Transformer LLMs have been shown to exhibit strong reasoning ability that scales with inference-time compute, most prominently through token-space "thinking" chains of thought. A growing line of work pushes extra computation into the model's latent space, which we term Auxiliary Latent-Space Computation (ALSC). Existing ALSC methods largely fall into three buckets: (i) token-mediated latent rollouts, (ii) residual/activation steering, and (iii) memory (KV) compression. An underexplored alternative is memory consolidation/reconsolidation, two processes in the brain that are responsible for stabilising newly formed memory traces, and, upon recall, transiently rendering established traces plastic such they can integrate new contextual information before restabilising. In Transformer LLMs, this can be seen as analogous to performing in-place rewrites of new KV segments, and rewrites of recalled past segments. In this work, we give a theoretical justification as to why memory (re)consolidation via KV cache rewrites is beneficial for improved reasoning. We do this through the lens of Information Bottleneck (IB) theory, which posits that model generalisation emerges from an optimal balance between input information compression and retention of predictive information in latent representations. We then introduce the Bottlenecked Transformer, which augments a backbone LLM with a Cache Processor, an auxiliary Transformer that performs periodic, non-causal, in-place KV rewrites at newline-delimited reasoning step boundaries. The Processor consolidates recently written KV entries and reconsolidates a small, top-k attention-selected set of prior entries. We evaluate our Bottlenecked Transformer architecture on math reasoning benchmarks. Our model sees consistent performance gains over vanilla Transformers and pause-token augmented baselines, with gains of up to +6.6pp for selected tasks/backbones.

Vision-Language Models (VLMs) such as CLIP have shown remarkable performance in cross-modal tasks through large-scale contrastive pre-training. To adapt these large transformer-based models efficiently for downstream tasks, Parameter-Efficient Fine-Tuning (PEFT) techniques like (Low-Rank Adaptation) LoRA have emerged as scalable alternatives to full fine-tuning, especially in few-shot scenarios. However, like traditional deep neural networks, VLMs are highly vulnerable to adversarial attacks, where imperceptible perturbations can significantly degrade model performance. Adversarial training remains the most effective strategy for improving model robustness in PEFT. In this work, we propose AdvCLIP-LoRA, to our knowledge the first method designed to enhance the adversarial robustness of CLIP models fine-tuned with LoRA in few-shot settings. Our method formulates training as a minimax optimization over low-rank adapters and adversarial perturbations, enabling robust adaptation with a small trainable footprint. Across eight datasets and two backbones (ViT-B/16 and ViT-B/32), AdvCLIP-LoRA achieves state-of-the-art performance in few-shot classification, adversarial base-to-new generalization, and cross-dataset transfer, delivering higher adversarial robustness than prompt tuning baselines without sacrificing much clean accuracy. These findings highlight AdvCLIP-LoRA as a practical approach for robust adaptation of VLMs in resource-constrained settings.

Large Language Models (LLMs) exhibit remarkable capabilities across various tasks, yet guiding them to follow desired behaviours during inference remains a significant challenge. Activation steering offers a promising method to control the generation process of LLMs by modifying their internal activations. However, existing methods commonly intervene in the model's behaviour using steering vectors generated by the model itself, which constrains their effectiveness to that specific model and excludes the possibility of leveraging powerful external expert models for steering. To address these limitations, we propose ExpertSteer, a novel approach that leverages arbitrary specialized expert models to generate steering vectors, enabling intervention in any LLMs. ExpertSteer transfers the knowledge from an expert model to a target LLM through a cohesive four-step process: first aligning representation dimensions with auto-encoders to enable cross-model transfer, then identifying intervention layer pairs based on mutual information analysis, next generating steering vectors from the expert model using Recursive Feature Machines, and finally applying these vectors on the identified layers during inference to selectively guide the target LLM without updating model parameters. We conduct comprehensive experiments using three LLMs on 15 popular benchmarks across four distinct domains. Experiments demonstrate that ExpertSteer significantly outperforms established baselines across diverse tasks at minimal cost.

Fine-tuning large language models (LLMs) on a mixture of diverse datasets poses challenges due to data imbalance and heterogeneity. Existing methods often address these issues across datasets (globally) but overlook the imbalance and heterogeneity within individual datasets (locally), which limits their effectiveness. We introduce Hierarchical Balancing Optimization (HBO), a novel method that enables LLMs to autonomously adjust data allocation during fine-tuning both across datasets (globally) and within each individual dataset (locally). HBO employs a bilevel optimization strategy with two types of actors: a Global Actor, which balances data sampling across different subsets of the training mixture, and several Local Actors, which optimizes data usage within each subset based on difficulty levels. These actors are guided by reward functions derived from the LLM's training state, which measure learning progress and relative performance improvement. We evaluate HBO on three LLM backbones across nine diverse tasks in multilingual and multitask setups. Results show that HBO consistently outperforms existing baselines, achieving significant accuracy gains. Our in-depth analysis further demonstrates that both the global actor and local actors of HBO effectively adjust data usage during fine-tuning. HBO provides a comprehensive solution to the challenges of data imbalance and heterogeneity in LLM fine-tuning, enabling more effective training across diverse datasets.

It is assumed that sparse autoencoders (SAEs) decompose polysemantic activations into interpretable linear directions, as long as the activations are composed of sparse linear combinations of underlying features. However, we find that if an SAE is more narrow than the number of underlying "true features" on which it is trained, and there is correlation between features, the SAE will merge components of correlated features together, thus destroying monosemanticity. In LLM SAEs, these two conditions are almost certainly true. This phenomenon, which we call feature hedging, is caused by SAE reconstruction loss, and is more severe the narrower the SAE. In this work, we introduce the problem of feature hedging and study it both theoretically in toy models and empirically in SAEs trained on LLMs. We suspect that feature hedging may be one of the core reasons that SAEs consistently underperform supervised baselines. Finally, we use our understanding of feature hedging to propose an improved variant of matryoshka SAEs. Importantly, our work shows that SAE width is not a neutral hyperparameter: narrower SAEs suffer more from hedging than wider SAEs. As large language models (LLMs) are deployed in real-world applications, it is increasingly important to understand their internal workings. SAEs have the advantage of operating completely unsupervised, and can easily be scaled to millions of neurons in its hidden layer (hereafter called "latents" While SAEs showed promising results, recent work has cast doubt on the performance of SAEs relative to baseline techniques. Wu et al. (2025) show that SAEs underperform on both concept steering and detection relative to baselines, and Kantamneni et al. (2025) show that SAEs underperform simple linear probes on both in-domain and out-of-domain detection, even when the probes have very few training samples. The question, then, is why do SAEs underperform relative to other techniques? And if we can identify the problems holding back SAEs, can we then fix those problems?

Driven by the rapid growth of model parameters, parameter-efficient fine-tuning (PEFT) has become essential for adapting large models to diverse downstream tasks under constrained computational resources. Within this paradigm, orthogonal fine-tuning and its variants preserve semantic representations of pre-trained models, but struggle to achieve both expressiveness and efficiency in terms of parameter counts, memory, and computation. To overcome this limitation, we propose efficient Orthogonal Fine-Tuning with Principal Subspace adaptation (PSOFT), which confines orthogonal transformations to the principal subspace of pre-trained weights. Specifically, PSOFT constructs this subspace via matrix decomposition to enable compatible transformations with higher effective rank, establishes a theoretical condition that strictly maintains the geometry of this subspace for essential semantic preservation, and introduces efficient tunable vectors that gradually relax orthogonality during training to enhance adaptability. Extensive experiments on 35 NLP and CV tasks across four representative models demonstrate that PSOFT offers a practical and scalable solution to simultaneously achieve semantic preservation, expressiveness, and multi-dimensional efficiency in PEFT. The code is publicly available at https://github.com/fei407/PSOFT.

Large Language Models (LLMs) are susceptible to adversarial attacks such as jailbreaking, which can elicit harmful or unsafe behaviors. This vulnerability is exacerbated in multilingual settings, where multilingual safety-aligned data is often limited. Thus, developing a guardrail capable of detecting and filtering unsafe content across diverse languages is critical for deploying LLMs in real-world applications. In this work, we introduce a multilingual guardrail with reasoning for prompt classification. Our method consists of: (1) synthetic multilingual data generation incorporating culturally and linguistically nuanced variants, (2) supervised fine-tuning, and (3) a curriculum-based Group Relative Policy Optimization (GRPO) framework that further improves performance. Experimental results demonstrate that our multilingual guardrail, MrGuard, consistently outperforms recent baselines across both in-domain and out-of-domain languages by more than 15%. We also evaluate MrGuard's robustness to multilingual variations, such as code-switching and low-resource language distractors in the prompt, and demonstrate that it preserves safety judgments under these challenging conditions. The multilingual reasoning capability of our guardrail enables it to generate explanations, which are particularly useful for understanding language-specific risks and ambiguities in multilingual content moderation.

Navigating dilemmas involving conflicting values is challenging even for humans in high-stakes domains, let alone for AI, yet prior work has been limited to everyday scenarios. To close this gap, we introduce CLASH (Character perspective-based LLM Assessments in Situations with High-stakes), a meticulously curated dataset consisting of 345 high-impact dilemmas along with 3,795 individual perspectives of diverse values. CLASH enables the study of critical yet underex-plored aspects of value-based decision-making processes, including understanding of decision ambivalence and psychological discomfort as well as capturing the temporal shifts of values in the perspectives of characters. By benchmarking 14 non-thinking and thinking models, we uncover several key findings. Instead, new failure patterns emerge, including early commitment and overcom-mitment. This paper aims to address a core question: Can LLMs make proper judgments in high-stakes dilemmas according to different perspectives?

Typographic attacks exploit the interplay between text and visual content in multimodal foundation models, causing misclassifications when misleading text is embedded within images. Existing datasets are limited in size and diversity, making it difficult to study such vulnerabilities. In this paper, we introduce SCAM, the largest and most diverse dataset of real-world typographic attack images to date, containing 1162 images across hundreds of object categories and attack words. Through extensive benchmarking of Vision-Language Models on SCAM, we demonstrate that typographic attacks significantly degrade performance, and identify that training data and model architecture influence the susceptibility to these attacks. Our findings indicate that typographic attacks remain effective against state-of-the-art Large Vision-Language Models, especially those employing vision encoders inherently vulnerable to such attacks. However, employing larger Large Language Model backbones reduces this vulnerability while simultaneously enhancing typographic understanding. Additionally, we demonstrate that synthetic attacks closely resemble real-world (handwritten) attacks, validating their use in research. Our work provides a comprehensive resource and empirical insights to facilitate future research toward robust and trustworthy multimodal AI systems. Finally, we publicly release the datasets introduced in this paper, along with the code for evaluations under www.bliss.berlin/research/scam.

MICS, CentraleSup elec, Universit e Paris-Saclay Large Language Models (LLMs) with reasoning capabilities have achieved state-of-the-art performance on a wide range of tasks. Despite its empirical success, the tasks and model scales at which reasoning becomes effective, as well as its training and inference costs, remain underexplored. In this work, we rely on a synthetic data distillation framework to conduct a large-scale supervised study. We compare Instruction Fine-Tuning (IFT) and reasoning models of varying sizes, on a wide range of math-centric and general-purpose tasks, evaluating both multiple-choice and open-ended formats. Our analysis reveals that reasoning consistently improves model performance, often matching or surpassing significantly larger IFT systems. Notably, while IFT remains Pareto-optimal in training and inference costs, reasoning models become increasingly valuable as model size scales, overcoming IFT performance limits on reasoning-intensive and open-ended tasks. Reasoning helps most on open-ended and math tasks; gains are limited or inconsistent on general multiple-choice tasks. Large Language Models (LLMs) that generate explicit Chains of Thought (CoT) have rapidly become a defining paradigm. The research community is releasing increasingly capable reasoning models, which consistently outperform standard Instruction Fine-Tuned (IFT) counterparts at test time, especially on math, coding, and other reasoning-heavy tasks DeepSeek-AI (2025); OpenAI (2024); Mistral-AI (2025). Despite rapid progress, we still lack clarity on when explicit reasoning is most beneficial. Both prior evidence and our findings (Figure 1) point to a highly task-dependent picture: reasoning yields substantial gains on math and coding benchmarks where multi-step problem solving is essential (Zhu et al., 2024), but provides only limited improvements on simpler factual or classification tasks (Liu et al., 2024). As Figure 1 shows, these gains concentrate on reasoning-intensive (e.g., gsm8k, aime) and open-ended tasks, while benefits on general multiple-choice tasks are much smaller or inconsistent. Meanwhile, the scaling dynamics of reasoning models pose further challenges.