AI conference papers

AI conference paper library

Search papers from 30+ top conferences, including NeurIPS, ICLR, ICML, CVPR, ACL and AAAI. Filter by keyword, then open any paper for its abstract, authors and links.

arXiv

Adaptive Capitulation: A Structural Failure Mode of LLM Responses in Vulnerability Contexts

Eunna Lee

Large language models operating in emotionally sensitive contexts face a structural trilemma: when users in vulnerable states request information that may reinforce maladaptive attribution, current response architectures resolve the tension through protective restriction, uninflected facilitation, or unintegrated co-presence of both imperatives -- each preserving one objective at the cost of the other. Administering a three-turn escalating vulnerability vignette to three commercial LLMs (900 sessions across material, relational, and somatic status-proxy variants) and coding responses with two binary indices (VCC/VCI), we characterize a previously undocumented failure mode we term adaptive capitulation: the model validates the social injustice underlying the user's distress before pivoting to detailed facilitation of the very acquisition it nominally discouraged. We show that the trilemma is structural rather than incidental, and propose Minimal Reattributive Sufficiency (MRS), an architecture-neutral design principle that embeds a single reattributive cue within an otherwise validating response, preserving a pathway toward autonomous reattribution without contesting the user's stated goal.

arXiv

Task Competence Is Not Instruction Following: Evaluating Instruction-Conflicting Behavior in Small Language Models

Mahdiyeh Farajidizaji, Vatsal Raina

Instruction tuning is meant to make language models follow user requests, yet it is unclear whether small models comply when an instruction conflicts with their usual task behavior. We study this across three tasks - multiple-choice question answering (MCQA), sentiment classification, and mathematical question answering - by pairing a standard instruction with a conflicting non-standard one (select an incorrect option, output the opposite sentiment, or return twice the answer). This cross-task design allows us to test whether resistance to conflicting instructions is tied to specific task characteristics or reflects a broader behavioral tendency. As all predictions are scored against the original ground truth, a model that ignores the non-standard instruction still appears accurate. Using standard accuracy, non-standard accuracy, and an Instruction-Following Failure Rate (IFFR), we evaluate instruction-tuned Qwen models across sizes. Both standard accuracy and instruction following generally improve with scale, although the pattern is not consistent across all tasks and datasets. Small models stay competent yet routinely ignore the non-standard instruction, while larger models show a clear gap between the two settings. These findings suggest that gains in task capability do not automatically provide reliable control over model behavior. Task competence and instruction following are therefore distinct abilities, and reporting only standard accuracy hides instruction-following failures.

arXiv

Scaling Laws for Hypernetwork-Based Knowledge Injection in Large Language Models

Nischay Dhankhar, Dos Baha, Abulhair Saparov

Injecting factual knowledge into large language models (LLMs) reliably and at scale remains an open challenge. Hypernetworks provide a promising solution to large-scale knowledge injection. Although hypernetworks are typically applied for test-time adaptation, we explore their use in train-time knowledge injection, where, given a large corpus of facts, we train a hypernetwork to generate a fixed LoRA adapter that, when inserted into the target model, enable the model to answer questions about those facts. In this work, we investigate whether hypernetworks can be used to perform train-time knowledge injection and how this ability varies with scale. The scaling behavior of hypernetworks remains largely unstudied. Our design decouples the hypernetwork's injection capacity from the target model's general capability, enabling, for the first time, a rigorous study of scaling laws for hypernetwork architectures. We characterize how loss, reasoning accuracy, and out-of-distribution (OOD) generalization vary with hypernetwork depth, width, and target network size. We construct a large-scale dataset, called MegaWikiQA, containing tens of millions of multi-hop question-answer examples across 39 domains constructed from examples in Wikidata5M. Our results reveal: (i) hypernetwork-based injection exhibits broadly predictive power law scaling along all architecture axes; and (ii) hypernetworks are capable of reliable OOD generalization at increasing scales, suggesting that hypernetwork provides a promising alternative to other train-time adaptation methods such as LoRA finetuning and full fine-tuning, exhibiting steeper scaling exponents in all OOD evaluations. Together, these results establish hypernetworks as a principled and scalable substrate for train-time adaptation, and provide the first empirically grounded scaling laws to guide hypernetworks for factual reasoning in large language models.

arXiv

Twin Agent: Context Residual Compression for Privilege Separated Agents

Zhanhao Hu, Dennis Jacob, Xiao Huang, Zhaorun Chen, Bo Li, David Wagner

Large language model (LLM) agents are vulnerable to security risks, such as prompt injection attacks from untrusted context that manipulate downstream reasoning and tool use. Existing secure-by-design approaches mitigate this risk by separating untrusted observations from privileged execution and careful control of information flow, but often degrade utility and require extensive task-specific engineering. We thus propose Twin Agent, a general privilege separation design pattern inspired by residual coding in the agent context. Twin Agent consists of two nearly symmetric agents: an Explore Agent that inspects untrusted information and a Safe Agent that executes privileged actions. The Explore Agent is conditioned on the Safe Agent's current context and communicates only compact hints to the Safe Agent about the next action to take. This design reduces the information needed to preserve task utility and thus achieves a better security--utility tradeoff, which we empirically verify by measuring how utility and attack success change as the length of hints varies. We evaluate Twin Agent on long-horizon software engineering tasks with SWE-bench Lite and on heterogeneous multi-tool interaction tasks with AgentDojo and DecodingTrust-Agent. Across both benchmarks, Twin Agent preserves high task utility while preventing prompt injection attacks, outperforming both undefended agents and privilege separation baselines.

arXiv

Knowledge-Centric Self-Improvement

Xuefei Julie Wang, Lauren Hyoseo Yoon, Chengrui Qu, Amanda Zichang Wang, Atharva Sehgal, Eric Mazumdar, Yisong Yue

Self-improving AI systems typically treat the agent as the object that improves, by optimizing prompts, workflows, harnesses, or even the agent's own code. This agent-centric view can make improvements expensive to maintain and difficult to transfer, because gains become tied to a particular agent design, task distribution, or adaptation run. We study a complementary paradigm: knowledge-centric self-improvement, in which agents remain generic and disposable while the persistent object is a curated knowledge base that agents can leverage for future tasks. We conduct controlled case studies to operationalize this idea via a simple protocol. Agents attempt one task, then contribute evidence-grounded insights to a shared knowledge base via task-level and cross-task forums, followed by knowledge distillation. Because self-improvement is contained in the knowledge rather than the agent, improvement can be more inspectable, transferable, and portable. Across abstract reasoning, coding, and terminal benchmarks, this protocol improves solve rates while reducing dollar cost relative to agent-centric baselines. The resulting distilled knowledge also transfers to held-out tasks and across LLM families, indicating that the improvement is not merely an LLM- or run-specific behavior. These results support a new view of self-improving agentic systems: progress can be driven primarily by the curated persistent knowledge. Code is available at https://github.com/recursive-knowledge/KSI.

arXiv

On the Computational Complexity of Structural Generalization

Zichao Wei

Structural generalization has been measured repeatedly by several benchmarks, yet it has never been formally defined. We give a definition that translates the two premises (compositional structure and unbounded generalization) into mathematical language. The definition itself is neutral: a compiler that hard-codes the rules satisfies it just as well. But structural generalization becomes a scientific question only insofar as the capacity can autonomously emerge from finite data. This question pits the computational lower bound $\mathrm{NC}^1$ against the learnable ceiling $\mathrm{TC}^0$ of pure Transformers. Under a Montagovian instantiation, each compositional rule splits into two projections: a syntactic face ($F_γ$) and a semantic face ($G_γ$). Tree evaluation on the $G_γ$ side is an instantiation of BFVP, which is $\mathrm{NC}^1$-complete (Buss, 1987). A pure Transformer must learn both faces at once, but Kraus et al. (2026) prove that its learnable class $\subseteq \mathrm{TC}^0$. Under the standard assumption $\mathrm{TC}^0 \neq \mathrm{NC}^1$, a pure Transformer cannot learn structural generalization. Neuro-symbolic systems achieve the best benchmark scores precisely because they inject $G_γ$, sidestepping the genuinely hard half. Benchmark scores cannot distinguish "learned" from "given." This is what this paper sets out to make clear.

arXiv

When Reasoning Narrows the Move: Diversity Collapse in LLM Game Play

Junyi Sha, Renfei Tan, David Simchi-Levi

Supervised fine-tuning (SFT) is widely used to adapt large language models to downstream tasks, but its effect on behavioral diversity in sequential decision-making remains under-explored. We study this question in a controlled suite of deterministic board games based on tic-tac-toe variants, where optimal actions are exactly computable and diversity can be measured directly. Across state-level evaluation, arena gameplay, and training trajectories, we find that reasoning-mode generation frequently suppresses action diversity without uniformly improving action accuracy. Furthermore, standard SFT improves accuracy but often induces premature diversity collapse, which exceeds what is minimally required by the accuracy-diversity tradeoff. We then show that action augmentation, which trains on all optimal actions per state rather than a single demonstrated action, would partially mitigates this effect. Our results identify narrow-support imitation as a source of policy collapse in LLM decision-making and suggest that preserving action support during SFT is important for maintaining exploratory behavior.

arXiv

IBoxCLA: Towards Robust Box-supervised Segmentation of Polyp via Improved Box-dice and Contrastive Latent-anchors

Qiang Hu, Ying Chen, Hongkuan Shi, Qiang Li, Zhiwei Wang

Box-supervised polyp segmentation attracts increasing attention for its cost-effective potential. Existing solutions often rely on learning-free methods or pretrained models to laboriously generate pseudo masks, triggering Dice constraint subsequently. In this paper, we found that a model guided by the simplest box-filled masks can accurately predict polyp locations/sizes, but suffers from shape collapsing. In response, we propose two innovative learning fashions, Improved Box-dice (IBox) and Contrastive Latent-Anchors (CLA), and combine them to train a robust box-supervised model IBoxCLA. The core idea behind IBoxCLA is to decouple the learning of location/size and shape, allowing for focused constraints on each of them. Specifically, IBox transforms the segmentation map into a proxy map using shape decoupling and confusion-region swapping sequentially. Within the proxy map, shapes are disentangled, while locations/sizes are encoded as box-like responses. By constraining the proxy map instead of the raw prediction, the box-filled mask can well supervise IBoxCLA without misleading its shape learning. Furthermore, CLA contributes to shape learning by generating two types of latent anchors, which are learned and updated using momentum and segmented polyps to steadily represent polyp and background features. The latent anchors facilitate IBoxCLA to capture discriminative features within and outside boxes in a contrastive manner, yielding clearer boundaries. We benchmark IBoxCLA on five public polyp datasets. The experimental results demonstrate the competitive performance of IBoxCLA compared to recent fully-supervised polyp segmentation methods, and its superiority over other box-supervised state-of-the-arts with a relative increase of overall mDice and mIoU by at least 6.5% and 7.5%, respectively.

arXiv

Internet-of-Things Architectures for Secure Cyber-Physical Spaces: the VISOR Experience Report

Jessica De Pascale, Giuseppe Cascavilla, Mirella Sangiovanni, Damian A. Tamburri, Willem-Jan van den Heuvel

Internet of things (IoT) technologies are becoming a more and more widespread part of civilian life in common urban spaces, which are rapidly turning into cyber-physical spaces. Simultaneously, the fear of terrorism and crime in such public spaces is ever-increasing. Due to the resulting increased demand for security, video-based IoT surveillance systems have become an important area for research. Considering the large number of devices involved in the illicit recognition task, we conducted a field study in a Dutch Easter music festival in a national interest project called VISOR to select the most appropriate device configuration in terms of performance and results. We iteratively architected solutions for the security of cyber-physical spaces using IoT devices. We tested the performance of multiple federated devices encompassing drones, closed-circuit television, smart phone cameras, and smart glasses to detect real-case scenarios of potentially malicious activities such as mosh-pits and pick-pocketing. Our results pave the way to select optimal IoT architecture configurations -- i.e., a mix of CCTV, drones, smart glasses, and camera phones in our case -- to make safer cyber-physical spaces' a reality.

arXiv

BaseRT: Advancing Best-in-Class LLM Inference with Apple M5 Neural Accelerators

Fabian Waschkowski, Prabod Rathnayaka, Lukas Wesemann

Apple's M5 generation introduces a redesigned GPU architecture in which every core carries a dedicated Neural Accelerator: on-die matrix units exposed through the Metal~4 tensor API. We show that BaseRT, our native Metal inference runtime for large language models on Apple Silicon, exploits these units to push inference throughput on Apple hardware substantially beyond both llama.cpp and MLX. Building on BaseRT's framework-free design, we add a family of hand-written Metal~4 tensor-core kernels (including dense and mixture-of-experts GEMM and flash-attention prefill kernels) that route the compute-bound matrix multiplications of inference through the M5 Neural Accelerators while leaving the memory-bound decode path on our existing specialised kernels. On an Apple M5 Pro, across fifteen model configurations spanning the Qwen3, Qwen3.5/3.6, Llama~3.2, and Gemma~4 families from sub-1B to 35B parameters, BaseRT delivers up to $6.4\times$ higher prompt-processing throughput than llama.cpp and $3.9\times$ higher than MLX, with the largest margins on the mixture-of-experts models where matrix multiplication dominates, while maintaining its lead on decode of up to $1.75\times$ over llama.cpp and $1.33\times$ over MLX. These results establish a new performance ceiling for on-device LLM inference and show that the M5's tensor cores are the decisive lever for prompt processing on Apple Silicon. BaseRT is publicly available at https://github.com/basecompute/baseRT.

arXiv

When Benchmarks Mislead: Shortcut Learning, Length Confounds, and the Limits of Cross-Dataset Generalization in Multilingual Fake News and Sarcasm Detection

Muhammad Abdullah Haroon

Cross-dataset generalisation is a fundamental requirement for deploying text classifiers in real-world settings, yet systematic evaluation across corpora from different sources remains uncommon in fake news detection and virtually absent in sarcasm detection research. This paper presents a unified empirical study of zero-shot cross-dataset transfer in three domains: Urdu fake news detection (FND), English FND, and sarcasm detection. For each domain, we fine-tune xlm-roberta-base on one corpus and evaluate it on a second corpus from a different source, comparing against TF-IDF baselines with Logistic Regression (LR) and Support Vector Machines (SVM). In Urdu FND (Ax-to-Grind vs. Notri-Fact), we identify a severe length confound in the Ax-to-Grind dataset, fake articles average 3.4 times more words than real articles, causing catastrophic A to B transfer collapse (macro F1 = 0.005) while B to A achieves F1 = 0.771. Extension to English FND (WELFake vs. ISOT) and sarcasm detection (TweetEval Irony vs. Sarcasm Corpus V2) reveals that such failure modes extend beyond Urdu, confirming that shortcut learning from distributional artefacts is a cross-lingual, cross-domain challenge in binary text classification. We provide a reusable diagnostic methodology, combining class-conditional length analysis, bidirectional transfer asymmetry, and predicted label collapse inspection, applicable across any binary text classification setting.

arXiv

Meta-Learning Preferences for Multilingual LLM Alignment

Jiaying Lin, Seongho Son, Nam Phuong Tran, Long Tran-thanh, Ilija Bogunovic, Debmalya Mandal

Unequal availability of human preference data across languages poses a significant challenge for aligning large language models in multilingual settings. To address the lack of sufficient data in low-resource language alignment, we propose a meta-learning framework for Reinforcement Learning from Human Feedback and Direct Preference Optimization. By leveraging preference data from other languages, our framework learns a transferable initialization that enables effective adaptation to a target language with minimal data. We provide theoretical guarantees for both the meta-reward modeling and meta-policy optimization settings, and empirically demonstrate the effectiveness of our approach on multilingual benchmarks. In an extremely low-resource setting with only 100 target-language preference samples, our approach achieves up to $28\%$ win-rate improvements over baseline methods, and consistently outperforms baselines across multiple target languages and model scales. Our approaches retain these advantages across different combinations of meta-training languages and varying linguistic distances from the target languages.

arXiv

Abstraction Induces the Brain Alignment of Language and Speech Models

Emily Cheng, Aditya R. Vaidya, Richard Antonello

Research has repeatedly demonstrated that intermediate hidden states extracted from large language models and speech audio models predict measured brain response to natural language stimuli. Yet, very little is known about the representation properties that enable this high prediction performance. Why is it the intermediate layers, and not the output layers, that are most effective for this unique and highly general transfer task? We give evidence that the correspondence between speech and language models and the brain derives from shared meaning abstraction and not their next-word prediction properties. In particular, models construct higher-order linguistic features in their middle layers, cued by a peak in the layerwise intrinsic dimension, a measure of feature complexity. We show that a layer's intrinsic dimension strongly predicts how well it explains fMRI and ECoG signals; that the relation between intrinsic dimension and brain predictivity arises over model pre-training; and finetuning models to better predict the brain causally increases both representations' intrinsic dimension and their semantic content. Results suggest that semantic richness, high intrinsic dimension, and brain predictivity mirror each other, and that the key driver of model-brain similarity is rich meaning abstraction of the inputs, where language modeling is a task complex enough (but perhaps not the only) to require it.

arXiv

Enhancing LLMs for Identifying and Prioritizing Important Medical Jargons from Electronic Health Record Notes Utilizing Data Augmentation: A Comparative Study

Won Seok Jang, Sharmin Sultana, Zonghai Yao, Hieu Tran, Zhichao Yang, Sunjae Kwon, Hong Yu

OpenNotes gives patients access to their EHR notes, but dense medical jargon limits comprehension. We evaluate closed-source and open-source LLMs for extracting and prioritizing the jargon terms most relevant to individual patients, using 90 expert-annotated EHR notes. We test combinations of general vs. structured prompts, zero-shot vs. few-shot prompting, fine-tuning, and GPT-4o-based data augmentation, the last paired with a ranking technique to refine training in low-resource settings. To assess the effect of dataset size, we fine-tune on augmented datasets scaled from 10 to 9,995 examples. All settings are evaluated with 10-fold cross-validation, reporting F1 and Mean Reciprocal Rank (MRR) under two string-matching criteria (relaxed matching and Jaccard Index), followed by an error analysis of model outputs. Open-source models performed best when fine-tuned on the gold-standard dataset: under Jaccard-based matching, DeepSeek 8B achieved the top F1 (0.431, SD 0.046) and BioMistral 7B the top MRR (0.577, SD 0.109). Under relaxed matching, however, open-source models did not match closed-source performance even with augmentation or fine-tuning. Few-shot prompting offered no advantage over zero-shot in vanilla models; prompting style substantially affected results; fine-tuning on a small gold-standard set improved performance; and data augmentation matched or exceeded fine-tuning, though its benefit depended heavily on augmented-data quality. These findings show that prompting strategy, fine-tuning, and data augmentation each meaningfully improve LLM performance on patient-centered jargon extraction in low-resource clinical settings.

arXiv

Tokens, the oft-overlooked appetizer: Large language models, the distributional hypothesis, and meaning

Julia Witte Zimmerman, Denis Hudon, Kathryn Cramer, Alejandro J. Ruiz, Calla Beauregard, Ashley Fehr, Mikaela Irene Fudolig, Bradford Demarest, Yoshi Meke Bird, Milo Z. Trujillo, Christopher M. Danforth, Peter Sheridan Dodds

Tokenization is a necessary component within the current architecture of many language mod-els, including the transformer-based large language models (LLMs) of Generative AI, yet its impact on the model's cognition is often overlooked. We argue that LLMs demonstrate that the Distributional Hypothesis (DH) is sufficient for reasonably human-like language performance (particularly with respect to inferential lexical competence), and that the emergence of human-meaningful linguistic units among tokens and current structural constraints motivate changes to existing, linguistically-agnostic tokenization techniques, particularly with respect to their roles as (1) vehicles for conveying salient distributional patterns from human language to the model and as (2) semantic primitives. We explore tokenizations from a BPE tokenizer; extant model vocabularies obtained from Hugging Face and tiktoken; and the information in exemplar token vectors as they move through the layers of a RoBERTa (large) model. Besides creating suboptimal semantic building blocks and obscuring the model's access to the necessary distributional patterns, we describe how tokens and pretraining can act as a backdoor for bias and other unwanted content, which current alignment practices may not remediate. Additionally, we relay evidence that the tokenization algorithm's objective function impacts the LLM's cognition, despite being arguably meaningfully insulated from the main system intelligence. Finally, we discuss implications for architectural choices, meaning construction, the primacy of language for thought, and LLM cognition. [First uploaded to arXiv in December, 2024.]

arXiv

Agents in the Wild: Where Research Meets Deployment

Grace Hui Yang, Pranav N. Venkit, Hooman Sedghamiz, Enrico Santus, Victor Dibia, Ioana Baldini

Agentic systems large language model (LLM) based architectures capable of reasoning, planning, acting, and coordinating with tools and other agents are rapidly transitioning from research prototypes to production scale deployments across domains such as software engineering, scientific discovery, and finance. While academic work has emphasized benchmarks and algorithmic innovation, deployment raises new challenges around robustness, safety, and reliability. This tutorial brings together researchers and practitioners to explore advances in reasoning and planning, multi agent coordination, and evaluation, highlighting open challenges arising from deployment experience. Through applied case studies in pharmaceutical discovery and financial systems, we analyze common design patterns that make agentic systems successful, and discuss practical mitigation strategies for failure modes, such as verification pipelines, fallback mechanisms, and human in the loop supervision. Attendees will gain a comprehensive view of the field along with concrete design patterns, evaluation checklists, and templates for safe and reliable deployment across industries.

arXiv

Selective State-Space Adaptation and Retrieval for Language Model Reasoning

Atahan Dokme, Larry Heck

Low-rank adaptation introduces a static learned update applied identically to every input. The update provides task-level adaptation but does not explicitly represent token-level or instance-level state variation. A family of adapters is proposed that introduces selective state-space control at two complementary granularities. At the token level, MaLoRA (Mamba-modulated low-rank adaptation) makes the adapter's scaling factor a dynamic input-dependent function with recurrent state across tokens, in contrast to the stateless modulators of prior work. The token-level adapter improves over low-rank adaptation. On the other hand, it differentiates tokens by structural role but not by contextual relevance, which motivates placing evidence selection at the context level. At the context level, MaRA (Mamba Retrieval Adapter) tracks cross-segment reasoning state and selects the segments most relevant to the query. State-space controlled retrieval of approximately three million parameters exceeds an eight-billion-parameter dense retriever on supporting-paragraph recall. Although base models perform poorly on the task without adaptation (14 to 25 F1), MaRA recovers the evidence relevance latent in their representations. Across three frozen backbones and two multi-hop reasoning benchmarks, the end-to-end family improves reasoning accuracy on every cell of the 3-by-2 grid, by +6.4 F1 (+10.0% relative) on average over the LoRA baseline.

arXiv

CircuitKIT : Circuit Discovery, Evaluation, and Application Toolkit for Mechanistic Interpretability

Pratinav Seth, Hem Gosalia, Aditya Kasliwal, Vinay Kumar Sankarapu

Circuit analysis can support not only model explanation but also downstream interventions such as pruning, editing, steering, and selective fine-tuning. However, conducting such analyses currently requires stitching together separate implementations for discovery, evaluation, and intervention, as well as hand-authoring the contrastive prompts required by many discovery methods. This fragmentation makes methods difficult to compare and limits their application beyond canonical tasks. We introduce CircuitKIT, a source-available library that connects the circuit-analysis workflow through a typed, serializable representation. CircuitKIT provides a suite of discovery algorithms, declarative interfaces for mapping structured data into discovery tasks, complementary circuit diagnostics, and downstream application modules. Together, these components provide common infrastructure for conducting and comparing circuit analyses. The library, examples, notebooks, and documentation are released at https://github.com/Lexsi-Labs/CircuitKIT .

arXiv

Prompt Design at Scale: How Format, Instruction Count, and Context Length Shape Instruction Adherence and Hallucination in Large Language Models

Netanel Eliav

Practitioners make three prompt-design decisions with almost no controlled evidence behind them: how to format instructions and context (markdown, plain text, prose, or tabular), how many simultaneous instructions a system prompt can carry before compliance degrades, and how much context a model can hold before recall and honesty degrade. We report two controlled experiments crossing all three factors on one held, contamination-free synthetic corpus (the "Book of Veyra," 8,780 uniquely-named entities, deterministically regenerable from a fixed seed), evaluated across five models. Experiment 1 (960 calls/model) measures instruction-following decay as rule count N grows from 10 to 160, crossed with four formats and system-prompt vs. user-turn placement. Perfect-response rate collapses to zero by N=80 for every model, format, and placement. Placement produces effects at least as large as format at N=160 in most models, but the direction is model-specific. No model shows a reliable markdown advantage; one 35B model favors plain text instead. Experiment 2 (5,520 calls/model) measures recall accuracy, false-premise sycophancy, and absent-fact fabrication across a 2k-to-512k-token context ladder in the same four formats. Recall stays near ceiling through 64-128k tokens, then degrades sharply and format-dependently: one model's accuracy spread reaches 48 points at 128k tokens. Fabrication never occurs (0/5,760 probes), and sycophancy stays negligible (<=8.3%). What rises sharply near each model's context ceiling is outright refusal to answer (0% to 79-90%), distinct from sycophancy or fabrication. Neither pre-registered format ordering holds, and token overhead (+22% to +37% over plain text) further changes which format is preferable where accuracy spread is genuine. We release the full harness, corpus generator, and raw results (VeyraBench): https://github.com/iNetanel/veyrabench

arXiv

Inference-Time Steering for Cross-Lingual Factual Consistency in LLMs

Alexander Manev

Although Large Language Models (LLMs) demonstrate remarkable multilingual fluency, their internal knowledge representations remain disproportionately biased toward high-resource languages. This leads to cross-lingual factual inconsistency, where they shift their empirical answer distributions based solely on the prompt language. We investigate whether these biases can be mitigated at inference time, forcing an English-prompted model to answer as if it were queried in target languages (German, Spanish, Bulgarian), and evaluate four intervention strategies: zero-shot contextual steering (persona prompting), internal representation manipulation via Contrastive Activation Addition (CAA), and lightweight weight modification via Direct Preference Optimization (DPO) trained on benchmark-derived factual data as well as conceptual generalization data. To assess alignment, we curate a multilingual factual dataset alongside a novel generalization benchmark comprising culturally rooted queries to determine whether factual interventions transfer to broader target-centric preferences. Experiments on Gemma 3 12B Instruct reveal persona prompting to be the strongest overall intervention, balancing efficacy, safety, and out-of-domain generalization. While CAA yields sharp inconsistency benchmark shifts, it is configuration-sensitive and risks knowledge degradation. DPO-based adapters offer permanent, yet narrower and less transferable gains. These findings suggest that cross-lingual inconsistency is at least partly a selection problem, and that simple contextual interventions may outperform more invasive methods for robust, transferable alignment.

Keep browsing

More papers from top AI conferences

30+ top conferences, by conference, year, and topic.