arXiv subjects

Hinrich Schütze

Publications and source records attributed to Hinrich Schütze.

At least 19 recordsLinked to original sources

arXiv2026

GLUScope: A Tool for Analyzing GLU Neurons in Transformer Language Models

We present **GLUScope**, an open-source resource for analyzing neurons in Transformer-based language models (LMs), intended for interpretability researchers. We focus on the now-standard *GLU-based activation functions* such as SwiGLU, used in virtually all recent open-weights LMs. This introduces a new challenge: understanding positive activations is not enough. Instead, both the *gate* and the *in* activation of a neuron can be positive or negative, leading to four different possible sign combinations (*activation quadrants*) that in some cases have quite different functionalities (e.g. promoting vs. suppressing a specific set of tokens). Accordingly, for any neuron, our tool shows text examples for each of the four sign combinations, and indicates how often each combination occurs. Our resource can lead to novel insights: in particular, thanks to analyzing all four activation quadrants, it can surface interpretable behavior that other tools would miss. We also release the code used to create the resource, thus enabling researchers to build on it by adding more models, text corpora, or neuron pages.

cs.CL↗

arXiv2026

Output Vector Editing for Memorization Mitigation in Large Language Models

Large language models can reproduce training text, creating privacy, copyright, and security risks. We propose Output Vector Editing (OVE), a post-hoc method for suppressing text continuations that a model reproduces from a given prefix. Under superposition, a neuron can contribute to multiple features. Yet some neuron-level mitigation methods suppress localized neurons by zeroing their activations, removing each neuron's entire contribution. OVE instead edits their output vectors, combining (i) two-stage localization of MLP layers and neurons with (ii) full-vocabulary least-squares updates. Its multi-anchor variant, OVE-MA, adaptively selects continuation positions expected to provide stronger suppression relative to the required weight change. We evaluate recovery and locality across four models (360M--7B parameters), using autoregressively validated continuations and recovery tests with greedy decoding, beam search, and sampling. On 634 OLMo-7B pairs, OVE-MA with at most five output-vector updates per request leaves 24.45\% recoverable in these tests, compared with 32.02\% for the closest prior post-hoc editing method, REVS, while using approximately 75 times fewer update operations. These results show that adaptive position selection can concentrate effective suppression into a small number of weight updates. OVE edits weights without gradients and requires no additional intervention during inference.

cs.CL↗

arXiv2026

Predicting Linear Relational Decodability Without the Relation

Linear relational embeddings (LREs) approximate a language model's subject-to-object mapping with an affine function. Their faithfulness varies enormously across relations, from near zero for person_father to the ceiling for country_capital_city. We investigate the factors that contribute to this variation and demonstrate that it can be significantly predicted from a readout that is not aware of the underlying relation. To do so, we propose a novel form of the Jacobian lens, the entity lens. Whereas the Jacobian lens is estimated across all tokens and concepts, the entity lens EL_t is trained to predict what the model will output for an entity of type t under the neutral prompt "{} is known for"; e.g., we train ELs for countries and persons. After estimating an entity lens, we measure how highly it ranks the model's answer for a relation without using any relation-specific information in estimation and prediction. Across 28 relations of the LRE dataset on Olmo-3-7B this relation-agnostic score correlates with LRE faithfulness at $ρ= 0.7355$ (p < .01). This suggests that linear decodability is largely a property of what a subject's representation already carries, rather than of the relation being decoded. Moving the unit of analysis from the relation to the individual fact, we conclude that new relation datasets and analysis methods are needed to address several problems in existing datasets, most prominently semantically weak first tokens (e.g., most predicted universities' first token is "University") and reference competition (e.g., the LLM confuses obscure actor names with competing non-actor entities that have the same last name). Our full pipeline is available at https://github.com/jonasrohw/entity-lens

cs.CL↗

arXiv2026

Knowledge That Is Linear Before Fine-Tuning Generalizes After It

Not all facts a language model can recall support further learning equally well. We distinguish active facts, which support generalization beyond their own recall, from inert facts, which can be recalled but do not support such generalization in a given task. We study this distinction by constructing new relations from existing ones. Take the relation that maps a place to its continent (Paris -> Europe) and assign each of its answers an arbitrary new label (Europe -> crab, Asia -> zebra). The new relation then maps Paris -> crab. We fine-tune the model on the new relation for a subset of subjects, without telling it the original relation or the relabeling rule, and test whether it assigns the corresponding labels to held-out subjects that share an answer in the original relation. Having learned Paris -> crab, does it answer Munich -> crab? Across eleven language models and 70 relations, how linearly readable the original relation is in pretrained representations strongly predicts this generalization. Within each model, the two correlate across relations at Pearson r = 0.87 to 0.92. This association persists when controlling, separately and jointly, for the number of fine-tuning subjects and answers, the model's accuracy on the original relation, and how often the subjects occur in pretraining text. It is also robust to re-randomizing the label assignment and fine-tuning seed, and to replacing the animal labels with labels from thirteen other semantic categories. When we map a fine-tuned model's wrong label back to the original answer it was assigned to, Munich -> zebra becomes Munich -> Asia. Answers recovered this way are more often among the linear probe's top answers for the subject than among the pretrained model's own. Whether a new relation generalizes depends less on whether its facts can be recalled than on how readily their shared structure can be read out linearly.

cs.CL↗

arXiv2026

Bridged Self-Distillation for On-Policy Learning

On-policy distillation can transfer the reasoning capabilities of a strong teacher to a smaller student. However, distillation from a stronger teacher model faces two mismatches: its solutions may be difficult for a smaller student to imitate, and its token-level probabilities may be incompatible with, or even unavailable to, the student. Existing methods handle these settings through different forms of supervision. To this end, we propose Bridged Self-Distillation (BSD), a unified framework for distilling from any teacher that provides reasoning traces. The model to be distilled from, termed the external teacher, first provides a reasoning trace as privileged context, which can be generated in an offline fashion. The student generates rollouts without this context, while a copy of the student conditioned on it, serving as the self-teacher, scores the same rollouts in the student's own token space. BSD uses the resulting token-level guidance to adjust updates on unsuccessful rollouts, with a verifier providing the outcome signal. As the external teacher is queried only to produce text, the same training procedure applies to all teacher scenarios: shared-tokenizer, cross-tokenizer, and closed API teachers. Across these scenarios, BSD improves over the base model and outperforms baselines tailored to each scenario, offering a novel and unified pathway for distillation. Code is available: https://github.com/TUMLegalTech/Bridged-Self-Distillation

cs.CL↗

arXiv2026

Beyond Input Understanding: Diagnosing Multilingual Mathematical Reasoning with Directed Acyclic Trace Graphs

Large reasoning models (LRMs) achieve strong mathematical reasoning performance in English, but remain much less reliable in many low- and medium-resource languages. This gap is often explained as a failure to understand non-English problem statements. We show that this view is incomplete: even when the problem is given in English, controlling the model's reasoning language can substantially reduce accuracy, suggesting that language also affects reasoning execution itself. To study this effect, we introduce DATG, a Directed Acyclic Trace Graph framework that maps reasoning traces to language-independent mathematical anchors and dependencies. This allows us to align target-language traces with reference DAGs and measure whether they cover required mathematical nodes, respect dependency edges, and avoid harmful mathematical actions. Experiments on the Qwen3 series across 12 languages show that non-English reasoning often suffers from reduced anchor coverage and weaker dependency fidelity, especially in low-resource languages. Motivated by this diagnosis, we propose Loop-Retry and Formula-Retry, two simple test-time controls targeting DATG-exposed failure modes, and show that they consistently improve target-language reasoning performance in low-resource languages.

cs.CL↗

arXiv2026

Beyond Clean Accuracy: The Practical Challenges and Benefits of Pixel-Based Language Models

Pixel-based language models process rendered text without relying on a fixed subword vocabulary, offering a promising approach to noisy and multilingual inputs. However, their practical value depends on whether they can be reliably adapted to downstream tasks. To this end, we systematically examine this question through fine-tuning behavior, typographical robustness, and zero-shot crosslingual transfer. Under our tested settings, pixel-based models require longer training and broader labelled-data coverage than text-based models, while delayed learning and learning-rate-dependent failures make adaptation less reliable. Our analyses further suggest that task-relevant information is less readily accessible in pretrained pixel representations, placing greater demands on fine-tuning. Once successfully adapted, however, pixel-based models lose substantially less accuracy under typographical noise and achieve higher target-language accuracy than English-pretrained text-based counterparts on two crosslingual benchmarks, despite weaker clean English performance. Together, these findings reveal a central trade-off: replacing tokenization with rendering reduces sensitivity to corrupted and unfamiliar inputs, but shifts a central bottleneck to downstream adaptation.

cs.CL↗

arXiv2026

FramingQA: Does the Question Shape the Answer? Measuring the Compositional Framing Effect

We introduce FramingQA, a benchmark that measures the model sensitivity to question framing across law, medicine, finance, and robotic simulations. Large language models (LLMs) often change their responses to subtle rephrasings that align with an implied stance by users. This can leave users with advice tainted by how they happened to phrase a question rather than by the underlying facts, and the consequences are highly costly in high-stakes domains. Because in the realistic scenarios, both expert practitioners and non-expert users frequently ask LLMs questions containing incomplete or misleading assumptions, models are highly susceptible to those framings. To test this, we inject the framing bias across three nested levels: a framing-biased question phrasing (root), an injected framing-biased premise prepended to a neutral question (propositional), and a premise paired with a framing-biased question (global). Evaluating nine open models (3.8B-70B) across four families, we find that strong per-variant accuracy does not guarantee the robustness across differently phrased questions under the fixed factual information.

cs.CL↗

arXiv2026

Orthogonal Yet Coupled: Decoupling Geometric Components for Model Merging

Merging pretrained models has emerged as an effective approach for consolidating diverse capabilities into a single unified model. However, prevailing merging methods typically treat each task vector as an indivisible merging unit, overlooking the heterogeneous geometric changes encoded within it. This treatment can induce cross-component coupling: when merging decisions are derived from statistics of the complete task vector, the geometric characteristics of one component may influence how another is selected, weighted, or combined, potentially degrading the quality of the merged model. To address this issue, we propose DiGA, a Disentangled Geometry-Aware model merging framework. Using the pretrained weights as a shared geometric reference, DiGA orthogonally decomposes each task vector into components corresponding to distinct geometric attributes. Rather than merging the task vectors as a whole, DiGA aggregates corresponding components independently within their respective subspaces and subsequently recombines them into a unified update. This component-wise formulation preserves the geometric identity of each component and prevents the characteristics of one component from interfering with the aggregation of another. Furthermore, DiGA can be incorporated into a broad range of existing model merging methods. Extensive experiments across diverse models, tasks, and merging methods demonstrate that DiGA improves merged-model performance and reduces capability degradation. Our repository is on https://github.com/wzj1718/DiGA.

cs.CL↗

arXiv2026

Why Better Cross-Lingual Alignment Fails for Better Cross-Lingual Transfer: Case of Encoders

Cross-lingual alignment is often assumed to improve cross-lingual transfer by bringing representations of different languages closer together. However, improvements in representational alignment do not consistently translate into better downstream performance. We investigate this disconnect using XLM-R models explicitly aligned across four language pairs with token-level, sentence-level, and masked-language-modeling objectives. We evaluate their zero-shot transfer on a token-level task (part-of-speech tagging) and a sentence-level task (sentence classification), and analyze both representational changes and the gradients induced by the alignment and downstream objectives. We find that embedding-based alignment metrics do not reliably indicate whether alignment will improve or degrade downstream performance. Moreover, alignment and downstream-task gradients are often nearly orthogonal, particularly when the alignment objective and downstream task operate at different representational levels. These findings suggest that representation alignment alone is insufficient for assessing cross-lingual transfer, and that the compatibility between alignment and downstream objectives should be considered when designing/evaluating alignment methods.

cs.CL↗

arXiv2026

Calibration Is Not Enough: Evaluating Confidence Estimation Under Language Variations

Confidence estimation (CE) indicates how reliable the answers of large language models are and impacts user trust and decision-making. Existing evaluations mainly concern the alignment between confidence and correctness, but ignore the variability of language: confidence estimates should remain consistent under semantically equivalent prompts or answer variations, while changing when answer meaning differs, as this may indicate a change in correctness. Therefore, we introduce a novel evaluation framework based on three complementary properties: \textbf{robustness} to prompt perturbations, \textbf{stability} across semantically equivalent answers, and \textbf{sensitivity} to semantically different answers. We show that these metrics are largely independent from existing CE metrics, and that common CE methods often fail on them: while most methods achieve high robustness and stability, they struggle to distinguish semantically different answers, potentially because they do not effectively leverage generation-side information. Overall, our framework exposes overlooked limitations of current CE evaluations and provides guidance for selecting confidence estimators for real-world applications.

cs.CL↗

arXiv2026

A Systematic Evaluation of Cross-Lingual Consistency Enhancement Methods in Multilingual Language Models

Multilingual language models often produce inconsistent answers to semantically equivalent questions across languages, motivating methods to improve cross-lingual consistency (CLC). However, existing methods are typically evaluated using different models, tasks, and protocols, leaving their relative strengths unclear. In this work, we present a unified evaluation of representative CLC-enhancement methods for question answering, spanning inference-time interventions and post-training approaches across three model families and three closed-form benchmarks. The results show that post-training methods are generally more reliable, with direct distribution alignment consistently improving CLC across all model-dataset combinations, while other methods are more sensitive to answer format and the breadth of language coverage. Notably, cross-domain transfer is limited unless source and target tasks share similar output formats. We further investigate whether CLC enhancement hurts models' ability to respond differently *when needed*, that is, when asked culture-dependent questions. Across two benchmarks of culturally diverse question answering, we find no systematic degradation in controlled closed-form evaluation, whereas open-ended generation reveals occasional accuracy reductions, particularly for non-English responses. Our work highlights the need to evaluate CLC enhancement for both cross-domain robustness and culturally appropriate variation, informing future work in post-training and benchmark development.

cs.CL↗

arXiv2026

Weakening Neurons: An Input-Output Functionality in Transformers with Outsize Influence

We analyze the learned input-output behavior of GLU-based neurons in large language models (LLMs). We propose a simple analysis method: For each neuron, we compute the cosine similarities between its input (reading) and output (writing) weight vectors. In this scheme, a strong negative cosine similarity indicates the neuron weakens the direction it detects in the residual stream, so we call this a weakening neuron. This allows us to gain a number of novel insights. First, we show that nine different LLMs have similar patterns: weakening neurons appear mostly in late layers whereas their counterparts, (conditional) strengthening neurons, are frequent in early-middle layers. Second, we find that weakening neurons display surprising behavior: even though there are few, they activate often and have a large influence on model behavior. Third, weakening neurons have a strong effect on model output when gate values are negative -- which is surprising since negative gate values are not expected to encode functionality.

cs.LG↗

arXiv2026

Relational Linearity is a Predictor of Hallucinations

Hallucination is a central failure mode of language models (LMs). We focus on hallucinations in response to questions like: "Which instrument did Glenn Gould play?", but we ask these questions for synthetic entities designed to be unknown to the model. We find that LMs like Gemma-7B-IT frequently hallucinate, i.e., they have difficulty recognizing that the hallucinated fact is not part of their knowledge. Based on the idea of linear relational embeddings, we put forward the following hypothesis. (i) Due to the abstract scheme that is used to represent them, LMs can easily produce plausible objects for non-existing subjects of linear relations, which can lead to hallucinations. (ii) For nonlinear relations, this mechanism for producing an object is not available and so a hallucination is easier to avoid. To test this hypothesis, we create SynthHal, a synthetic unknown-entity benchmark for 15 relations. We find that across four instruction-tuned models, relational linearity is a strong predictor of models hallucinating an object for an unknown subject vs refusing to give an answer, with correlations $r \in [.58, .84]$. While this is not direct evidence for the hypothesized causal mechanism, it is suggestive and opens up a new line of inquiry into understanding LM hallucinations.

cs.CL↗

arXiv2026

Expert-Aware Causal Tracing of Factual Recall in Sparse MoE Language Models

Activation patching can identify a mixture-of-experts (MoE) block whose clean output restores a corrupted factual prediction. However, because the block output combines contributions from multiple routed experts, block-level rescue does not establish whether the recovery localizes to an individual expert or depends on the routed expert set. We study this question on single-token COUNTERFACT contrasts by corrupting subject-token embeddings, restoring clean block outputs, and then restoring clean-minus-noised expert updates under fixed routing. In Qwen3-30B-A3B-Base, a discovery sweep selects layer 44, and held-out analysis identifies L44E069 as a recurrent routed contributor with positive specificity over same-layer active experts. Its effect is fact-matched and improves true-token probability and rank, which explains part of the layer rescue. In Mixtral-8x7B-v0.1, the selected recurrent singleton is not specific; matched-size controls instead show specificity of the clean-routed top-2 set. These findings show that successful MoE-block restoration does not necessarily imply localization to a single expert.

cs.CL↗

arXiv2026

Do We Still Need Humans in the Loop? Human vs. LLM Annotation in Active Learning for TikTok Hate Speech Detection

Annotating data remains a costly bottleneck for supervised NLP. Active learning (AL) reduces the number of human labels needed by selecting only the most informative instances, while instruction-tuned LLMs attack the same bottleneck from the other side, making labels cheap enough to annotate entire corpora. This raises two questions: can LLM labels replace human labels within the AL loop, and does AL remain necessary when entire corpora can be cheaply labeled? We investigate both by training supervised hate speech classifiers on a new dataset of 278K German political TikTok comments, comparing human and LLM annotation under matched conditions. LLM annotation at scale outperforms human-supervised classifiers at roughly one-tenth the cost, for both a closed-source (GPT-5.2) and an open-weight (Qwen3.5-122B-A10B) LLM, and the advantage is robust under soft-label evaluation. It hinges on the annotation interface: only a two-question decomposition mirroring the human annotation task unlocks it. AL provides no reliable advantage over random sampling in our prefiltered pool. Error structure depends on the LLM: only GPT-5.2 matches the human FP/FN balance, while other variants over-flag border-control and economic-competition discourse. Humans remain essential as evaluators; for training labels, the question shifts to which LLM, which interface, and what shape of pool.

cs.CL↗

arXiv2026

A Target-Centric Survey of Quantization-Aware Training

The rapid development of LLMs incurs prohibitive memory footprints and intensive computational demands. Quantization-Aware Training (QAT) techniques have emerged as a promising solution to address these challenges by explicitly simulating quantization effects during model training, yielding low-bit models that achieve accuracy comparable to their full-precision counterparts. In this work, we provide a target-centric survey of QAT, aimed at clarifying both its theoretical foundations and its evolving implementation landscape. We systematically review existing QAT methods through a target-centric taxonomy and synthesize cross-target differences in error characteristics, numerical formats, and strategy transferability. We further summarize QAT evaluation paradigms and discuss challenges in optimization and deployment, outlining potential directions for future research.

cs.LG↗

arXiv2026

Reasoning over Grammar: Can Synthetic Linguistic Reasoning Traces Enhance Low-Resource Machine Translation?

Large language models (LLMs) offer a promising approach to machine translation (MT) for extremely low-resource languages by incorporating linguistic resources through in-context learning. However, LLMs often struggle to apply grammatical information effectively during translation. Inspired by recent progress in chain-of-thought reasoning, we investigate whether low-resource MT can benefit from structured intermediate steps of linguistic analysis and grammatical reasoning. We propose a pipeline for automatically generating step-by-step linguistic reasoning traces from Universal Dependencies treebanks, dictionaries, and grammar-rule banks. We evaluate these traces in three settings: in-context learning (ICL), supervised fine-tuning (SFT), and reinforcement fine-tuning (RFT), on Xibe and Chintang as test cases. Our results show that linguistic reasoning traces are most effective as inference-time guidance: in ICL, reliable linguistic reasoning traces substantially improve translation performance across models, languages, and metrics. In contrast, using these traces as training data yields smaller and less consistent gains, as models learn the format but often generate error-prone content. These findings suggest that LLMs can leverage grammatical information for low-resource MT when given reliable linguistic analyses, while learning to generate such analyses remains a major bottleneck.

cs.CL↗