Searcharxiv⌕ Search

arXiv subjects

Ivan Titov

Publications and source records attributed to Ivan Titov.

At least 19 recordsLinked to original sources

Protecting Safety during Reasoning Training with Low-Rank Fisher Regularisation

Large language models are trained to have safety guardrails, but these can be compromised while fine-tuning for new capabilities, such as reasoning. One way to mitigate this degradation is to modify the fine-tuning objective to preserve safety behaviour, but doing so effectively and efficiently remains challenging. We revisit Elastic Weight Consolidation (EWC), which penalises parameter updates based on sensitivity estimates from the diagonal of the Fisher information matrix (FIM). While EWC provides some protection, the diagonal approximation ignores interactions between parameters. We address this limitation by exploiting recent findings that safety-relevant behaviour is located in a low-rank subspace within models. Our method, Subspace-Projected Elastic Weight Consolidation (SP-EWC), projects gradients onto this subspace to capture non-diagonal curvature efficiently. We investigate three different ways of identifying the safety subspace: through safety fine-tuning updates (LoRA), safety-related activations (ActSVD), and contrastive representations of safety and utility data (cPCA). We find that this choice substantially affects performance, with SP-EWC using cPCA comparing favourably against existing methods by better maintaining safety guardrails without introducing excessive over-refusal.

cs.LG↗

Knowledge Offloading: Decomposing LLMs into Sparse Backbones and Memory Modules

LLMs encode both general capabilities and domain-specific knowledge in a single set of parameters. We ask whether this capacity can be reorganized: keeping broadly useful computation in a shared backbone, while moving specialized knowledge into external memory modules. We propose \emph{knowledge offloading} (KOFF), a framework for decomposing a pretrained LLM into a sparse shared backbone and domain-specific memories. Starting from a frozen base model, we either jointly or sequentially learn a structured pruning mask and lightweight recovery modules, implemented as LoRA adapters and learned key-value caches. Across Llama and Qwen models from 3B to 14B, we find that 12\% to 20\% of capacity can be moved out of the shared backbone while largely preserving perplexity and MMLU performance (up to 90\% of the full model). We show that LoRA and learned KV memories are complementary, and analyses suggest that the offloaded knowledge is specialized: mismatched memories do not yield the same gains, and language-specific neurons are preferentially removed while language-neutral neurons largely remain in the backbone. These results suggest that the knowledge of a pretrained LLM can be reallocated between a shared core and swappable external memories.

cs.LG↗

Post-hoc Disentanglement of Textual and Acoustic Features in Pre-trained Speech Encoders

Pre-trained speech models build entangled internal representations, which capture a variety of features (e.g., pitch, loudness, syntax, or semantics of an utterance) in a distributed encoding. This entanglement makes it difficult to track how such representations rely on textual versus acoustic information when used in downstream applications, limiting their interpretability and transparency. In this paper, we explore an approach to post-hoc disentanglement that builds on the Information Bottleneck principle. We take representations from pre-trained speech encoders as input, and separate them into two distinct components in a cascaded fashion: first we learn a vector encoding content (i.e., what can be transcribed as text), and nothing else; second, we learn a vector encoding all complementary acoustic features relevant to a target task. We apply and evaluate our proposed setup to emotion recognition and speaker identification target tasks, quantifying the relative contribution of textual and acoustic features at each model layer. Finally, we use the approach for disentangled feature attribution as it allows us to identify the most salient speech frames from both the textual and acoustic perspectives.

cs.CL↗

The Missing Minimal Pair: Stereotype Evaluation in LLMs

A common approach to measuring bias in Large Language Models is to compare the log-likelihoods of two contrastive stereotype sentences. We argue that such single-pair comparisons are often unreliable: simply rewriting the same stereotype with an alternative attribute can yield logically inconsistent preferences. To address this, we propose a dual minimal pair setup that introduces two axes of comparison for robust stereotype evaluation. First, we present a data-augmentation framework that fills critical gaps in existing stereotype datasets by generating paraphrases and alternate attributes. We apply our framework on a set of English, Russian, Spanish and Chinese stereotypes. Second, we introduce two evaluation metrics tailored to the dual minimal pair setup. One of these metrics provides a new perspective on bias by modeling the mutual information (MI) between social groups and stereotyped attributes. This MI-based metric is better suited for aggregation and enables more robust comparisons of stereotype strength across different languages and models. Our code is available at https://github.com/stepanat/missing-minimal-pair/.

cs.CL↗

Geometric Self-Distillation for Reasoning Generalization

On-policy distillation provides dense teacher supervision on a language model's own trajectories. In self-distillation with privileged context, this supervision comes from the model itself, conditioned on a hint or solution trace hidden from the student. When the teacher's preferences hinge on privileged information, it can assign higher probability to continuations the student cannot infer from its own context. Matching these preferences throughout training can induce predictive drift and degrade out-of-distribution (OOD) reasoning. We propose GeoSD, a self-distillation method that controls this drift through two complementary geometric terms. A Hellinger loss weights each teacher preference by the student--teacher overlap, reducing the influence of tokens to which the student assigns low probability. Because these influences can still accumulate, a Fisher--Rao penalty regulates predictive distance from a copy of the student refreshed periodically during training. Both terms compare next-token distributions in Fisher--Rao geometry and are jointly optimized with a preconditioner motivated by the natural gradient. Across three model families, GeoSD retains strong in-distribution gains while improving average mathematical OOD accuracy by 5.7--8.6 points over the base model. OOD gains hold across five model scales from 1.7B to 32B and transfer to code generation, where GeoSD improves code accuracy by 1.9 points on average despite distilling on mathematics alone. Our analysis of mathematical reasoning shows that standard matching rapidly concentrates probability mass at high-entropy states and that its samples confidently agree on incorrect answers. In contrast, GeoSD preserves alternative token mass and reduces false consensus.

cs.LG↗

Shared Doubt: Zero-Shot Cross-Lingual Confidence Estimation for Language Models

Confidence estimation (CE), i.e. quantifying the reliability of a model's prediction, has attracted great interest in the context of large language models (LLMs). However, most studies focus on English, ignoring the multilingual reality of LLM usage, while many CE methods degrade or require retraining across languages. To address this gap, we investigate whether multilingual LLMs encode shared, language-transferable confidence features in open-ended question answering. We use a lightweight linear probe that predicts answer correctness directly from intermediate representations. Trained monolingually, the probe generalizes zero-shot to unseen, linguistically diverse languages without target-language supervision. Multiple ablations and learned layer weights reveal that confidence features concentrate in middle layers across languages, suggesting a shared confidence subspace. While zero-shot cross-lingual performance depends on similarity to the source language, the probe provides a strong baseline without any retraining and compares favorably to other popular confidence estimation methods.

cs.CL↗

S2a-reducibility and differentiation in Martin-Löf random reals

Solovay reducibility is studied intensively as a tool to compare the approximability and the degree of randomness of left-c.e. reals. By definition, a real is left-c.e. if it has a left-c.e. approximation, that is, it is the limit of an effective nondecreasing sequence of rationals. If reals $α$ and $β$ have left-c.e. approximations $a_0, a_1, \ldots$ and $b_0, b_1, \ldots$, respectively, such that the approximation ratios \[ \frac{α-a_n}{β-b_n} \] are bounded from above by a constant, the real $α$ is Solovay reducible to $β$. The latter is the case for any such $α$ and $β$ and their left-c.e. approximations whenever $β$ is Martin-Löf random by the Kučera-Slaman Theorem [DOI:10.1137/S0097539799357441]. This result was substantially strengthened by Barmpalias and Lewis-Pye [DOI:10.1016/j.jcss.2017.06.002], who demonstrated that, under the given assumptions, the approximation ratios are not only bounded but actually converge to a limit, which does not depend on the considered left-c.e. approximations. There is a quest for a suitable extension of Solovay reducibility to the class of all reals. Promising candidates include S2a-reducibility on the set of computably approximable reals by Zheng and Rettinger [DOI:10.1007/978-3-540-27798-9_39] and monotone Solovay reducibility by Titov [DOI:10.1007/978-3-031-95908-0_33]. For the latter, Titov [DOI:10.1017/jsl.2025.10157] demonstrated that the theorems of Kučera and Slaman and of Barmpalias and Lewis-Pye extend to all reals. He conjectured further [DOI:10.1017/jsl.2025.10157, Conjecture 3.2] that similar extensions hold for S2a-reducibility in terms of its functional characterization by Kumabe, Miyabe, and Suzuki [DOI:10.3233/COM-230486]. In this work, we refute this conjecture by proving that the analogue of the Barmpalias-Lewis-Pye Limit Theorem does not hold for S2a-reducibility.

math.LO↗

Does Out-of-Sight Equal Out-of-Mind in CoT Monitorability?

Chain-of-thought (CoT) reasoning offers a window into the decision-making of large language models (LLMs), which can be monitored for target behaviors by reading the reasoning trace, motivating work on CoT monitorability. Latent CoT approaches, however, replace the explicit tokens with a small number of continuous states, lowering inference costs but removing the readable trace this monitoring relies on. Monitoring then requires alternative access to the model, such as probing its activations or verbalizing the latent states back into text, but how much monitorability these alternatives preserve is unclear. We study this question with a hint-based intervention setup, a proxy for behaviors where models exploit biasing input cues, e.g., an inadvertently leaked answer or a belief stated by the user, without acknowledging them. Taking hint-reliance as the monitorability target, we compare monitors across reasoning modes, from explicit CoT to weakly- and strongly-supervised latent CoT, on math reasoning and question answering. We find that, in this setup, monitorability depends more on properties of the task (such as whether the correct answer constrains the supporting reasoning) and the level of access to model internals than on the reasoning mode.

cs.CL↗

Consensus as Privileged Context for Label-Free Self-Distillation

Sampling multiple solutions and returning the majority answer is among the most reliable ways to improve the reasoning accuracy of large language models without labels, and a growing family of methods converts this consensus signal into training supervision. However, existing approaches use consensus only in restricted forms: as a filter that selects solutions for fine-tuning, as a preference between answers, or as a scalar reward for reinforcement learning, discarding most of the information that the agreeing solutions contain. We present CANON (Consensus-ANchored self-distillatiON), a label-free training method that turns consensus into dense, token-level supervision. For each unlabeled prompt, CANON samples multiple solutions, extracts the majority answer, and conditions a frozen snapshot of the model on a solution that reaches it; this consensus-anchored teacher then supervises the model on its own rollouts at every token. Experiments on mathematical and scientific reasoning benchmarks show that CANON improves pass@1 by up to 12 points, outperforming label-free reinforcement learning by 6 points at a seventh of its compute and approaching a teacher conditioned on gold solutions; trained on pooled unlabeled data, it transfers to held-out benchmarks, matching training methods that use gold labels. Analysis suggests that the improvements are not pure distribution sharpening: after training, the model solves problems it previously never solved in 32 attempts, and its majority vote itself becomes more accurate.

cs.LG↗

Truthful or Fabricated? Using Causal Attribution to Mitigate Reward Hacking in Explanations

Chain-of-thought explanations are widely used to inspect the decision process of large language models (LLMs) and to evaluate the trustworthiness of model outputs, making them important for effective collaboration between LLMs and humans. We demonstrate that preference optimization - a key step in the alignment phase - can inadvertently reduce the faithfulness of these explanations. This occurs because the reward model (RM), which guides alignment, is tasked with optimizing both the expected quality of the response and the appropriateness of the explanations (e.g., minimizing bias or adhering to safety standards), creating potential conflicts. The RM lacks a mechanism to assess the consistency between the model's internal decision process and the generated explanation. Consequently, the LLM may engage in "reward hacking" by producing a final response that scores highly while giving an explanation tailored to maximize reward rather than accurately reflecting its reasoning. To address this issue, we propose enriching the RM's input with a causal attribution of the prediction, allowing the RM to detect discrepancies between the generated self-explanation and the model's decision process. In controlled settings, we show that this approach reduces the tendency of the LLM to generate misleading explanations.

cs.CL↗

Polarized neutron scattering as a probe for vortex-type spin correlations in iron oxide multicore assemblies

We report an experimental investigation of the magnetic microstructure of iron oxide multicore assemblies by means of polarized small-angle neutron scattering (SANS). Guided by a recently developed analytical theory for vortex-state magnetic nanoparticles, we provide a quantitative comparison between the measured and calculated cross sections, revealing signatures that are consistent with vortex-type magnetization configurations at low applied magnetic fields. In particular, the field evolution and the characteristic isotropic ring-type feature of the spin-flip scattering intensity at intermediate momentum transfers are in line with the formation of flux-closure states. The latter are stabilized by the interplay of exchange, Zeeman, and magnetostatic energies. The methodology allows for a statistically significant characterization of vortex states in densely packed nanoparticle systems, thereby complementing surface-sensitive techniques that are commonly limited to the observation of spin structures in individual particles.

cond-mat.mes-hall↗

Operationalising the Superficial Alignment Hypothesis via Task Complexity

The superficial alignment hypothesis (SAH) posits that large language models learn most of their knowledge during pre-training, and that post-training merely surfaces this knowledge. The SAH, however, lacks a precise definition, which has led to (i) different and seemingly orthogonal arguments supporting it, and (ii) important critiques to it. We propose a new metric called task complexity: the length of the shortest program that achieves a target performance on a task. In this framework, the SAH simply claims that pre-trained models drastically reduce the complexity of achieving high performance on many tasks. Our definition unifies prior arguments supporting the SAH, interpreting them as different strategies to find such short programs. Experimentally, we estimate the task complexity of mathematical reasoning, machine translation, and instruction following; we then show that these complexities can be remarkably low when conditioned on a pre-trained model. Further, we find that pre-training enables access to strong performances on our tasks, but it can require programs of gigabytes of length to access them. Post-training, on the other hand, collapses the complexity of reaching this same performance by several orders of magnitude. Overall, our results highlight that task adaptation often requires surprisingly little information -- often just a few kilobytes.

cs.LG↗

Blending Supervised and Reinforcement Fine-Tuning with Prefix Sampling

Existing LLMs-post-training techniques are broadly categorized into supervised fine-tuning (SFT) and reinforcement fine-tuning (RFT). Each paradigm presents a distinct trade-off: (1) SFT excels at mimicking demonstration data, but can lead to problematic generalization as a form of behavior cloning. (2) Conversely, RFT can significantly enhance a model's performance but is prone to learning unexpected behaviors, and its performance is sensitive to the initial policy. In this paper, we propose a unified view of these methods and introduce Prefix-RFT, a hybrid approach that synergizes learning from both demonstration and exploration. Using mathematical reasoning problems as a test bed, we empirically demonstrate that Prefix-RFT is simple yet effective. Not only does it surpass the performance of standalone SFT and RFT, but it also outperforms parallel mixed-policy RFT methods. Our analysis highlights the complementary nature of SFT and RFT, validating that Prefix-RFT effectively harmonizes them. Further ablation studies confirm the method's robustness to variations in the quality and quantity of demonstration data.

cs.LG↗

The Cancellation Hypothesis in Critic-Free RL: From Outcome Rewards to Token Credits

A commonly accepted explanation of critic-free RL for LLMs, based on sequence-level rewards, is that it reinforces successful rollouts with a positive advantage while penalizing failed ones. In contrast, we study critic-free RL from a token-level perspective, revealing the token-flipping phenomenon: positive and negative rollouts exhibit remarkably similar proportions of tokens whose probabilities are boosted or suppressed during RL training. To explain this phenomenon, we further show that a token's change in probability is not fully determined by its own advantage; coupled gradient interactions with other tokens also play a non-negligible role. Specifically, these token coupling effects occur primarily between identical tokens that are both predicted with low confidence. Building upon this analysis, we propose the cancellation hypothesis: as a result of coupling, opposing signals cancel out for tokens shared by positive and negative rollouts, while tokens more specific to successful rollouts receive stronger reinforcement, thereby inducing hidden token-level credit assignment from rollout-level rewards. We support this hypothesis with complementary empirical evidence. (1) Compared with training on only positive rollouts, critic-free RL shifts updates from template and formatting tokens toward reasoning tokens; (2) Tokens boosted by critic-free RL consistently demonstrate higher value than suppressed tokens, regardless of whether they originate from positive or negative rollouts. Guided by this view, we implement two batching interventions to encourage or preserve cancellation in critic-free RL training: query-preserved mini-batching and reward-balanced batching. Despite their simplicity, these interventions improve RLVR training across multiple model scales, supporting cancellation as both an explanatory principle and a practical design criterion for critic-free RL training.

cs.LG↗

Finding Culture-Sensitive Neurons in Vision-Language Models

Despite their impressive performance, vision-language models (VLMs) still struggle on culturally situated inputs. To understand how VLMs process culturally grounded information, we study the presence of culture-sensitive neurons, i.e., neurons whose activations show preferential sensitivity to inputs associated with particular cultural contexts. We examine whether such neurons are important for culturally diverse visual question answering and where they are located. Using the CVQA benchmark, we identify neurons of culture selectivity and perform diagnostic tests by deactivating the neurons flagged by various identification methods. Experiments on three VLMs across 25 cultural groups demonstrate the existence of neurons whose ablation disproportionately harms performance on questions about the corresponding cultures, while having limited effects on others. Moreover, we introduce a new margin-based selector Contrastive Activation Margin (ConAct) and show that it outperforms probability- and entropy-based methods in identifying neurons associated with cultural selectivity. Finally, our layer-wise analyses reveal that such neurons are not uniformly distributed: they cluster in specific decoder layers in a model-dependent way.

cs.LG↗

Speedability of computably approximable reals and their approximations

An approximation of a real is a sequence of rational numbers that converges to the real. An approximation is left-c.e. if it is computable and nondecreasing and is d.c.e. if it is computable and has bounded variation. A real is computably approximable if it has some computable approximation, and left-c.e. and d.c.e. reals are defined accordingly. An approximation $\{a_s\}_{s \in ω}$ is speedable if there exists a nondecreasing computable function $f$ such that the approximation $\{a_{f(s)}\}_{s \in ω}$ converges in a certain formal sense faster than $\{a_s\}_{s \in ω}$. This leads to various notions of speedability for reals, e.g., one may require for a computably approximable real that either all or some of its approximations of a specific type are speedable. Merkle and Titov established the equivalence of several speedability notions for left-c.e. reals that are defined in terms of left-c.e. approximations. We extend these results to d.c.e. reals and d.c.e. approximations, and we prove that in this setting, being speedable is equivalent to not being Martin-Löf random. Finally, we demonstrate that every computably approximable real has a computable approximation that is speedable.

math.LO↗

Anthropomimetic Uncertainty: What Verbalized Uncertainty in Language Models is Missing

Human users increasingly communicate with large language models (LLMs), but LLMs suffer from frequent overconfidence in their output, even when its accuracy is questionable, which undermines their trustworthiness and perceived legitimacy. Therefore, there is a need for language models to signal their confidence in order to reap the benefits of human-machine collaboration and mitigate potential harms. Verbalized uncertainty is the expression of confidence with linguistic means, an approach that integrates perfectly into language-based interfaces. Most recent research in natural language processing (NLP) overlooks the nuances surrounding human uncertainty communication and the biases that influence the communication of and with machines. We argue for anthropomimetic uncertainty, the principle that intuitive and trustworthy uncertainty communication requires a degree of imitation of human linguistic behaviors. We present a thorough overview of the research in human uncertainty communication, survey ongoing research in NLP, and perform additional analyses to demonstrate so-far underexplored biases in verbalized uncertainty. We conclude by pointing out unique factors in human-machine uncertainty and outlining future research directions towards implementing anthropomimetic uncertainty.

cs.CL↗

Analyzing LLM Instruction Optimization for Tabular Fact Verification

Instruction optimization provides a lightweight, model-agnostic approach to enhancing the reasoning performance of large language models (LLMs). This paper presents the first systematic comparison of instruction optimization, based on the DSPy optimization framework, for tabular fact verification. We evaluate four out-of-the-box prompting techniques that cover both text-only prompting and code use: direct prediction, Chain-of-Thought (CoT), ReAct with SQL tools, and CodeAct with Python execution. We study three optimizers from the DSPy framework -- COPRO, MiPROv2, and SIMBA -- across four benchmarks and three model families. We find that instruction optimization consistently improves verification accuracy, with MiPROv2 yielding the most stable gains for CoT, and SIMBA providing the largest benefits for ReAct agents, particularly at larger model scales. Behavioral analyses reveal that SIMBA encourages more direct reasoning paths by applying heuristics, thereby improving numerical comparison abilities in CoT reasoning and helping avoid unnecessary tool calls in ReAct agents. Across different prompting techniques, CoT remains effective for tabular fact checking, especially with smaller models. Although ReAct agents built with larger models can achieve competitive performance, they require careful instruction optimization.

cs.CL↗