arXiv subjects

Carsten Eickhoff

Publications and source records attributed to Carsten Eickhoff.

At least 19 recordsLinked to original sources

arXiv2026

Mol-JEPA: A multimodal Joint Embedding Predictive Architecture for Molecules

Despite recent advances in molecular foundation models, several limitations remain, such as chemically invalid augmentations, modality collapse, and incomplete representation of biochemical environments. To address these challenges, we present \textbf{Mol-JEPA}, a scalable framework for learning molecular world models. Rather than relying on suboptimal molecular perturbations, our model uses modality masking to exploit information from molecular structures, cellular phenotypes, binding affinities, ADMET profiles, quantum chemistry simulations and other drug discovery data. Across various benchmarks, we show that the representations learned by Mol-JEPA deliver strong performance, demonstrating the value of incorporating biochemical context through latent space prediction.

cs.LG↗

arXiv2026

A Mechanistic Analysis of Gender Sensitivity in Dense Retrieval Models

While gender bias in dense retrieval models is well documented, with prior work showing that models often score male-gendered documents higher than female or neutral variants, the internal mechanisms producing these disparities are poorly understood. In this paper, we mechanistically analyze bi-encoder models to localize gender sensitivity, finding that the signal originates in input embeddings and propagates through a small set of late-layer attention heads that carry both gender and term-matching signals. Guided by these findings, we test steering interventions at both identified points and find distinct effects: embedding-level steering non-specifically neutralizes score differences, while attention-level steering produces directional shifts. Our findings provide a mechanistic basis for targeted debiasing and highlight the challenge of disentangling gender from relevance signals in shared model components.

cs.IR↗

arXiv2026

Isotropy Cliffs: The Geometric Signature of Decision-Making in Large Language Models

We investigate the geometry of decision-making in Multiple Choice Question Answering (MCQA) through the lens of isotropy. Analyzing five open-weight models across diverse datasets, we identify decision-critical transition layers characterized by a shift in isotropy, coinciding with a major representational change and the emergence of task-relevant clusters. We demonstrate that this synchronized geometric behavior is strongly correlated with downstream accuracy ($r\approx0.84$), displaying its relevance for successful decision-making. Furthermore, we show that this transition is robust to prompt variations, suggesting that it reflects a general mechanism of model behavior.

cs.AI↗

arXiv2026

Vision-Default, Prior-Override: Causal Mechanisms of Perception-Knowledge Conflict in Vision-Language Models

Vision-language models must reconcile visual evidence with memorized world knowledge when the two conflict. How they resolve this conflict shapes the reliability of multimodal systems, yet prior work characterizes it behaviorally without a component-level causal account. We combine activation patching across three granularities (residual stream, attention heads, and MLP sublayers) with model-component ablation studies and mechanistic analysis. Across three VLM families, we find that visual grounding emerges by default, whereas prior grounding depends on a small set of causally necessary attention heads (2.5-4.8%) concentrated in the second half of the network. These heads enable answers from stored world knowledge (e.g., "red" for a strawberry) despite conflicting visual input. Ablating them flips predictions from knowledge-grounded to visually grounded answers in 68-96% of cases under prior-knowledge prompts, but changes only 0.8-7.5% of visually grounded predictions, establishing an asymmetric causal structure. The identified heads decompose into routing heads, which modulate information flow, and writing heads, which directly project answer tokens into the residual stream. This structure is consistent across model families and scales, revealing a sparse causal circuit underlying perception-knowledge conflict in VLMs.

cs.CL↗

arXiv2026

MATCHA: Matching Text via Contrastive Semantic Alignment

Reliable evaluation is essential for understanding large language model (LLM) performance, yet today's go-to metrics, namely token-overlap scores (e.g., ROUGE) and embedding-based measures (e.g., BERTScore), often misjudge semantic similarity of documents. Our study shows that both token-overlap metrics and embedding-based metrics routinely assign nearly identical scores to texts that directly contradict each other, thereby potentially masking fundamental errors. We introduce MATCHA, an automatic metric that jointly rewards semantic agreement with a reference and penalizes contradictions. MATCHA employs a dual-view perspective that measures (i) proximity to the gold text and (ii) distance from an adversarially generated counterfactual contradiction. In eight public benchmarks, MATCHA outperforms popular metrics, compared with human annotations on question-answering, image caption generation, natural language inference, summarization, and semantic textual similarity tasks. On the TruthfulQA dataset (i.e., a dataset without a training set, where no embedding-based metrics could locally train on), this improvement in terms of matching texts with a reference reaches 18.38% over ROUGE-L and 20.82% over BERTScore. Both quantitative comparison and qualitative human assessments confirm the efficacy and validity of MATCHA and uncover fundamental weaknesses in pre-existing metrics. Compared with 23 embedding models, including top state-of-the-art ones, used as a metric similar to BERTScore, MATCHA remains the most accurate in distinguishing correct from incorrect statements solely based on a reference. Our code and metric are publicly available (https://github.com/Siran-Li/MATCHA).

cs.CL↗

arXiv2026

Understanding Wacky Weights: A Dissection of SPLADE's Learned Term Importance

Learned sparse retrieval models such as SPLADE combine the effectiveness of neural architectures with the efficiency of inverted indices. As these models assign weights to terms from a fixed vocabulary, interpretability is often touted as a major benefit of these models. However, the emergence of wacky weights, i.e., expansion terms that appear semantically unrelated to the input, limits interpretability. While prior research has anecdotally observed this phenomenon, there is a lack of systematic understanding regarding their origins, prevalence, and contribution to retrieval effectiveness. In this paper, we reproduce SPLADE-v2 to systematically investigate wacky weights across the SPLADE family of models. We present a comprehensive dissection of wacky weights, providing a formal definition of wackiness based on the lexical utility of expansion terms. Furthermore, we introduce a novel measure to compare the prevalence of these tokens across models with varying vocabularies and sparsity levels. Beyond reproducing the original SPLADE-v2, we train it with various loss functions, datasets, and backbone transformers to isolate the factors contributing to wackiness. Our results show that larger vocabularies are associated with a higher prevalence of wacky tokens, while stricter sparsity regularizers are associated with lower prevalence. Finally, we find that wacky weights are used primarily for in-domain effectiveness rather than out-of-domain generalization.

cs.IR↗

arXiv2026

Scientific Domain Knowledge Improves Vision-Language Fundus Models

Vision-language models hold considerable promise for ophthalmology, but it remains unclear which training data source best conveys expert domain knowledge. Existing ophthalmic models are trained on fixed text templates, medical reports, or general biomedical literature, sources that have never been compared under matched conditions. To include domain-specific literature in this comparison, we present PubMed-Ophtha, a hierarchical dataset with high domain density of 102,023 panels with their subcaptions from 15,842 open-access articles in PubMed Central. We then finetuned identical CLIP models on each source, using a general biomedical literature model as baseline, and found that domain-specific literature achieved the best average performance across 110 clinical tasks, reaching a mean linear probing AUROC of 88.63% ahead of medical reports (85.68%). Restricting the dataset to fundus images, to the image count of the medical report dataset, or to articles unrelated to the evaluation datasets did not reduce performance, indicating that the gains likely stem from domain density. We release the dataset, the finetuned models, and the full generation pipeline.

cs.CV↗

arXiv2026

Is There Knowledge Left to Extract? Evidence of Fragility in Medically Fine-Tuned Vision-Language Models

Vision-language models (VLMs) are increasingly adapted through domain-specific fine-tuning, yet it remains unclear whether this improves reasoning beyond superficial visual cues, particularly in high-stakes domains like medicine. We evaluate four paired open-source VLMs (LLaVA vs. LLaVA-Med; Gemma vs. MedGemma) across four medical imaging tasks of increasing difficulty: brain tumor, pneumonia, skin cancer, and histopathology classification. We find that performance degrades toward near-random levels as task difficulty increases, indicating limited clinical reasoning. Medical fine-tuning provides no consistent advantage, and models are highly sensitive to prompt formulation, with minor changes causing large swings in accuracy and refusal rates. To test whether closed-form VQA suppresses latent knowledge, we introduce a description-based pipeline where models generate image descriptions that a text-only model (GPT-5.1) uses for diagnosis. This recovers a limited additional signal but remains bounded by task difficulty. Analysis of vision encoder embeddings further shows that failures stem from both weak visual representations and downstream reasoning. Overall, medical VLM performance is fragile, prompt-dependent, and not reliably improved by domain-specific fine-tuning.

cs.CV↗

arXiv2026

When to Call an Apple Red: Humans Follow Introspective Rules, VLMs Don't

Understanding when Vision-Language Models (VLMs) will behave unexpectedly, whether models can reliably predict their own behavior, and if models adhere to their introspective reasoning are central challenges for trustworthy deployment. To study this, we introduce the Graded Color Attribution (GCA) dataset, a controlled benchmark designed to elicit decision rules and evaluate participant faithfulness to these rules. GCA consists of line drawings that vary pixel-level color coverage across three conditions: world-knowledge recolorings, counterfactual recolorings, and shapes with no color priors. Using GCA, we ask both VLMs and human participants to state a threshold rule: the share of an object's pixels that must be a given color for the object to receive that color label. We then compare these rules with their subsequent color attribution decisions. Our findings reveal that models systematically violate their own introspective rules. For example, GPT-5-mini violates its stated introspection rules in nearly 60% of cases on objects with strong color priors. Human participants remain faithful to their stated rules, with any apparent violations being explained by a well-documented tendency to overestimate color coverage. In contrast, we find that VLMs can accurately estimate color coverage, yet directly contradict their own reasoning in their final responses. Across all models and strategies for eliciting introspective rules, world-knowledge priors systematically degrade faithfulness in ways that do not mirror human cognition. Our findings challenge the view that VLM reasoning failures are difficulty-driven and suggest that VLM introspective self-knowledge is miscalibrated, with direct implications for high-stakes deployment.

cs.CL↗

arXiv2026

When Silence Is Golden: Can LLMs Learn to Abstain in Temporal QA and Beyond?

Large language models (LLMs) rarely admit uncertainty, often producing fluent but misleading answers, rather than abstaining (i.e., refusing to answer). This weakness is even evident in temporal question answering, where models frequently ignore time-sensitive evidence and conflate facts across different time-periods. In this paper, we present the first empirical study of training LLMs with an abstention ability while reasoning about temporal QA. Existing approaches such as calibration might be unreliable in capturing uncertainty in complex reasoning. We instead frame abstention as a teachable skill and introduce a pipeline that couples Chain-of-Thought (CoT) supervision with Reinforcement Learning (RL) guided by abstention-aware rewards. Our goal is to systematically analyze how different information types and training techniques affect temporal reasoning with abstention behavior in LLMs. Through extensive experiments studying various methods, we find that RL yields strong empirical gains on reasoning: a model initialized by Qwen2.5-1.5B-Instruct surpasses GPT-4o by $3.46\%$ and $5.80\%$ in Exact Match on TimeQA-Easy and Hard, respectively. Moreover, it improves the True Positive rate on unanswerable questions by $20\%$ over a pure supervised fine-tuned (SFT) variant. Beyond performance, our analysis shows that SFT induces overconfidence and harms reliability, while RL improves prediction accuracy but exhibits similar risks. Finally, by comparing implicit reasoning cues (e.g., original context, temporal sub-context, knowledge graphs) with explicit CoT supervision, we find that implicit information provides limited benefit for reasoning with abstention. Our study provides new insights into how abstention and reasoning can be jointly optimized, providing a foundation for building more reliable LLMs.

cs.CL↗

arXiv2026

UbuntuGuard: A Culturally-Grounded Policy Benchmark for Equitable AI Safety in African Languages

Current guardian models are predominantly Western-centric and optimized for high-resource languages, leaving low-resource African languages vulnerable to evolving harms, cross-lingual failures, and cultural misalignment. Moreover, most guardian models rely on rigid, predefined safety categories that fail to generalize across diverse linguistic and sociocultural contexts. Achieving robust safety requires flexible, runtime-enforceable policies and benchmarks that reflect local norms, harm scenarios, and cultural expectations. We introduce UbuntuGuard, the first policy-based safety benchmark for African languages built from adversarial queries authored by 155 domain experts across sensitive fields, including healthcare. From these expert-crafted queries, we derive context-specific safety policies and reference responses that capture culturally grounded risk signals, enabling policy-aligned evaluation of guardian models. We evaluate 15 models, comprising seven general-purpose LLMs and eight guardian models across three distinct variants: static, dynamic, and multilingual. Our findings reveal that existing English-centric benchmarks overestimate real-world multilingual safety, cross-lingual transfer provides partial but insufficient coverage, and dynamic models, while better equipped to leverage policies at inference time, still struggle to fully localize African-language contexts. These findings highlight the urgent need for multilingual, culturally grounded safety benchmarks to enable the development of reliable and equitable guardian models for low-resource languages.

cs.CL↗

arXiv2026

Less is More: Label-Guided Summarization of Procedural and Instructional Videos

Video summarization helps turn long videos into clear, concise representations that are easier to review, document, and analyze, especially in high-stakes domains like surgical training. Prior work has progressed from using basic visual features like color, motion, and structural changes to using pre-trained vision-language models that can better understand what's happening in the video (semantics) and capture temporal flow, resulting in more context-aware video summarization. We propose a three-stage framework, PRISM: Procedural Representation via Integrated Semantic and Multimodal analysis, that produces semantically grounded video summaries. PRISM combines adaptive visual sampling, label-driven keyframe anchoring, and contextual validation using a large language model (LLM). Our method ensures that selected frames reflect meaningful and procedural transitions while filtering out generic or hallucinated content, resulting in contextually coherent summaries across both domain-specific and instructional videos. We evaluate our method on instructional and activity datasets, using reference summaries for instructional videos. Despite sampling fewer than 5% of the original frames, our summaries retain 84% semantic content while improving over baselines by as much as 33%. Our approach generalizes across procedural and domain-specific video tasks, achieving strong performance with both semantic alignment and precision.

cs.CV↗

arXiv2026

Mechanisms of Prompt-Induced Hallucination in Vision-Language Models

Large vision-language models (VLMs) are highly capable, yet often hallucinate by favoring textual prompts over visual evidence. We study this failure mode in a controlled object-counting setting, where the prompt overstates the number of objects in the image (e.g., asking a model to describe four waterlilies when only three are present). At low object counts, models often correct the overestimation, but as the number of objects increases, they increasingly conform to the prompt regardless of the discrepancy. Through mechanistic analysis of three VLMs, we identify a small set of attention heads whose ablation substantially reduces prompt-induced hallucinations (PIH) by at least 40% without additional training. Across models, PIH-heads mediate prompt copying in model-specific ways. We characterize these differences and show that PIH ablation increases correction toward visual evidence. Our findings offer insights into the internal mechanisms driving prompt-induced hallucinations, revealing model-specific differences in how these behaviors are implemented.

cs.CV↗

arXiv2026

The Persona Paradox: Medical Personas as Behavioral Priors in Clinical Language Models

Persona conditioning can be viewed as a behavioral prior for large language models (LLMs) and is often assumed to confer expertise and improve safety in a monotonic manner. However, its effects on high-stakes clinical decision-making remain poorly characterized. We systematically evaluate persona-based control in clinical LLMs, examining how professional roles (e.g., Emergency Department physician, nurse) and interaction styles (bold vs.\ cautious) influence behavior across models and medical tasks. We assess performance on clinical triage and patient-safety tasks using multidimensional evaluations that capture task accuracy, calibration, and safety-relevant risk behavior. We find systematic, context-dependent, and non-monotonic effects: Medical personas improve performance in critical care tasks, yielding gains of up to $\sim+20\%$ in accuracy and calibration, but degrade performance in primary-care settings by comparable margins. Interaction style modulates risk propensity and sensitivity, but it's highly model-dependent. While aggregated LLM-judge rankings favor medical over non-medical personas in safety-critical cases, we found that human clinicians show moderate agreement on safety compliance (average Cohen's $\kappa = 0.43$) but indicate a low confidence in 95.9\% of their responses on reasoning quality. Our work shows that personas function as behavioral priors that introduce context-dependent trade-offs rather than guarantees of safety or expertise. The code is available at https://github.com/rsinghlab/Persona\_Paradox.

cs.AI↗

arXiv2025

From Topology to Retrieval: Decoding Embedding Spaces with Unified Signatures

Studying how embeddings are organized in space not only enhances model interpretability but also uncovers factors that drive downstream task performance. In this paper, we present a comprehensive analysis of topological and geometric measures across a wide set of text embedding models and datasets. We find a high degree of redundancy among these measures and observe that individual metrics often fail to sufficiently differentiate embedding spaces. Building on these insights, we introduce Unified Topological Signatures (UTS), a holistic framework for characterizing embedding spaces. We show that UTS can predict model-specific properties and reveal similarities driven by model architecture. Further, we demonstrate the utility of our method by linking topological structure to ranking effectiveness and accurately predicting document retrievability. We find that a holistic, multi-attribute perspective is essential to understanding and leveraging the geometry of text embeddings.

cs.LG↗

arXiv2025

APP: Accelerated Path Patching with Task-Specific Pruning

Circuit discovery is a key step in many mechanistic interpretability pipelines. Current methods, such as Path Patching, are computationally expensive and have limited in-depth circuit analysis for smaller models. In this study, we propose Accelerated Path Patching (APP), a hybrid approach leveraging our novel contrastive attention head pruning method to drastically reduce the search space of circuit discovery methods. Our Contrastive-FLAP pruning algorithm uses techniques from causal mediation analysis to assign higher pruning scores to task-specific attention heads, leading to higher performing sparse models compared to traditional pruning techniques. Although Contrastive-FLAP is successful at preserving task-specific heads that existing pruning algorithms remove at low sparsity ratios, the circuits found by Contrastive-FLAP alone are too large to satisfy the minimality constraint required in circuit analysis. APP first applies Contrastive-FLAP to reduce the search space on required for circuit discovery algorithms by, on average, 56\%. Next, APP, applies traditional Path Patching on the remaining attention heads, leading to a speed up of 59.63\%-93.27\% compared to Path Patching applied to the dense model. Despite the substantial computational saving that APP provides, circuits obtained from APP exhibit substantial overlap and similar performance to previously established Path Patching circuits

cs.LG↗

arXiv2025

Benchmarking is Broken -- Don't Let AI be its Own Judge

The meteoric rise of AI, with its rapidly expanding market capitalization, presents both transformative opportunities and critical challenges. Chief among these is the urgent need for a new, unified paradigm for trustworthy evaluation, as current benchmarks increasingly reveal critical vulnerabilities. Issues like data contamination and selective reporting by model developers fuel hype, while inadequate data quality control can lead to biased evaluations that, even if unintentionally, may favor specific approaches. As a flood of participants enters the AI space, this "Wild West" of assessment makes distinguishing genuine progress from exaggerated claims exceptionally difficult. Such ambiguity blurs scientific signals and erodes public confidence, much as unchecked claims would destabilize financial markets reliant on credible oversight from agencies like Moody's. In high-stakes human examinations (e.g., SAT, GRE), substantial effort is devoted to ensuring fairness and credibility; why settle for less in evaluating AI, especially given its profound societal impact? This position paper argues that the current laissez-faire approach is unsustainable. We contend that true, sustainable AI advancement demands a paradigm shift: a unified, live, and quality-controlled benchmarking framework robust by construction, not by mere courtesy and goodwill. To this end, we dissect the systemic flaws undermining today's AI evaluation, distill the essential requirements for a new generation of assessments, and introduce PeerBench (with its prototype implementation at https://www.peerbench.ai/), a community-governed, proctored evaluation blueprint that embodies this paradigm through sealed execution, item banking with rolling renewal, and delayed transparency. Our goal is to pave the way for evaluations that can restore integrity and deliver genuinely trustworthy measures of AI progress.

cs.AI↗

arXiv2025

PiCME: Pipeline for Contrastive Modality Evaluation and Encoding in the MIMIC Dataset

Multimodal deep learning holds promise for improving clinical prediction by integrating diverse patient data, including text, imaging, time-series, and structured demographics. Contrastive learning facilitates this integration by producing a unified representation that can be reused across tasks, reducing the need for separate models or encoders. Although contrastive learning has seen success in vision-language domains, its use in clinical settings remains largely limited to image and text pairs. We propose the Pipeline for Contrastive Modality Evaluation and Encoding (PiCME), which systematically assesses five clinical data types from MIMIC: discharge summaries, radiology reports, chest X-rays, demographics, and time-series. We pre-train contrastive models on all 26 combinations of two to five modalities and evaluate their utility on in-hospital mortality and phenotype prediction. To address performance plateaus with more modalities, we introduce a Modality-Gated LSTM that weights each modality according to its contrastively learned importance. Our results show that contrastive models remain competitive with supervised baselines, particularly in three-modality settings. Performance declines beyond three modalities, which supervised models fail to recover. The Modality-Gated LSTM mitigates this drop, improving AUROC from 73.19% to 76.93% and AUPRC from 51.27% to 62.26% in the five-modality setting. We also compare contrastively learned modality importance scores with attribution scores and evaluate generalization across demographic subgroups, highlighting strengths in interpretability and fairness. PiCME is the first to scale contrastive learning across all modality combinations in MIMIC, offering guidance for modality selection, training strategies, and equitable clinical prediction.

cs.LG↗