Eghbal A. Hosseini

Eghbal A. Hosseini, Yuxuan Li, Yasaman Bahri et al. · deepmind, stanford

Large Language Models (LLMs) have been shown to organize the representations of input sequences into straighter neural trajectories in their deep layers, which has been hypothesized to facilitate next-token prediction via linear extrapolation. Language models can also adapt to diverse tasks and learn new structure in context, and recent work has shown that this in-context learning (ICL) can be reflected in representational changes. Here we bring these two lines of research together to explore whether representation straightening occurs \emph{within} a context during ICL. We measure representational straightening in Gemma 2 models across a diverse set of in-context tasks, and uncover a dichotomy in how LLMs' representations change in context. In continual prediction settings (e.g., natural language, grid world traversal tasks) we observe that increasing context increases the straightness of neural sequence trajectories, which is correlated with improvement in model prediction. Conversely, in structured prediction settings (e.g., few-shot tasks), straightening is inconsistent -- it is only present in phases of the task with explicit structure (e.g., repeating a template), but vanishes elsewhere. These results suggest that ICL is not a monolithic process. Instead, we propose that LLMs function like a Swiss Army knife: depending on task structure, the LLM dynamically selects between strategies, only some of which yield representational straightening.

Eghbal A. Hosseini, Evelina Fedorenko

Predicting upcoming events is critical to our ability to interact with our environment. Transformer models, trained on next-word prediction, appear to construct representations of linguistic input that can support diverse downstream tasks. But how does a predictive objective shape such representations? Inspired by recent work in vision (Henaff et al., 2019), we test a hypothesis about predictive representations of autoregressive transformers. In particular, we test whether the neural trajectory of a sentence becomes progressively straighter as it passes through the network layers. The key insight is that straighter trajectories should facilitate prediction via linear extrapolation. We quantify straightness using a 1-dimensional curvature metric, and present four findings in support of the trajectory straightening hypothesis: i) In trained models, the curvature decreases from the early to the deeper layers of the network. ii) Models that perform better on the next-word prediction objective exhibit greater decreases in curvature, suggesting that this improved ability to straighten sentence trajectories may be the driver of better language modeling performance. iii) Given the same linguistic context, the sequences that are generated by the model have lower curvature than the actual continuations observed in a language corpus, suggesting that the model favors straighter trajectories for making predictions. iv) A consistent relationship holds between the average curvature and the average surprisal of sentences in the deep model layers, such that sentences with straighter trajectories also have lower surprisal. Importantly, untrained models do not exhibit these behaviors. In tandem, these results support the trajectory straightening hypothesis and provide a possible mechanism for how the geometry of the internal representations of autoregressive models supports next word prediction.

Andrew Kyle Lampinen, Yuxuan Li, Eghbal Hosseini et al.

Language model representations often contain linear directions that correspond to high-level concepts. Here, we study the dynamics of these representations: how representations evolve along these dimensions within the context of (simulated) conversations. We find that linear representations can change dramatically over a conversation; for example, information that is represented as factual at the beginning of a conversation can be represented as non-factual at the end and vice versa. These changes are content-dependent; while representations of conversation-relevant information may change, generic information is generally preserved. These changes are robust even for dimensions that disentangle factuality from more superficial response patterns, and occur across different model families and layers of the model. These representation changes do not require on-policy conversations; even replaying a conversation script written by an entirely different model can produce similar changes. However, adaptation is much weaker from simply having a sci-fi story in context that is framed more explicitly as such. We also show that steering along a representational direction can have dramatically different effects at different points in a conversation. These results are consistent with the idea that representations may evolve in response to the model playing a particular role that is cued by a conversation. Our findings may pose challenges for interpretability and steering -- in particular, they imply that it may be misleading to use static interpretations of features or directions, or probes that assume a particular range of features consistently corresponds to a particular ground-truth value. However, these types of representational dynamics also point to exciting new research directions for understanding how models adapt to context.

Eghbal A. Hosseini, Brian Cheung, Evelina Fedorenko et al.

Neural networks exhibit a remarkable degree of representational convergence across diverse architectures, training objectives, and even data modalities. This convergence is predictive of alignment with brain representation. A recent hypothesis suggests this arises from learning the underlying structure in the environment in similar ways. However, it is unclear how individual stimuli elicit convergent representations across networks. An image can be perceived in multiple ways and expressed differently using words. Here, we introduce a methodology based on the Generalized Procrustes Algorithm to measure intra-modal representational convergence at the single-stimulus level. We applied this to vision models with distinct training objectives, selecting stimuli based on their degree of alignment (intra-modal dispersion). Crucially, we found that this intra-modal dispersion strongly modulates alignment between vision and language models (cross-modal convergence). Specifically, stimuli with low intra-modal dispersion (high agreement among vision models) elicited significantly higher cross-modal alignment than those with high dispersion, by up to a factor of two (e.g., in pairings of DINOv2 with language models). This effect was robust to stimulus selection criteria and generalized across different pairings of vision and language models. Measuring convergence at the single-stimulus level provides a path toward understanding the sources of convergence and divergence across modalities, and between neural networks and human neural representations.

Jack King, Evelina Fedorenko, Eghbal A. Hosseini

In autoregressive large language models (LLMs), temporal straightening offers an account of how the next-token prediction objective shapes representations. Models learn to progressively straighten the representational trajectory of input sequences across layers, potentially facilitating next-token prediction via linear extrapolation. However, a direct link between this trajectory and token-level behavior has been missing. We provide such a link by relating contextual curvature-a geometric measure of how sharply the representational trajectory bends over recent context-to next-token entropy. Across two models (GPT-2 XL and Pythia-2.8B), contextual curvature is correlated with entropy, and this relationship emerges during training. Perturbation experiments reveal selective dependence: manipulating curvature through trajectory-aligned interventions reliably modulates entropy, while geometrically misaligned perturbations have no effect. Finally, regularizing representations to be straighter during training modestly reduces token-level entropy without degrading validation loss. These results identify trajectory curvature as a task-aligned representational feature that influences behavioral uncertainty in LLMs.

Lukas Wolf, Greta Tuckute, Klemen Kotar et al.

Training on multiple modalities of input can augment the capabilities of a language model. Here, we ask whether such a training regime can improve the quality and efficiency of these systems as well. We focus on text--audio and introduce Whisbert, which is inspired by the text--image approach of FLAVA (Singh et al., 2022). In accordance with Babylm guidelines (Warstadt et al., 2023), we pretrain Whisbert on a dataset comprising only 100 million words plus their corresponding speech from the word-aligned version of the People's Speech dataset (Galvez et al., 2021). To assess the impact of multimodality, we compare versions of the model that are trained on text only and on both audio and text simultaneously. We find that while Whisbert is able to perform well on multimodal masked modeling and surpasses the Babylm baselines in most benchmark tasks, it struggles to optimize its complex objective and outperform its text-only Whisbert baseline.