformers can learn to select the correct lag of an interleaved Markov chain in context via a circuit the authors call a selective induction head, whose asymptotic optimality proof is incomplete.
-
Meta-Learning Approaches for Speaker-Dependent Voice Fatigue Models
Meta-learning, especially a transformer-based sequence model, outperforms cross-sectional and mixed-effects baselines for predicting time since sleep from speech, though the evaluation protocol may overstate deploymen...
-
Reverse Convolution and Its Applications to Image Restoration
The abstract and body of this submission are two unrelated papers; the reverse-convolution claims appear nowhere in the full text.
Discussion (0). Continue with ORCID to comment.