Contextual Tamil Spelling and Grammar Correction Using Progressively Fine-Tuned Sequence-to-Sequence Transformers

Tamil spell and grammar correction is challenging because Tamil is an agglutinative low-resource language with rich verbal morphology, complex sandhi (phonetic transformation) rules at word boundaries, and a script of 247 distinct letters. Prior work targets word-level surface errors with rule-based methods, statistical n-gram models, Minimum Edit Distance, or hybrid pipelines with a transformer re-ranker; such methods cannot reliably handle contextual errors - subject-verb agreement, tense consistency, or cross-word sandhi - which require sentence-level understanding. We propose an end-to-end sequence-to-sequence formulation and fine-tune mT5-small and mBART-50 on a synthetic corpus of up to 657,720 noisy-clean Tamil sentence pairs spanning ten error categories. Both backbones follow the same four-stage progressive schedule, each stage targeting one weakness: surface noise (v2), contextual grammar (v3), single-site sandhi (v4), and multi-site cross-word sandhi (v5). On a 1,000-sentence balanced diagnostic set verified disjoint from all training data, our best model, mBART-50 v5, reaches 69.3% top-1 exact-match accuracy, with 87.5% on sandhi and 43.5% on subject-verb agreement. The schedule is what produces these gains: subject-verb accuracy rises from 1.0% to 52.5% once contextual pairs are introduced, and sandhi from 0% to 87.5% once multi-site sandhi pairs are. We additionally quantify a precision-recall trade-off this literature has not reported: sandhi recall is paid for monotonically in identity accuracy. Finally, Tamil-LLaMA-7B-Instruct reaches 19.0% zero-shot and 24.7% with three demonstrations against a 20.0% copy baseline, showing that a Tamil-adapted instruction model does not transfer to specialised sentence-level correction without task-specific supervision.

Improving precipitation forecasts in an AI weather model using observational data

Artificial intelligence weather prediction (AIWP) systems now surpass state-of-the-art physical models for medium-range weather forecasting. Current global AIWP models are trained almost exclusively using one reanalysis dataset, ERA5, but it has known biases, particularly for precipitation. Here we fine-tune a graph-transformer architecture with IMERG precipitation data at 0.25° resolution. The resulting model improves medium-range continuous ranked probability scores by up to 19%, while also demonstrating superior skill for tropical storms and drizzle events. Our model exceeds the Brier skill score of state-of-the-art operational models on extreme rainfall prediction by 57% globally; however, a physics-based operational model remains more reliable for the heaviest precipitation events. Our results demonstrate that incorporating observations-based precipitation data directly into training can substantially improve precipitation forecasts.

Routing Is Not Enough: Diagnosing Intra-Adapter Subspace Contention in MoE+LoRA Fine-Tuning

Multi-domain fine-tuning often combines MoE routing with LoRA, assuming that token-level routing separates domain-specific updates. We test this assumption in MoE+LoRA using Python code paired with biomedical text and mathematical reasoning. Although these domains show near-disjoint expert routing, adding biomedical data substantially increases code perplexity, indicating that routing separation alone may not prevent negative transfer. To localize the failure, we introduce Jaccard routing overlap and adapter-gradient cosine similarity, which measure expert sharing and update compatibility, respectively. These diagnostics indicate that interference arises mostly from nearly orthogonal domain gradients competing within the same low-rank adapter subspace. We address this issue with SpawnLoRA, which dynamically adds gated sub-adapters inside MoE experts when adapter-level contention is detected, while keeping the router fixed. We evaluate SpawnLoRA on Phi-tiny-MoE-instruct and OLMoE-1B-7B across multiple mixture settings and find that it effectively reduces negative transfer compared with standard and rank-adaptive LoRA. These results demonstrate that structural separation inside experts provides benefits beyond routing or rank expansion alone.

LoRA-TSD: Tangent-Space Spectral Descent for LoRA via Muon-Style Updates

Low-rank adaptation (LoRA) is the standard way to fine-tune large models, yet when its two factors are trained independently, the update ignores the geometry of the low-rank weight change it induces. We introduce LoRA-TSD, an optimizer that treats every LoRA step as a tangent vector of the fixed-rank matrix manifold and takes the spectral-norm steepest-descent step of Muon inside that tangent space, mapping the result back to the factors through a retraction native to the LoRA parametrization. The step avoids expensive operations on full weight matrices, and its retraction is up to $2.8\times$ cheaper than the truncated-SVD retraction used by prior manifold methods. We prove that the Frobenius-norm version of our surrogate recovers LoRA-Pro, and we identify the tangent-projected gradient, the Riemannian gradient of the manifold, as the stationarity measure natural to LoRA training and computable from the factor gradients alone. Under this measure we give the first global convergence guarantees for both LoRA-Pro and LoRA-TSD, with rates that drive the factor-gradient norms to zero. Across six commonsense and natural-language-inference benchmarks with Llama-3.2-1B, Llama-3.1-8B and Qwen3-32B, LoRA-TSD outperforms every competing LoRA optimizer and stays robust to the adapter rank. Code is available at https://github.com/brain-lab-research/LoRA-TSD.

Adapting a Foundation Model for Lunar Surface Height Estimation

Digital elevation models (DEMs) can provide accurate height information, making it invaluable for analyzing the lunar surface. As the European Space Agency (ESA) prepares for future lunar missions that aim to land on the Moon, a precise method for height estimation will be essential for hazardous terrain that could endanger the landing approach. Traditional approaches to generate DEMs from imagery, such as shape from shading (SfS) and stereophotogrammetry (SPG) have been proven highly valuable for this task. However, due to advancements in machine learning, especially computer vision, the focus has shifted towards monocular depth estimation via deep learning. The lunar surface is covered by rocks and craters, and classic hazard detection methods rely solely on 2D image data. Our goal is to address this issue by developing a relative lunar surface height estimator that can provide additional information for hazard localization. In this letter, we present a methodology that builds on the well-known zero-shot relative depth estimation model Depth Anything V2 (DAV2). Other works have been using it as a state-of-the-art comparison for their proposed lunar DEM estimation method, but without adaptations to the target domain. Thus, it may underperform. Therefore, we propose a fine-tuning strategy with publicly available SPG-derived DEM data of the lunar surface. Our results demonstrate a significant improvement in performance compared to the zero-shot model, effectively transforming DAV2 into a reliable relative depth estimator of the lunar surface.

Task-Level Natural Language Priors as Learning Signals for Low-Resource LLM Training

Large language models (LLMs) often struggle when low-resource training data are ambiguous or incomplete. Task-level natural-language priors can provide useful guidance in such settings, but existing approaches usually treat these priors as input context rather than as learning signals during training. We propose Prior-Guided Tuning (PGT), a training perspective that incorporates natural-language priors as auxiliary learning signals for low-resource LLM training. Under this perspective, we introduce Contrastive Prior Steering (CPS), which keeps the original supervised objective intact while adding positive and negative prior-conditioned auxiliary losses to encourage task-consistent learning and discourage plausible but misleading alternatives. Experiments on AmbiMath, Jigsaw, and MNLI/HANS show that CPS consistently improves over plain and prompt fine-tuning. On AmbiMath, CPS achieves 97.6% average exact-match accuracy. On Jigsaw, CPS improves average Macro F1 by 9.5 percentage points over standard fine-tuning, and with 1/10 of the experimental training data slightly exceeds full-data plain fine-tuning. On HANS, CPS improves non-entailment accuracy by 8.3 and 5.2 percentage points for LLaMA 3.1 8B and Qwen 2.5 7B, respectively, while maintaining comparable in-domain MNLI accuracy. These results support our central claim: task-level natural-language priors can provide useful guidance as auxiliary learning signals for low-resource LLM training. Our code and data will be publicly available.

SCULPT: Training Edge Vision Models for Post-Training Quantization Readiness

Edge vision models are difficult to deploy on resource-constrained hardware, making low-bit post-training quantization (PTQ) attractive. In practice, standard FP32 training often produces heavy-tailed activation distributions whose outliers destabilize activation quantization: preserving the full range wastes quantization bins on rare extremes, while aggressive clipping causes information loss. Existing solutions typically rely on quantization-aware training (QAT), which adds training complexity and bit-width coupling, or advanced PTQ procedures that repair the model after training. We present SCULPT (Statistical Clipping and Uniform Loss for Post-Training), a training-time method that improves PTQ readiness during ordinary FP32 fine-tuning. SCULPT combines a topology-aware activation regularizer that suppresses quantization-hostile skewness and kurtosis with a stable percentile-based clipping mechanism that learns deployment-ready activation bounds. Unlike QAT, SCULPT does not simulate quantization during optimization; unlike post hoc outlier-repair PTQ methods, it does not require runtime activation transformations. The learned clipping bounds can be exported directly into a standard PTQ workflow for low-bit deployment, including INT8 and lower-bit settings such as W4A8.

HEAT: Faster Fully Homomorphic Inference via Approximations-Weights Co-Adaptation

Fully homomorphic encryption (FHE) allows a server to run a language model directly on encrypted user prompts, but current approaches remain prohibitively slow. Ciphertexts natively support only addition, multiplication, and rotation, and multiplications may be composed only to a bounded depth before a costly bootstrapping operation is needed to continue. Every nonlinearity must therefore be approximated by an iterative method, and each iteration uses multiplications. A higher iteration count buys precision but exhausts the available depth faster and triggers more bootstraps, which dominate latency. Existing approaches fix the iteration counts uniformly across the model rather than tailoring them to each site's error tolerance. We introduce Homomorphic Encryption-Aware Training (HEAT), a fine-tuning method that makes the per-nonlinearity iteration counts learnable, enabling them and the model weights to co-adapt during training. HEAT optimizes iterations with respect to the task objective, allowing the model to adapt to approximation errors encountered during inference without architectural changes or retraining from scratch. On encrypted GPT-2 decoding, HEAT reduces iterations by $3.1\times$, bootstraps by $1.6\times$, and end-to-end latency by $1.4\times$, while improving decode agreement over the calibrated baseline.

Hearing the Whispers: Black-Box Membership Inference Attacks on Finetuned TTS Models

Text-to-Speech (TTS) foundation models are increasingly fine-tuned on private datasets to synthesize highly personalized voices, introducing severe privacy risks by exposing both biometric identities and sensitive speech content. Existing black-box membership inference attacks (MIAs) follow a two-stage pipeline of query generation and representation engineering, both of which face unique challenges when adapted to TTS. For query generation, dual conditioning on synthesis text and reference speech creates a large and underexplored query design space with no established criterion for identifying an effective query. For representation engineering, the multi-level speech characteristics and temporal variability of speech make low-level representations and direct comparisons inadequate for capturing membership signals. To address these challenges, we present the first black-box MIA framework explicitly tailored to TTS models at both the speaker and record levels. For query generation, we characterize the feasible query space and establish two criteria, scorable extent and memorization elicitation, for evaluating five representative queries, identifying recitation as the strongest. For representation engineering, we obtain multi-level speech representations from embedding models and temporally align the generated and target audio for fine-grained comparison. Evaluations across three state-of-the-art TTS models (CosyVoice2, F5-TTS, and XTTS-v2) fine-tuned on two benchmark datasets (VCTK and British Dialect) reveal severe privacy leakage: speaker-level AUC remains above 0.80 and approaches 1.0 in the strongest settings, while record-level AUC ranges from 0.80 to 0.90 and remains effective even in challenging scenarios where both members and non-members are of the same speakers. We further identify speech characteristics associated with disproportionate vulnerability to memorization.

Closing Cost-Quality Gap in Document VLMs: Difficulty-Aware Data Curation and Quality-Adjusted Deployment Economics

Extracting structured fields from hundreds of millions of documents annually remains costly in regulated industries: bespoke OCR cascades cover only a fraction of workflows, privacy rules preclude external models, and existing open-source VLMs that clear quality thresholds cost more to serve than human annotation. We present a deployed document-understanding system built on a Mixture-of-Experts VLM (35B total, 3B active), fine-tuned on in-house production data mixed with open-domain documents curated by a Difficulty-Aware pipeline for layout diversity, fact-extractability, and cross-model consistency. Fitting on a single H100 and serving heterogeneous workflows via prompting, the model leads all deployable (non-reasoning) baselines up to an order of magnitude larger. A quality-adjusted cost analysis, with confirmation and correction costs calibrated from production telemetry, shows it reduces expected costs by over 80% against the human baseline and by more than 50% against the best competing open-source model, while larger baselines remain economically unviable.

When Safety Routing Breaks: Understanding Alignment Fragility under Benign Fine-Tuning

Benign fine-tuning severely weakens the safety alignment of large language models (LLMs), so we study why refusal behavior is so fragile. While prior work often attributes this failure to gradient conflict, we propose a fundamentally different Fisher-geometric explanation: safety Fisher is low-rank, and alignment makes the safety geometry flatter while preserving an output-routing pathway. After 100 benign fine-tuning examples, this pathway is selectively re-sharpened in output-side MLP modules, explaining the asymmetric fragility: safety can collapse to high attack success rates, while general utility degrades mildly. The routing view also explains why few safety examples can restore refusal behavior, indicating that internal safety-relevant representations are preserved. Finally, we show that LoRA and ASAM mitigate early collapse by suppressing output-side sharpness, but their protection weakens at larger fine-tuning scales. Overall, safety failure is best understood as a disruption of a low-rank output-routing mechanism

Prompt-Robust Language Models: Which Training Strategies Work?

Despite their strong performance, large language models remain highly sensitive to prompt formulation. Prior work addresses this through refined data construction or through dedicated robustness objectives. We reproduce and compare these strategies under controlled conditions, and measure how effective they are in addressing models' prompt sensitivity. We find the current robustness fine-tuning methods improve over standard fine-tuning and in-context learning, but the best-to-worst prompt gap remains as high as 40-57% of performance. Moreover, the recent robustness-enhancing methods we test - CoIN for contrastive alignment and PPCL for consistency regularization - often fail to outperform the simplest data construction strategy: training on one template per batch. Our diagnostics explain these results. The auxiliary objectives move the quantity they penalize, but do not generalize beyond it. Additionally, data construction strategies differ due to the conflicting signs of per-template gradients on 57-64% of parameters. Thus, batches that mix formulations force the optimizer to reconcile competing updates instead of finding a shared, prompt-agnostic one.

Membership Inference in Fine-tuned Diffusion Language Models via Token-level Memorization Asymmetry

Diffusion language models (DLMs) have recently emerged as an alternative modeling paradigm to autoregressive LMs, offering advantages such as parallel generation and bidirectional context modeling. Despite growing interest in their generative capabilities, the privacy risks of DLMs remain underexplored. We identify a phenomenon termed token-level memorization asymmetry through theoretical analysis of diffusion training dynamics. Building on this finding, we propose Q-Skew, a quantile-weighted skewness-based indicator for membership inference on finetuned DLMs. Experiments across multiple fine-tuning datasets and models show that our method outperforms existing baselines. Moreover, we show that Q-Skew can also facilitate other privacy violations, such as PII extraction. Our findings reveal a previously underexplored privacy attack surface and highlight the need for systematic privacy evaluation of DLMs.

Text Capability Loss in Vision-Language Adaptation: An Attention-Sink Diagnosis

Fine-tuning a pretrained LLM into a vision-language model (VLM) can erode the backbone's text capability, with the damage concentrated on tasks that require following exact output rules, such as instruction following, chain-of-thought reasoning graded on a strictly parsed final answer, and similar evaluations with strict graders. We trace this gap to attention-sink corruption: VL fine-tuning perturbs the early sink position that anchors a large fraction of attention probability, and how well the base LLM preserves its sink tracks how much of the affected capability survives adaptation. Building on this view, we introduce Sink Strength, a single scalar computed on the base LLM in a few seconds on a single GPU that predicts post-VL degradation without any VL training. It consistently tracks relative degradation across the six VLM-LLM pairs and multiple format-sensitive tasks. Complementing this diagnostic, we find that post-pretraining QK-RMSNorm injection fails to reproduce the protection of native QK-RMSNorm, while several off-the-shelf weight-merging settings fail to recover the lost capability after VL training. These negative results underscore the value of screening backbones with Sink Strength before VL training and narrow the intervention space toward head-selective training-time protection.

Online Self-Weighted Fine-Tuning

Standard supervised fine-tuning (SFT) assigns the same explicit loss weight to every expert demonstration, regardless of the model's changing competence over training queries. Reinforcement learning (RL) based methods adapt update strength using model-generated rollouts, but often require substantially more sampling and can be unstable on hard tasks. We propose \textbf{Online Self-Weighted Fine-Tuning (OSW-FT)}, a simple method that augments SFT with online, trajectory-level weighting. For each query, OSW-FT estimates the model's current success rate using a small number of inference-only rollouts and rescales the standard SFT loss accordingly. The optimization direction remains anchored to the expert trajectory, while the update magnitude adapts online. For binary-verifiable reasoning, we connect this weighting to SFT and RL at the gradient level, inspired by variance-reduction principles. The resulting estimator is unbiased for the exact OSW-FT surrogate update for any finite rollout count, and we analyze convergence with respect to the corresponding surrogate objective. Evaluated across Qwen3 series ranging from 0.6B to 4B on multiple challenging benchmarks (e.g., AIME), OSW-FT consistently improves over SFT on small-to-medium scale models. OSW-FT offers a favorable compute-performance trade-off as a practical approach for fine-tuning small-to-medium LLMs on binary-verifiable reasoning tasks with only \textbf{2 online rollouts}.

CRAFT: Fine-Tuning Pre-hoc Explainability in AI-native 6G RAN

The next generation of mobile networks is envisioned as fully AI-native, with AI-RAN architectures embedding small language models (SLMs) to perform reasoning over real-time telemetry. The state-of-the-art training paradigms for telecom LLMs, exemplified by RANSTRUCT-style supervised fine-tuning (SFT) on curated instruction data, are limited to post hoc rationalization. Here, the explanations, when produced at all, are generated after or independently of the decision, leaving the decision process unauditable. Pre-hoc reasoning, where a causal reasoning trace is produced before the output label, is preferable, and the broader LLM reasoning literature has made real progress toward it via RL methods such as Group Relative Policy Optimization (GRPO). Here we observe that transplanting this recipe into the telecom setting runs into a cold-start barrier: SLMs either learn to output the desired format or learn to predict the label, but rarely both. We identify this barrier and propose CRAFT, which stands for Cold-start Reasoning Alignment via Fine-Tuning, a data-centric method to autonomously generate a verified dataset of (input, trace, label) triplets. CRAFT fine-tunes SLMs on this verified data using low-rank adaptation (LoRA), requiring substantially less compute and wall-clock time than GRPO-based methods. On the TRACTOR and IC xApp telecom datasets, CRAFT achieves up to 86.5% and 94.6% for accuracy and F1 with no parse failures, while direct GRPO and SFT+GRPO fail to exceed 28% and 53.5% F1 with multiple parse failures. We further show that CRAFT-initialized policies serve as a robust foundation for subsequent GRPO fine-tuning, as under diverse reward functions the performance remains consistent with no parse failures. Finally, we demonstrate that CRAFT consumes 59% less energy than GRPO-based baselines, making it a sustainable path to deployable, auditable AI in 6G RAN.

MemeBridge: A Dataset for Benchmarking and Mitigating the Bidirectional Cultural Gap in Meme Interpretation

Communicating across cultures is inherently challenging, especially through culturally dense and ambiguous formats like memes. While people expect large language models (LLMs) to hold promise for bridging such gaps, existing benchmark datasets often fail to capture the cultural context necessary for accurate interpretation. To address this, we introduce MemeBridge, a curated dataset centered on U.S.-originated memes, designed to capture two complementary perspectives: (1) how Chinese participants interpret these memes, and (2) how U.S. participants anticipate how people from other cultures might misunderstand them. Here, context refers to implicit cultural knowledge, including background beliefs, norms, and shared assumptions that shape meme comprehension. The dataset was constructed via a multi-stage crowdsourcing pipeline with rigorous validation, including human agreement checks and GPT-based classification verification. Each meme is annotated with sentiment, emotion, cultural significance, and knowledge type, providing rich supervision for downstream tasks. Notably, we observe that the anticipated misunderstandings from U.S. participants are often inaccurate, highlighting the asymmetries in cultural understanding and the challenges of adopting perspectives beyond one's own. This bidirectional framing, which focuses on both expression and perception, enables more nuanced benchmarking of cross-cultural comprehension. Our probing of multiple LLMs reveals that while models developed in different cultural contexts exhibit partial cross-cultural understanding, they often struggle with sophisticated interpretations. By contrast, fine-tuning with MemeBridge improves model performance, underscoring the value of culturally grounded resources for training and evaluating LLMs in globally diverse settings.

WHALE: A Simple Recipe for Joint Harness-Weight Optimization

Agent performance depends jointly on the model parameters and the executable harness code that manages context and control flow. Optimizing either component in isolation can leave the system bottlenecked by its frozen counterpart: weight updates can change which harness is effective, while harness updates can change which model capabilities are exposed. Existing joint-adaptation methods optimize weights and textual prompts but leave the broader harness fixed. We propose Weight-Harness Alternating LEarning (WHALE), a simple recipe that alternates two phases: updating the model under the current harness, then searching for a better harness under the updated model. We instantiate these two phases with online rejection-sampling fine-tuning and Meta-Harness, respectively. When to switch is a key design choice: to separate real improvements from noise without over-optimizing against a changing counterpart, WHALE uses either fixed phase durations or an adaptive patience rule over training signals. Using Qwen3.5-2B/4B agents across three domains (search question answering, mathematical reasoning, and chess puzzles), WHALE outperforms weight-only, harness-only, and Fast-Slow Training by 4.15-24.38 percentage points in best mean@8 accuracy. Either component can be the bottleneck: harness search matches peak weight-only accuracy with far fewer rollouts in SearchQA, but improves math accuracy only after a weight update. Small interleaved updates also outperform stagewise weight-then-harness optimization in accuracy and rollout cost. The code is available at https://github.com/krafton-ai/WHALE.

Fine-Tuning Low-Bit Models with Gradient in Quantized Code Space

Fine-tuning Low-bit models aims to adapt a quantized model while keeping the final deployed checkpoint in the same low-bit form. This setting is practically important as it reduces memory and inference cost for storage and deployment. Under this constraint, adaptation becomes an optimization problem over quantization codes and scales. Existing continuous low-bit training is efficient, but it can be distorted by straight through estimation error or by post-quantize gap; discrete search is deployment-faithful, but it is often too inefficient under a finite training budget. We propose code surrogate gradient as the first order signal in deployable code space to acceleate optimization, and performing guided search to preserve deployment faithfulness. Experiments across arithmetic reasoning, instruction following, and structured language understanding show that GradCodes consistently improves fine-tuning low-bit models across different quantization datatypes. Code is provided at https://github.com/ovo67/GradCodes.

ECGQuest: Benchmarking and Fine-Tuning Language Models for Electrocardiography

Electrocardiogram (ECG) interpretation requires knowledge of cardiology, electrophysiology, clinical diagnosis, ECG waveforms, signal acquisition, and instrumentation. Existing language-model benchmarks, however, primarily assess broad medical knowledge or interpretation of individual ECG signals and images rather than the broader contextual knowledge required for ECG interpretation. We developed ECGQuest, a literature-grounded resource for evaluating and fine-tuning ECG-specific language models. A GPT-4o-based pipeline generated questions from 23 ECG references and Computing in Cardiology proceedings from 2003-2025. The final dataset contains 10,904 unique True/False questions paired with their negated forms (21,808 Q&A pairs). We evaluated three commercial and 20 open-source language models on a held-out test set in a zero-shot setting. Five open-source models with 7-14B parameters were fine-tuned using Low-Rank Adaptation, with BERT and BiomedBERT included as supervised encoder baselines. Generalization was assessed on ECG-related subsets of MedMCQA and MedQA converted to binary True/False questions using official answer keys. Zero-shot accuracy on ECGQuest ranged from 49.5% to 74.4%, with GPT-5 performing best. General-purpose models outperformed medically specialized models, several models showed strong True/False bias, and encoder baselines performed near chance. Fine-tuning improved all open-source models by 6.5-14.1%. Fine-tuned DeepSeek-R1-Distill-Qwen-14B reached 76.3% accuracy, while a five-model voting ensemble reached 78.5%. On MedMCQA and MedQA, fine-tuning mainly benefited weaker or class-biased models and did not consistently improve strong base models. ECGQuest provides a reproducible benchmark for contextual ECG knowledge and shows that parameter-efficient fine-tuning can make smaller language models competitive with substantially larger commercial models.

Evaluating and Mitigating Anti-LGBTQ Biases in German and Multilingual Language Models

While gender and racial biases in language models have been widely studied, anti-LGBTQ biases remain underexplored, particularly beyond English. Existing benchmarks often do not capture cultural and linguistic variation and rely on gender representations. This paper introduces a multilingual German-English benchmark dataset for the evaluation of anti-LGBTQ biases in language models. It combines community-sourced stereotypes from German-speaking queer individuals with a German translation of WinoQueer. The data is used to evaluate eight language models across sizes and architectures and explore mitigation through fine-tuning on community and progressive media content. Results show that language models reproduce anti-queer stereotypes, with variation across identities and models. Differences between the translated and community-based data highlight the importance of cultural adaptation for multilingual bias evaluation. Fine-tuning reduces bias on average, but not consistently across models and identities. Warning: This text contains examples of anti-queer hateful language and stereotypes.

GAFT: Geo-Anchored Fine-Tuning for Hazard Identification from Rare Failures

Off-road navigation can fail when physical structures induce irrecoverable states such as high-centering or entrapment, requiring human interventions. Identifying these structures is crucial, yet challenging. Such failure events are rare and costly to collect, resulting in limited training data. Moreover, the collected data associate frames with outcomes, but do not indicate the visual cues responsible for the failure. Learning directly from these data can therefore exploit scenario-specific visual cues, leading to poor generalization. We propose \textbf{Geo-Anchored Fine-Tuning (GAFT)}, a parameter-efficient method that adapts a vision foundation model with a geometry-derived prior. It guides LoRA adaptation by aligning a spatial attention-rollout map with the geometry prior, while preserving pretrained representations. On an intervention-verified forest hazard benchmark, across ten independently trained adaptations, GAFT consistently outperforms frozen DINOv2 and supervised PEFT baselines, improving the repeated leave-one-scenario-out mean $F_2$ from 0.0607 to 0.3757 with statistical significance under paired analysis. Within these independently trained models, the best-performing GAFT model achieves a repeated-LOSO $F_2$ of 0.570. Code and benchmark: https://github.com/Xu-Yanran/geo_anchored_fine_tuning

Hidden Threat in Synthetic Data: Covert Targeted Bias Injection through Benign Text

Synthetic data is increasingly used to train large language models (LLMs), yet its security implications remain poorly understood. Prior work on subliminal learning suggests that models can inherit behavioral traits from seemingly unrelated training data. In this work, we investigate whether such mechanisms can be exploited to inject targeted social biases into aligned models through semantically benign synthetic data. We construct a pipeline in which a misaligned teacher model generates filtered synthetic datasets across domains such as creative writing and code generation, which are then used to fine-tune aligned student models. Our experiments show that benign-looking synthetic data can act as a covert channel for transmitting targeted biases while largely preserving the student model's general task capabilities. These results reveal a previously underexplored security risk in synthetic data-driven LLM training pipelines and highlight the need for improved safeguards. As one possible step toward this goal, we suggest that log-linearity-based scoring may provide a useful signal for screening seemingly benign synthetic data.

OCR-Based Field Extraction for Archaeological Pottery Metadata: The CENTURIA Dataset

Pottery is a primary source for reconstructing the chronological and economic dimensions of past societies. Archaeologists often document ceramic finds through technical drawings and handwritten metadata. This metadata is critical for dating, provenance attribution, and cross-site comparison, but remains inaccessible to computational analysis, requiring manual transcription of every record. We investigate whether state-of-the-art document analysis models can address this task, and introduce CENTURIA, a dataset of 507 pottery records from the Roman site of Carnuntum, providing transcriptions, bounding boxes, and structured field-level labels across seven metadata categories. Benchmarking five OCR models reveals a substantial domain gap: zero-shot transcription error reaches 15-32% SpACER-M, far exceeding rates on printed archival documents, with domain-specific fields recovered in fewer than 3% of cases. LoRA fine-tuning on just 57 samples, reflecting a realistic archival annotation budget, closes this gap, reducing transcription error to below 1.5% and recovering overall field-level accuracy above 87%. Our results show that a small expert-validated fine-tuning set suffices to convert handwritten pottery documentation into structured, searchable metadata ready for archaeological databases.

Do Small Models Use the Law You Give Them? Measuring Context Use on a Bilingual Bangladesh Legal Benchmark

Fine-tuning can improve legal question-answering accuracy without improving how models use law supplied in context. We study this distinction in bilingual Bangladeshi legal QA, where observed errors can arise from answer scoring, retrieval, or failure to use relevant law. We construct a hierarchy-preserving statutory corpus, 2,165 reviewed bilingual fine-tuning examples, and a 150-item supplied-law control. We evaluate six instruction-tuned models: Llama-3.2-1B, Llama-3.2-3B, Qwen3.5-0.8B, Qwen3.5-2B, Qwen3.5-4B, and Gemma-4-E2B, with three LoRA seeds per model. To separate effects, we combine constrained option-letter scoring, cyclic option rotation, and controlled removal of the governing provision. On 398 Bar Council outputs, an exact-line parser attributes an accuracy gain of 50.0\% to the Qwen3.5-2B seed-42 adapter, whereas option scoring yields only $3.0\%$. For Gemma-4-E2B, the two scoring methods favor different systems. When the governing provision is guaranteed to be present, five of six reference models improve by $14.7\%-19.3\%$ under the four-order criterion. Removing that provision reduces accuracy by $8.0\%-15.3\%$ for models and by $13.8\%-14.9\%$ points for their adapters. However, difference-in differences estimates show no increase in reliance on the governing provision after fine-tuning. Results show that legal adaptation claims require separating scorer, retriever, and model effects. Our Code and data are available at https://anonymous.4open.science/r/bangladesh-legal-qa-11E3

Stratified Consistency Distillation for Natural Language Formalization

Neurosymbolic reasoning has shown promising success in addressing complex reasoning tasks by combining large language models (LLMs) and symbolic solvers. While this approach shows promise, a fundamental challenge remains: improving the accuracy of translations from natural language to logical formulas. Current methods predominantly rely on prompt engineering, which is difficult to scale across different domains and input formats. Drawing inspiration from the success of fine-tuning in other model adaptation and alignment applications, we propose a fine-tuning-based Stratified Consistency Distillation approach: (1) We generate K logical translations per input using a frontier LLM and cluster them by semantic equivalence (2) Based on the entropy level, we apply majority voting (low entropy), LLM-as-a-Judge (medium entropy), or unification/abstention (high entropy), and (3) fine-tune a smaller model using the selected pseudo-labels. Our experiments show significant and consistent improvements in both Pass@K and our novel Equivalent Logical Similarity metrics, demonstrating the potential of advancing logical translation through consistency distillation.

Hallucination Mitigation for Large Vision-Language Models via Implicit Feature Stabilization

Large Vision-Language Models (LVLMs) are prone to hallucinations: they fluently describe objects, attributes, and scenes that are not in the image. We connect part of this failure to a measurable property of their representations, feature instability, where mild semantics-preserving perturbations of the input cause large changes in the learned embeddings; hallucination rates rise together with this variability. Existing stability-motivated remedies are explicit, in the sense that they intervene at inference time through latent steering or constrained decoding, and pay for it on every query. We propose implicit stabilization instead: perturbation-invariance is built into the model weights during fine-tuning, and nothing extra runs at deployment. Our framework, INFUSE, first stabilizes visual and textual representations around perturbation-averaged and ground-truth anchors, then aligns the stabilized representations across modalities with bidirectional contrastive objectives. We prove that the anchor's root-mean-square deviation from the perturbation-mean representation shrinks at rate $1/\sqrt{K}$ in the number of views, and that under a Lipschitz decoder, this bounds how much any perturbation can change the model's hallucination behavior. On LLaVA-1.5, LLaVA-1.6, and Qwen3-VL-8B-Instruct, INFUSE reduces AMBER CHAIR by 46-63% relative to each base model, improves ObjHal, MMHal, HallusionBench, and POPE, and preserves VQA-v2 and TextVQA, all with no inference-time overhead.

Attention Sensitivity Is Not Enough: Dissociating Attention-Level and Behavioural In-Context Learning under Fine-Tuning

In-context learning (ICL) lets large language models adapt to new tasks from demonstrations, and fine-tuning can erode this behaviour. Many preservation diagnostics inspect attention: if attention changes when demonstrations change, the model is treated as context-sensitive. This paper asks how far that proxy can be trusted once it is optimised. We formalise \emph{In-Context Sensitivity} (ICS), the average row distance between last-token attention on matched and mismatched demonstration prefixes, and pair it with \emph{ICL-GAP}, the behavioural accuracy gap between the same prefixes. In a controlled four-arm ablation on Llama-2-7B, an ICS-maximising regulariser ($\armKL$) drives ICS to $1.413$, within $0.5\%$ of its geometric ceiling. The behavioural readout tells a different story: ICL-GAP stays near zero and MMLU accuracy moves from $0.371$ to $0.279$, a Goodhart dissociation of the bounded attention proxy. Endpoint statistics locate the mechanism: attention grows sharp and near-disjoint across prefixes yet routes to formatting and demonstration-body tokens rather than labels. A random-label protocol confirms that the behavioural probe family retains dynamic range at the same checkpoints. In a constructive sweep, behaviour gating partially mitigates the effect, while objectives anchored to pretrained computation hold the high-MMLU, moderate-ICS region that divergence maximisers leave. The main lesson is diagnostic: attention-level ICL proxies earn their place as training targets only after validation against behavioural gaps.

Cloud and On-Premises Deployment of Uzbek Legal RAG via Targeted Retriever Fine-Tuning

Deploying large language models for legal question answering raises challenges that general-purpose leaderboards do not capture, particularly for low-resource languages and under hard operational constraints. We report on building and operating a retrieval-augmented (RAG) legal assistant for Uzbek that must run in two regimes: a managed cloud service that maximizes answer quality within a per-token cost ceiling, and an on-premises deployment for clients whose legal data may not leave their infrastructure, restricting us to open-weight models on limited local hardware under latency constraints. Because no evaluation existed for this setting, we build two domain benchmarks: a retrieval benchmark of 178 expert-annotated legal queries with gold provision spans, and an end-to-end benchmark of 504 expert-curated question--answer pairs scored by an LLM judge whose ratings we validate against human judgments and against an independent-family judge. Applying these benchmarks under each regime, we find the open-versus-proprietary gap is small and cheaply closed by fine-tuning. Therefore, we train UTE-1, which is a state-of-the-art text embedder among open models for Uzbek. We also demonstrate that closing the performance gap via fine-tuning is both impractical due to the intensive hardware demands of long-context legal Q\&A and unnecessary, given that legal acts change frequently. We support this by reporting a negative result from a QLoRA experiment. We distill practical guidance for similar deployments, drawn from a system serving real users in production. We release our benchmarks, evaluation code and the fine-tuned embedder (UTE-1) \href{https://metric-ai-lab.github.io/Uzbek-Legal-RAG/}{at this https URL} to support future work on low-resource legal NLP.

Emergent Misalignment Is Not Magical

Fine-tuning large language models (LLMs) on narrowly harmful datasets can lead to misalignment broadly, a phenomenon known as emergent misalignment (EM). EM poses a challenge for AI safety and our understanding of LLMs. Prior work often frames EM as an unexpected behavior, and explains it by appealing to general misalignment directions or anthropomorphizing it as acquiring an evil persona. However, the mechanisms behind these framings remain obscure. In this work, we show that EM is a predictable and data-dependent generalization phenomenon. By examining the base model's representation of EM training data and evaluation prompts, we find that evilness after EM training is highly predictable from representational distance: the closer an evaluation prompt is to training data centroid, the more evilness it elicits from EM models after training (with an average Spearman correlation of -0.73 across 12 model-dataset settings). Building upon this analysis, we further demystify EM by showing that (1) its effectiveness changes significantly based on training data format; (2) there is not a general misalignment direction that transfers across different EM models; (3) the effect of EM is fundamentally different from persona changes. Furthermore, we extend the EM generalization metric from a scalar distance to a dataset-specific generalization direction, which robustly predicts EM models' evilness under semantics-preserving prompt perturbations including appending random tokens and paraphrasing, where other methods do not reliably generalize.