Marius Mosbach

Latent reasoning via continuous chain-of-thoughts (Latent CoT) has emerged as a promising alternative to discrete CoT reasoning. Operating in continuous space increases expressivity and has been hypothesized to enable \textit{superposition}: the ability to maintain multiple candidate solutions simultaneously within a single representation. Despite theoretical arguments, it remains unclear whether language models actually leverage superposition when given continuous reasoning tokens. We investigate this question across three regimes: a \textit{training-free} regime that constructs latent thoughts as convex combinations of token embeddings, a \textit{fine-tuned} regime where a base model is adapted to produce latent thoughts, and a \textit{from-scratch regime} where a model is trained entirely with latent thoughts to solve a given task. Using \logitlens and entity-level probing to analyze internal representations, we find that only models trained \textit{from scratch} exhibit signs of using superposition. In the training-free and fine-tuned regimes, we find that the superposition either collapses or is not used at all, with models discovering shortcut solutions instead. We argue that this is due to two complementary phenomena: i) pretraining on natural language data biases models to commit to a token in the last layers ii) capacity has a huge effect on which solutions a model favors. Together, our results offer a unified explanation for when and why superposition arises in continuous chain-of-thought reasoning, and identify the conditions under which it collapses.

LLMs memorize sensitive training data, including personally identifiable information (PII), creating a pressing need for reliable post hoc removal methods. Unlearning has emerged as a promising solution, with state-of-the-art (SOTA) methods often following a localize-first, unlearn-second paradigm that targets specific model parameters. However, existing benchmarks evaluate unlearning solely at the output level, leaving open the question of whether unlearning truly erases knowledge from a model’s parameters or merely obfuscates it, a concern reinforced by the success of resurfacing attacks. To bridge this gap, we introduce LACUNA: the first unlearning testbed with ground-truth parameter-level localization. LACUNA injects PII of synthetic individuals into predefined parameters of 1B and 7B OLMo-based models via masked continual pretraining, enabling direct evaluation of whether unlearning targets the weights responsible for knowledge storage. We use LACUNA to benchmark current SOTA unlearning methods and find that, despite strong output-level performance, existing methods are highly imprecise and susceptible to resurfacing attacks. We further show that when localization is successful, even a simple gradient-based unlearning method achieves strong erasure and robustness to resurfacing attacks, highlighting the importance of precise unlearning. We release LACUNA to complement behavioral evaluations and drive further advances in robust, localization-based unlearning.

Fine-tuning LLM-based text embedders via contrastive learning maps inputs and outputs into a new representational space, discarding the LLM's output semantics. We propose LLM2Vec-Gen, a self-supervised alternative that instead produces embeddings directly in the LLM's output space by learning to represent the model's potential response. Specifically, trainable special tokens are appended to the input and optimized to compress the LLM's own response into a fixed-length embedding, guided by an unsupervised embedding teacher and a reconstruction objective. Crucially, the LLM backbone remains frozen and training requires only unlabeled queries. LLM2Vec-Gen achieves state-of-the-art self-supervised performance on the Massive Text Embedding Benchmark (MTEB), improving by 8.8% over the unsupervised embedding teacher. Since the embeddings preserve the LLM's response-space semantics, they inherit capabilities such as safety alignment (up to 22.6% reduction in harmful content retrieval) and reasoning (up to 35.6% improvement on reasoning-intensive retrieval). Finally, the learned embeddings are also interpretable: they can be decoded back into text to reveal their semantic content. Our code, data, and pretrained models are publicly available at [github.com/McGill-NLP/llm2vec-gen](https://github.com/McGill-NLP/llm2vec-gen).

Understanding and controlling the behavior of large language models (LLMs) is an increasingly important topic in multilingual NLP. Beyond prompting or fine-tuning, \textit{language steering} , i.e., manipulating internal representations during inference, has emerged as a more efficient and interp...

A large language model’s (LLM’s) out-of-distribution (OOD) generalisation is crucial to its deployment. Previous work assessing LLMs’ generalisation performance, however, typically focuses on a single out-of-distribution dataset. This approach may fail to precisely evaluate the capabilities of th...