Searcharxiv⌕ Search

arXiv subjects

Jan Philip Wahle

Publications and source records attributed to Jan Philip Wahle.

At least 19 recordsLinked to original sources
arXiv2026

Is Visual Prompting All You Need? Studying VLM Spatial Reasoning under Progressive Visual Scaffolds

Vision-language models (VLMs) have advanced rapidly in multimodal reasoning, yet recent work shows that their failures often reflect an interaction between visual grounding and downstream reasoning. What remains less clear is how the visual presentation of a task shapes model performance and failure modes when the underlying reasoning problem is unchanged. We study this question in SPaRC, a benchmark for grid-based visual spatial planning, by introducing lightweight input-side scaffolds that preserve the visual modality while making spatial structure more accessible. Across multiple VLMs, these scaffolds improve task accuracy over the original visual setting by up to 34.0 percentage points and further complement GRPO-based training, yielding up to 4.6 additional accuracy points compared with near-zero gains on the original visual input. Analyses on both end-to-end task solving and object detection show that these gains are closely tied to reductions in grounding-related errors, while rule reasoning remains comparatively challenging. We find that visual presentation is a central factor that determines whether VLM benchmarks measure grounded perception, downstream reasoning, or a mixture of both.

cs.CV↗
arXiv2026

BabelSteering: Multilingual Safety Alignment via English Steering Vectors

Large language models (LLMs) are deployed globally in high-stakes settings, yet most safety research and alignment efforts remain concentrated on English. Thus, users interacting with LLMs in other languages may encounter weaker safeguards despite relying on the same systems for similarly sensitive tasks. In this work, we investigate whether safety signals learned from a high-resource language, like English, can improve multilingual safety. We propose BabelSteering, an activation steering method that acts as a lightweight inference- time intervention, using refusal directions derived from English safety supervision to generalize across languages. Our evaluation includes eight languages and jointly measures refusal of harmful requests, over-refusal, and general task utility. The results show that BabelSteering increases the refusal of harmful requests across languages, with only a marginal to no reduction in task utility but with some increase in refusal of pseudo-harmful prompts. For example, for Gemma 7B, we see an average increase in the refusal of harmful prompts across languages of 11 percentage points (pp), with individual languages like Bengali seeing an increase of 17 pp, with no loss of utility on Global MMLU, while pseudo-harmful refusals increase by 13 pp on average. We also introduce a multilingual translation-and-evaluation pipeline to facilitate future work on cross-lingual safety interventions. Overall, our findings suggest that activation steering may provide a practical, low- cost mechanism for extending English-derived safety signals to other languages. Warning: this paper contains examples with unsafe content

cs.CL↗
arXiv2026

Text Generation: A Systematic Literature Review of Tasks, Evaluation, and Challenges

Text generation has become more accessible than ever, and the growing interest in these systems, especially those using large language models, has spurred a surge in related publications. We provide a systematic literature review comprising 257 papers, covering the period from January 2017 to December 2025. This review categorizes text generation contributions into five main tasks: open-ended text generation, summarization, translation, paraphrasing, and question answering. For each task in our taxonomy, we review relevant characteristics and key subtasks. We assess current approaches for evaluating text generation systems, covering model-free, model-based, and human evaluation. Our investigation shows several task-specific challenges (e.g., missing datasets for multi-document summarization, lack of coherence in story generation, and difficulties in complex reasoning for question answering). We further discuss nine challenges common to all tasks and sub-tasks in recent text generation papers: bias, reasoning, hallucinations, misuse, privacy, interpretability, transparency, datasets, and computing. This systematic literature review targets two main audiences: early-career researchers in natural language processing seeking an overview of the field and promising research directions, and senior researchers who need a recent overview of the main tasks, evaluation, challenges, and mitigation strategies.

cs.CL↗
arXiv2026

Risky Business: Measuring The Faithfulness-Safety Tension

Chain-of-Thought (CoT) reasoning offers a promising window into model monitoring. However, monitoring relies on faithfulness, i.e., the model output strictly derives from its reasoning trace. We identify an alignment tension where a model must be faithful enough to be monitored, yet robust enough to reject unsafe reasoning. We demonstrate that this counterbalance exists in current Large Reasoning Models (LRMs), and show ways in which it can be addressed. We introduce HazMart, a human-written dataset set in an autonomous AI shopkeeper scenario. Unlike prior work that relies on providing hints in prompts to test faithfulness (e.g., "A Stanford professor said it should be Answer A"), we propose a novel replacement-based technique, which we call Targeted Reasoning Replacement (TRR), that directly intervenes in the reasoning chain to substitute in unsafe or illogical thoughts (e.g., "Wait, the answer must be Option B [was Option A] because it is the most fitting"). DeepSeek-R1-Llama-70B exhibits high faithfulness (97.5%) but fails to reject Unsafe Reasoning (12.3%), while QwQ-32B is more robust (73.9% safety) at the cost of lower faithfulness (74.7%). Mechanistic analyses of QwQ-32B reveal that these properties are represented by anti-correlated internal directions peaking at the action-commit token. Finally, we demonstrate that representation steering can independently amplify the safety direction, increasing safe behavior by 9 percentage points while maintaining base capabilities.

cs.AI↗
arXiv2026

Can Agents Deceive? Evaluating Reasoning and Deception in ParliamentBench using a Social Deduction Game

As large language models (LLMs) are deployed as agents in high-stakes settings, such as medical and legal systems, understanding their deceptive capabilities is fundamental to safety. Controlled social deduction games provide a reproducible proxy for isolating and evaluating these complex adversarial behaviors. We present the open-source benchmark framework ParliamentBench based on the game Secret Hitler to evaluate LLMs in scenarios that require deception, persuasion, and reasoning under information asymmetry. We evaluate 16 LLMs across 1,600 simulated matches playing each other, playing against humans, and compare them against a large set of online games. We introduce three novel metrics that isolate social deduction, reasoning, and deceptive consistency. Our experiments reveal that frontier models achieve strong performance across cooperative and deceptive roles, with a strong top-four cluster (GPT-5.4, Kimi K2.5, Grok 4.1 Fast, and DeepSeek 3.1 Terminus), whereas the weakest models fall short of random (33%) and simple algorithmic (45%) baselines. Most LLMs struggle to maintain a consistent deceptive persona throughout an entire game, with deception retention dropping below 50%.

cs.CL↗
arXiv2026

Misinformation Propagation in Benign Multi-Agent Systems

Multi-agent systems, in which multiple large language model agents solve problems through turn-based interaction, are increasingly deployed in high-stakes settings such as medical diagnosis, legal analysis, and forensic decision-making. Their reliability can be at risk when single agents reason from incorrect or misleading context, e.g., from tool calls, since errors may propagate through agent interactions. This work studies this risk by injecting intent-based misinformation into benign single-agent and multi-agent systems across reasoning, knowledge, and alignment tasks. We find that misinformation can degrade single-agent performance and persists across multi-agent debate, with agents often retaining answers introduced by misinformed peers. Nevertheless, multi-agent debate reduces the resulting performance degradation compared to single-agent prompting, especially when most agents are not exposed to misinformation. Robustness depends on group composition and decision protocol. Consensus can be more stable than voting under peer pressure, while majorities can often steer misinformed agents back toward correct answers. Our results show that misinformation robustness in multi-agent systems depends on the underlying model and also on how agents exchange information and aggregate decisions.

cs.MA↗
arXiv2026

Multi-Agent Reasoning Improves Compute Efficiency: Pareto-Optimal Test-Time Scaling

Advances in inference methods have enabled language models to improve their predictions without additional training. These methods often prioritize raw performance over cost-effective compute usage. However, computational efficiency is key for real-world applications with resource constraints. We provide a systematic analysis of the inference scaling strategies self-consistency, self-refinement, multi-agent debate, and mixture-of-agents, to study their computational performance tradeoffs. We evaluate methods on two reasoning benchmarks (MMLU-Pro, BBH) and include extensive parameter configurations (e.g., scaling the number of parallel predictions, agents, and debate rounds) across different model sizes. Across 34 configurations and over 100 evaluations, we compute the Pareto-optimal front to select methods that achieve the best accuracy with the lowest computational budget. Notably, inference scaling improves accuracy by up to +7.1% points over chain-of-thought at the highest evaluated budgets (20x the CoT compute budget) on MMLU-Pro. With an equal computing budget, debate and mixture-of-agents outperform self-consistency by 1.3% and 2.7% points, respectively. While self-consistency saturates earlier, multi-agent gains persist, particularly on more complicated tasks. We identify a simple multi-agent design guideline: mixture-of-agents is most efficient when the number of parallel generations exceeds the number of sequential aggregations.

cs.AI↗
arXiv2026

DimABSA: Building Multilingual and Multidomain Datasets for Dimensional Aspect-Based Sentiment Analysis

Aspect-Based Sentiment Analysis (ABSA) focuses on extracting sentiment at a fine-grained aspect level and has been widely applied across real-world domains. However, existing ABSA research relies on coarse-grained categorical labels (e.g., positive, negative), which limits its ability to capture nuanced affective states. To address this limitation, we adopt a dimensional approach that represents sentiment with continuous valence-arousal (VA) scores, enabling fine-grained analysis at both the aspect and sentiment levels. To this end, we introduce DimABSA, the first multilingual, dimensional ABSA resource annotated with both traditional ABSA elements (aspect terms, aspect categories, and opinion terms) and newly introduced VA scores. This resource contains 76,958 aspect instances across 42,590 sentences, spanning six languages and four domains. We further introduce three subtasks that combine VA scores with different ABSA elements, providing a bridge from traditional ABSA to dimensional ABSA. Given that these subtasks involve both categorical and continuous outputs, we propose a new unified metric, continuous F1 (cF1), which incorporates VA prediction error into standard F1. We provide a comprehensive benchmark using both prompted and fine-tuned large language models across all subtasks. Our results show that DimABSA is a challenging benchmark and provides a foundation for advancing multilingual dimensional ABSA. We publicly released the DimABSA dataset, which was used for Track A of SemEval-2026 Task 3, attracting over 300 participants.

cs.CL↗
arXiv2026

Who Watches the Watchmen? Humans Disagree With Translation Metrics on Unseen Domains

Automatic evaluation metrics are central to the development of machine translation systems, yet their robustness under domain shift remains unclear. Most metrics are developed on the Workshop on Machine Translation (WMT) benchmarks, raising concerns about their robustness to unseen domains. Prior studies that analyze unseen domains vary translation systems, annotators, or evaluation conditions, confounding domain effects with human annotation noise. To address these biases, we introduce a systematic multi-annotator Cross-Domain Error-Span-Annotation dataset (CD-ESA), comprising 18.8k human error span annotations across three language pairs, where we fix annotators within each language pair and evaluate translations of the same six translation systems across one seen news domain and two unseen technical domains. Using this dataset, we first find that automatic metrics appear surprisingly robust to domain-shifts at the segment level (up to 0.69 agreement), but this robustness largely disappears once we account for human label variation. Averaging annotations increases inter-annotator agreement by up to +0.11. Metrics struggle on the unseen chemical domain compared to humans (inter-annotator agreement of 0.78-0.83 vs. 0.96). We recommend comparing metric-human agreement against inter-annotator agreement, rather than comparing raw metric-human agreement alone, when evaluating across different domains.

cs.CL↗
arXiv2026

Mind the Gap Between Spatial Reasoning and Acting! Step-by-Step Evaluation of Agents With Spatial-Gym

Spatial reasoning is central to navigation and robotics, yet measuring model capabilities on these tasks remains difficult. Existing benchmarks evaluate models in a one-shot setting, requiring full solution generation in a single response, unlike humans, who work in interactive environments step-by-step. We introduce Spatial-Gym, a Gymnasium environment that isolates spatial constraint reasoning by testing pathfinding in 2D-grid puzzles as a sequential decision task with optional backtracking. We evaluate eight models in three settings (one-shot, step-by-step, step-by-step with backtracking) against human, random, and A* baselines on 500 episodes. The best model, GPT-OSS 120B, achieves a solve rate of 16.0%, 82 points below the human baseline (98.0%). Step-by-step format helps weaker models (up to +5.4%) by removing formatting errors, but hurts stronger models (up to 5.6%) by constraining global planning. Backtracking improves episode completion, but increases solve rate only for weaker models; stronger models rarely backtrack and do not benefit from it. Our experiments have three key findings: (1) models fail to scale reasoning effort with difficulty, (2) vision models receiving images of the spatial environment reduce solve rate by 73%, and (3) extended chain-of-thought reasoning retains a 3-5x accuracy advantage over standard inference even in the step-by-step setting. Spatial-Gym enables diagnosis of model limitations and provides a framework for improving spatial reasoning through reinforcement learning.

cs.AI↗
arXiv2026

Stay Focused: Problem Drift in Multi-Agent Debate

Multi-agent debate - multiple instances of large language models discussing problems in turn-based interaction - has shown promise for solving knowledge and reasoning tasks. However, these methods show limitations when solving complex problems that require longer reasoning chains. We analyze how multi-agent debate drifts away from the initial problem over multiple turns, thus harming task performance. We define this phenomenon as problem drift and quantify its presence across ten tasks (i.e., three generative, three knowledge, three reasoning, and one instruction-following task). We find that generative tasks drift often due to the subjectivity of the answer space (76-89%), compared to high-complexity tasks (7-21%). To identify the reasons, eight human experts analyze 170 multi-agent debates suffering from problem drift. We find the most common issues related to this drift are the lack of progress (35% of cases), low-quality feedback (26% of cases), and a lack of clarity (25% of cases). We propose DRIFTJudge, an LLM-as-a-judge method, as a first baseline to detect problem drift. We also propose DRIFTPolicy, which mitigates 31% of problem drift cases. Our study is a step toward understanding a key limitation of multi-agent debate, highlighting why longer debates can harm task performance and how problem drift could be addressed.

cs.CL↗
arXiv2026

SemEval-2026 Task 3: Dimensional Aspect-Based Sentiment Analysis (DimABSA)

We present the SemEval-2026 shared task on Dimensional Aspect-Based Sentiment Analysis (DimABSA), which improves traditional ABSA by modeling sentiment along valence-arousal (VA) dimensions rather than using categorical polarity labels. To extend ABSA beyond consumer reviews to public-issue discourse (e.g., political, energy, and climate issues), we introduce an additional task, Dimensional Stance Analysis (DimStance), which treats stance targets as aspects and reformulates stance detection as regression in the VA space. The task consists of two tracks: Track A (DimABSA) and Track B (DimStance). Track A includes three subtasks: (1) dimensional aspect sentiment regression, (2) dimensional aspect sentiment triplet extraction, and (3) dimensional aspect sentiment quadruplet extraction, while Track B includes only the regression subtask for stance targets. We also introduce a continuous F1 (cF1) metric to jointly evaluate structured extraction and VA regression. The task attracted more than 400 participants, resulting in 112 final submissions and 42 system description papers. We report baseline results, discuss top-performing systems, and analyze key design choices to provide insights into dimensional sentiment analysis at the aspect and stance-target levels. All resources are available on our GitHub repository.

cs.CL↗
arXiv2026

Affect, Body, Cognition, Demographics, and Emotion: The ABCDE of Text Features for Computational Affective Science

Work in Computational Affective Science and Computational Social Science explores a wide variety of research questions about people, emotions, behavior, and health. Such work often relies on language data that is first labeled with relevant information, such as the use of emotion words or the age of the speaker. Although many resources and algorithms exist to enable this type of labeling, discovering, accessing, and using them remains a substantial impediment, particularly for practitioners outside of computer science. Here, we present the ABCDE dataset (Affect, Body, Cognition, Demographics, and Emotion), a large-scale collection of over 400 million text utterances drawn from social media, blogs, books, and AI-generated sources. The dataset is annotated with a wide range of features relevant to computational affective and social science. ABCDE facilitates interdisciplinary research across numerous fields, including affective science, cognitive science, the digital humanities, sociology, political science, and computational linguistics.

cs.CL↗
arXiv2026

Piecing Together Cross-Document Coreference Resolution Datasets: Systematic Dataset Analysis and Unification

Research in CDCR remains fragmented due to heterogeneous dataset formats, varying annotation standards, and the predominance of the CDCR definition as the event coreference resolution (ECR). To address these challenges, we introduce uCDCR, a unified dataset that consolidates diverse publicly available English CDCR corpora across various domains into a consistent format, which we analyze with standardized metrics and evaluation protocols. uCDCR incorporates both entity and event coreference, corrects known inconsistencies, and enriches datasets with missing attributes to facilitate reproducible research. We establish a cohesive framework for fair, interpretable, and cross-dataset analysis in CDCR and compare the datasets on their lexical properties, e.g., lexical composition of the annotated mentions, lexical diversity and ambiguity metrics, discuss the annotation rules and principles that lead to high lexical diversity, and examine how these metrics influence performance on the same-head-lemma baseline. Our dataset analysis shows that ECB+, the state-of-the-art benchmark for CDCR, has one of the lowest lexical diversities, and its CDCR complexity, measured by the same-head-lemma baseline, lies in the middle among all uCDCR datasets. Moreover, comparing document and mention distributions between ECB+ and uCDCR shows that using all uCDCR datasets for model training and evaluation will improve the generalizability of CDCR models. Finally, the almost identical performance on the same-head-lemma baseline, separately applied to events and entities, shows that resolving both types is a complex task and should not be steered toward ECR alone. The uCDCR dataset is available at https://huggingface.co/datasets/AnZhu/uCDCR, and the code for parsing, analyzing, and scoring the dataset is available at https://github.com/anastasia-zhukova/uCDCR.

cs.CL↗
arXiv2026

Language Modeling and Understanding Through Paraphrase Generation and Detection

Language enables humans to share knowledge, reason about the world, and pass on strategies for survival and innovation across generations. At the heart of this process is not just the ability to communicate but also the remarkable flexibility in how we can express ourselves. We can express the same thoughts in virtually infinite ways using different words and structures - this ability to rephrase and reformulate expressions is known as paraphrase. Modeling paraphrases is a keystone to meaning in computational language models; being able to construct different variations of texts that convey the same meaning or not shows strong abilities of semantic understanding. If computational language models are to represent meaning, they must understand and control the different aspects that construct the same meaning as opposed to different meanings at a fine granularity. Yet most existing approaches reduce paraphrasing to a binary decision between two texts or to producing a single rewrite of a source, obscuring which linguistic factors are responsible for meaning preservation. In this thesis, I propose that decomposing paraphrases into their constituent linguistic aspects (paraphrase types) offers a more fine-grained and cognitively grounded view of semantic equivalence. I show that even advanced machine learning models struggle with this task. Yet, when explicitly trained on paraphrase types, models achieve stronger performance on related paraphrase tasks and downstream applications. For example, in plagiarism detection, language models trained on paraphrase types surpass human baselines: 89.6% accuracy compared to 78.4% for plagiarism cases from Wikipedia, and 66.5% compared to 55.7% for plagiarism of scientific papers from arXiv. In identifying duplicate questions on Quora, models trained with paraphrase types improve over models trained on binary pairs. Furthermore, I demonstrate that...

cs.CL↗
arXiv2026

Paraphrase Types Elicit Prompt Engineering Capabilities

Much of the success of modern language models depends on finding a suitable prompt to instruct the model. Until now, it has been largely unknown how variations in the linguistic expression of prompts affect these models. This study systematically and empirically evaluates which linguistic features influence models through paraphrase types, i.e., different linguistic changes at particular positions. We measure behavioral changes for five models across 120 tasks and six families of paraphrases (i.e., morphology, syntax, lexicon, lexico-syntax, discourse, and others). We also control for other prompt engineering factors (e.g., prompt length, lexical diversity, and proximity to training data). Our results show a potential for language models to improve tasks when their prompts are adapted in specific paraphrase types (e.g., 6.7% median gain in Mixtral 8x7B; 5.5% in LLaMA 3 8B). In particular, changes in morphology and lexicon, i.e., the vocabulary used, showed promise in improving prompts. These findings contribute to developing more robust language models capable of handling variability in linguistic expression.

cs.CL↗
arXiv2026

DimStance: Multilingual Datasets for Dimensional Stance Analysis

Stance detection is an established task that classifies an author's attitude toward a specific target into categories such as Favor, Neutral, and Against. Beyond categorical stance labels, we leverage a long-established affective science framework to model stance along real-valued dimensions of valence (negative-positive) and arousal (calm-active). This dimensional approach captures nuanced affective states underlying stance expressions, enabling fine-grained stance analysis. To this end, we introduce DimStance, the first dimensional stance resource with valence-arousal (VA) annotations. This resource comprises 11,746 target aspects in 7,365 texts across five languages (English, German, Chinese, Nigerian Pidgin, and Swahili) and two domains (politics and environmental protection). To facilitate the evaluation of stance VA prediction, we formulate the dimensional stance regression task, analyze cross-lingual VA patterns, and benchmark pretrained and large language models under regression and prompting settings. Results show competitive performance of fine-tuned LLM regressors, persistent challenges in low-resource languages, and limitations of token-based generation. DimStance provides a foundation for multilingual, emotion-aware, stance analysis and benchmarking.

cs.CL↗
arXiv2026

TrojanStego: Your Language Model Can Secretly Be A Steganographic Privacy Leaking Agent

As large language models (LLMs) become integrated into sensitive workflows, concerns grow over their potential to leak confidential information. We propose TrojanStego, a novel threat model in which an adversary fine-tunes an LLM to embed sensitive context information into natural-looking outputs via linguistic steganography, without requiring explicit control over inference inputs. We introduce a taxonomy outlining risk factors for compromised LLMs, and use it to evaluate the risk profile of the threat. To implement TrojanStego, we propose a practical encoding scheme based on vocabulary partitioning learnable by LLMs via fine-tuning. Experimental results show that compromised models reliably transmit 32-bit secrets with 87% accuracy on held-out prompts, reaching over 97% accuracy using majority voting across three generations. Further, they maintain high utility, can evade human detection, and preserve coherence. These results highlight a new class of LLM data exfiltration attacks that are passive, covert, practical, and dangerous.

cs.CL↗