Andrew Lampinen
@lampinen.bsky.social
Interested in cognition and artificial intelligence. Research Scientist at Google DeepMind. Previously cognitive science at Stanford. Posts are mine.
lampinen.github.io
lampinen.github.io
Apologies for being quiet on here lately — been focusing on the more important things in life :)
November 9, 2025 at 11:34 PM
Apologies for being quiet on here lately — been focusing on the more important things in life :)
We show that even when models generalize well from parametric learning in standard (nontrivial) evaluations, there are selective, consistent failures of latent learning. Only models with retrieval generalize well on the key tests of latent learning. 6/
September 22, 2025 at 4:21 AM
We show that even when models generalize well from parametric learning in standard (nontrivial) evaluations, there are selective, consistent failures of latent learning. Only models with retrieval generalize well on the key tests of latent learning. 6/
To illustrate this point, we explore latent learning across a wide range of benchmarks (from codebook translation to BC and RL navigation) — and compare baseline language models or agents to those equipped with oracle retrieval. 5/
September 22, 2025 at 4:21 AM
To illustrate this point, we explore latent learning across a wide range of benchmarks (from codebook translation to BC and RL navigation) — and compare baseline language models or agents to those equipped with oracle retrieval. 5/
But models can readily use latent information in their context. We therefore suggest that natural intelligence solves the latent learning problem via the complementary strengths of episodic memory: reinstating experiences into context makes latent information accessible. 4/
September 22, 2025 at 4:21 AM
But models can readily use latent information in their context. We therefore suggest that natural intelligence solves the latent learning problem via the complementary strengths of episodic memory: reinstating experiences into context makes latent information accessible. 4/
we argue that parametric learning methods are too tied to the explicit training task, and fail to effectively encode latent information relevant to possible future tasks, and we suggest that this explains a wide range of findings, from navigation to the reversal curse. 3/
September 22, 2025 at 4:21 AM
we argue that parametric learning methods are too tied to the explicit training task, and fail to effectively encode latent information relevant to possible future tasks, and we suggest that this explains a wide range of findings, from navigation to the reversal curse. 3/
When we've compared these in past work e.g. Supplement fig. A.13 here proceedings.neurips.cc/paper/2020/h... we've seen pretty similar results between the two. Haven't run it in exactly this setting though. There are also some arguments that 1/2
August 5, 2025 at 8:18 PM
When we've compared these in past work e.g. Supplement fig. A.13 here proceedings.neurips.cc/paper/2020/h... we've seen pretty similar results between the two. Haven't run it in exactly this setting though. There are also some arguments that 1/2
I don't know of any reviews unfortunately! Fig. 16 in our TMLR paper (openreview.net/forum?id=aY2...) shows an instance — training classifiers on the penultimate reps to decode a label predicted by both easy and hard features; at high predictivity the classifier prefers the easy feature, even 1/2
August 5, 2025 at 6:28 PM
I don't know of any reviews unfortunately! Fig. 16 in our TMLR paper (openreview.net/forum?id=aY2...) shows an instance — training classifiers on the penultimate reps to decode a label predicted by both easy and hard features; at high predictivity the classifier prefers the easy feature, even 1/2
We also present a worst-case study I find conceptually interesting: homomorphic encryption. It’s possible to do systematic computation over representations whose content is always encrypted, and thus difficult to decode by design!
August 5, 2025 at 2:36 PM
We also present a worst-case study I find conceptually interesting: homomorphic encryption. It’s possible to do systematic computation over representations whose content is always encrypted, and thus difficult to decode by design!
We briefly discuss (some of) the origins of these biases — they are driven by both learning dynamics and the fact that there are in some sense a larger variety of “natural” ways to represent a nonlinear feature.
August 5, 2025 at 2:36 PM
We briefly discuss (some of) the origins of these biases — they are driven by both learning dynamics and the fact that there are in some sense a larger variety of “natural” ways to represent a nonlinear feature.
These biases can lead to dramatic downstream effects that cause unexpected conclusions from analyses. For example, RSA may identify two models computing the same, complex task as much less representationally-similar than either of them is to a model computing a much simpler task (right panel)!
August 5, 2025 at 2:36 PM
These biases can lead to dramatic downstream effects that cause unexpected conclusions from analyses. For example, RSA may identify two models computing the same, complex task as much less representationally-similar than either of them is to a model computing a much simpler task (right panel)!
Representations were systematically biased towards certain kinds of features. For example, a model reliably computing easy (linear) and hard (nonlinear) features has 55% repr. variance explained by the easy one, 5% by the hard, with similar biases in top PCs and individual units.
August 5, 2025 at 2:36 PM
Representations were systematically biased towards certain kinds of features. For example, a model reliably computing easy (linear) and hard (nonlinear) features has 55% repr. variance explained by the easy one, 5% by the hard, with similar biases in top PCs and individual units.
We constructed controlled datasets with many input features, and trained deep learning models to compute functions of those features (e.g. linear ones like identifying a feature, or nonlinear ones like XOR). We then analyzed the patterns of representational activity they learned.
August 5, 2025 at 2:36 PM
We constructed controlled datasets with many input features, and trained deep learning models to compute functions of those features (e.g. linear ones like identifying a feature, or nonlinear ones like XOR). We then analyzed the patterns of representational activity they learned.
In neuroscience, we often try to understand systems by analyzing their representations — using tools like regression or RSA. But are these analyses biased towards discovering a subset of what a system represents? If you're interested in this question, check out our new commentary! Thread:
![What do representations tell us about a system? Image of a mouse with a scope showing a vector of activity patterns, and a neural network with a vector of unit activity patterns
Common analyses of neural representations: Encoding models (relating activity to task features) drawing of an arrow from a trace saying [on_____on____] to a neuron and spike train. Comparing models via neural predictivity: comparing two neural networks by their R^2 to mouse brain activity. RSA: assessing brain-brain or model-brain correspondence using representational dissimilarity matrices](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:e6ewzleebkdi2y2bxhjxoknt/bafkreiav2io2ska33o4kizf57co5bboqyyfdpnozo2gxsicrfr5l7qzjcq@jpeg)
August 5, 2025 at 2:36 PM
In neuroscience, we often try to understand systems by analyzing their representations — using tools like regression or RSA. But are these analyses biased towards discovering a subset of what a system represents? If you're interested in this question, check out our new commentary! Thread:
Looking forward to attending CogSci this week! I'll be giving a talk (see below) at the Reasoning Across Minds and Machines workshop on Wednesday at 10:25 AM, and will be around most of the week — feel free to reach out if you'd like to meet up!
July 28, 2025 at 6:07 PM
Looking forward to attending CogSci this week! I'll be giving a talk (see below) at the Reasoning Across Minds and Machines workshop on Wednesday at 10:25 AM, and will be around most of the week — feel free to reach out if you'd like to meet up!
We argued this was necessary even for formal domains like mathematics, which are fundamentally about the insight behind the logic — as mathematicians have long pointed out.
![Box 2 | Syntax & meaning in mathematics
Newell & Simon motivated the physical symbol system hypothesis by this claim: “Logic, and by
incorporation all of mathematics, was a game played with meaningless tokens according to certain
purely syntactic rules. All meaning had been purged.” [1]. In the reasoning systems they created,
symbols thus became synonymous with discrete entities, symbolic processing entailed manipulating
with formal algebras, and the properties of our formal symbolic systems were inherited by GOFAI.
But, is this how humans perform logical reasoning? Fields medalist Paul Cohen responded directly: “I
never was able to successfully analyze proofs as a combinatorial ‘game’ played with symbols on paper,”
instead, to reason productively “one must essentially forget that all proofs are eventually transcribed
in this formal language” [77]. Mac Lane [78] explains why: “strict formalism can’t explain which of
many formulas matter [...] the choice of form is determined by ideas and experience.” The meaning
of symbols and their relationships gives mathematicians intuitions for why a theorem might be true,
and thereby ideas of how to prove it (as do examples; 79). Furthermore, a major goal of theories and
proofs is to convey understanding. Thus, mathematicians are often unsatisfied with a formal proof of
a theorem that does not clearly communicate understanding of the underlying ideas [36, 80]. The
meaning behind symbols is the essence of mathematics, not a burden to be discarded.
McClelland et al. [65] draw on similar arguments to suggest that neural network models are uniquely
well-posed to capture the meaning-laden processes of human mathematical reasoning](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:e6ewzleebkdi2y2bxhjxoknt/bafkreiclp7zqka5jr6his2jjazoenm7yzkiici3zi6wmbt2jq6lpldeb5q@jpeg)
July 21, 2025 at 10:20 PM
We argued this was necessary even for formal domains like mathematics, which are fundamentally about the insight behind the logic — as mathematicians have long pointed out.
Augmented finetuning (green bars above) substantially outperforms other methods. We then test a larger dataset of thousands of documents generated from an underlying semantic structure, and still see strong benefits from ICL and ICL-augmented finetuning (below). 5/
May 2, 2025 at 5:02 PM
Augmented finetuning (green bars above) substantially outperforms other methods. We then test a larger dataset of thousands of documents generated from an underlying semantic structure, and still see strong benefits from ICL and ICL-augmented finetuning (below). 5/
But it’s not just reversals; ICL consistently generalizes better than finetuning in areas like syllogistic deduction too. Motivated by these findings, we propose a method to improve finetuning generalization: prompt language models to augment the data using ICL! 4/
May 2, 2025 at 5:02 PM
But it’s not just reversals; ICL consistently generalizes better than finetuning in areas like syllogistic deduction too. Motivated by these findings, we propose a method to improve finetuning generalization: prompt language models to augment the data using ICL! 4/
Across various datasets, we find that ICL generalizes much better to certain kinds of tests than finetuning. For example, in the setting of the Reversal Curse dataset, just putting the whole dataset in context achieves almost 100% generalization! 3/
May 2, 2025 at 5:02 PM
Across various datasets, we find that ICL generalizes much better to certain kinds of tests than finetuning. For example, in the setting of the Reversal Curse dataset, just putting the whole dataset in context achieves almost 100% generalization! 3/
We use controlled experiments to explore the generalization of ICL and finetuning in data-matched settings; if we have some documents containing new knowledge, does the LM generalize better from finetuning on them, or just putting all of them in context? 2/
May 2, 2025 at 5:02 PM
We use controlled experiments to explore the generalization of ICL and finetuning in data-matched settings; if we have some documents containing new knowledge, does the LM generalize better from finetuning on them, or just putting all of them in context? 2/
I trained models with different spurious predictivity in training. If the spurious feature is absent, the model fails to generalize. But it turns out there's a "sweet spot" where the model learns the spurious feature, then uses it to learn the harder one, and thus generalizes!
May 1, 2025 at 12:32 AM
I trained models with different spurious predictivity in training. If the spurious feature is absent, the model fails to generalize. But it turns out there's a "sweet spot" where the model learns the spurious feature, then uses it to learn the harder one, and thus generalizes!
So I made a simpler analog: a toy dataset where a transformer is trained to predict a hard relational feature (whether two cued tokens match; like syntactic agreement) and a simpler spurious feature (identities of the cued tokens). In test, the spurious feature is always useless.
May 1, 2025 at 12:32 AM
So I made a simpler analog: a toy dataset where a transformer is trained to predict a hard relational feature (whether two cued tokens match; like syntactic agreement) and a simpler spurious feature (identities of the cued tokens). In test, the spurious feature is always useless.
This is loosely inspired by the idea that, in a complex sentence with multiple noun-verb relations, a system could use either the syntax (a difficult, relational feature), or the semantics (easy, but spurious and unreliable) to figure out which nouns go with which verbs.
May 1, 2025 at 12:32 AM
This is loosely inspired by the idea that, in a complex sentence with multiple noun-verb relations, a system could use either the syntax (a difficult, relational feature), or the semantics (easy, but spurious and unreliable) to figure out which nouns go with which verbs.
and deriving reductive explanations from our models using tools like studying how data properties alter the behaviors & mechanisms models learn, as a kind of rational analysis, and studying model computations; these can build towards formal, normative theories.
February 28, 2025 at 5:14 PM
and deriving reductive explanations from our models using tools like studying how data properties alter the behaviors & mechanisms models learn, as a kind of rational analysis, and studying model computations; these can build towards formal, normative theories.
Finally, we turn to how complex naturalistic experimental paradigms and opaque models can lead to conceptual understanding, through experiments that augment the multidimensional variation of natural stimuli with theory-driven parametric manipulations…
February 28, 2025 at 5:14 PM
Finally, we turn to how complex naturalistic experimental paradigms and opaque models can lead to conceptual understanding, through experiments that augment the multidimensional variation of natural stimuli with theory-driven parametric manipulations…
We then move to practicalities of building generalizable models that scale to naturalistic settings — and draw concrete recommendations from AI, including the benefits of frictionless reproducibility and dynamic, cumulative benchmarks.
February 28, 2025 at 5:14 PM
We then move to practicalities of building generalizable models that scale to naturalistic settings — and draw concrete recommendations from AI, including the benefits of frictionless reproducibility and dynamic, cumulative benchmarks.