arXiv subjects

Nikita Kozodoi

Publications and source records attributed to Nikita Kozodoi.

12 recordsLinked to original sources

arXiv2026

Model Merging Does Not Require Uniform LoRA Ranks

Model merging combines separately fine-tuned experts into a single model without retraining. Standard practice trains every expert at the same LoRA rank, which couples the rank choice across experts even though each domain has its own best rank. We show this coupling is a property of the merge method, not of merging. We train 36 experts on three domains (mathematics, instruction following, multilingual) at four ranks and three model scales, then merge with six methods at uniform ranks and at each domain's own best rank. Giving experts different ranks costs averaging and utility-based selection nothing, and gains them up to 1.8 pp of multi-domain accuracy, while the conflict-resolution methods are inconsistent and can lose up to 4 pp. The split has a simple cause. Merging keeps only a fixed share of the pooled spectral components, and methods that choose them by size discard many that would have reduced loss. They also tend to discard the same expert: the instruction adapter has the smallest singular values, so it is dropped almost entirely and IFEval falls by 16 to 25 pp. Choosing components by their effect on loss instead keeps all three experts, and leaves accuracy nearly flat across an 8$\times$ rank range (0.7 pp, against 6.7 pp for TIES at 4B scale). Rank can therefore be chosen per expert whenever the merge operator does not select by magnitude.

cs.LG↗

arXiv2026

NaijaFacts: A Framework for Disentangling Language from Knowledge in Facts about Nigeria

Multilingual benchmarks for African languages mostly translate generic questions into local languages, measuring linguistic competence but not whether models know facts about Africa. We introduce \naija, a benchmark of factual and statistical questions about Nigeria, authored by native speakers and grounded in authoritative official statistics, with every question available in English, Nigerian Pidgin, \yoruba, Igbo, and Hausa. Answering similar questions for the UK and US yields a country by language matrix that disentangles a language effect (the model knows the fact but degrades in a low-resource language) from a knowledge effect (the fact is absent in every language). Across eight frontier and Nigeria-adapted models, the knowledge effect dominates: English-query accuracy is 49--50\% for the US/UK but 20\% for Nigeria. The language effect is smaller and tracks web representation, degrading from English to \yoruba{} (9 points on Nigeria, up to 24 on US/UK questions). We find scale does not help: the strongest models reach 66--79\% on the UK and the US but none exceed 31\% on Nigerian statistics. Extending to all 54 African countries with our best performing frontier model shows the same profile continent-wide, 96\% on culture and 98\% on politics against 26\% on statistics, so the gap is not specific to Nigeria.

cs.CL↗

arXiv2026

Generation and Verification: Human-Agent Collaboration for Safety-Critical Automotive Software Engineering

Research on coding agents mainly focuses on generation quality. Yet, safety-critical automotive software makes two additional demands. It prioritizes verification, since every change is reviewed, traced to a requirement, and checked against industry standards such as MISRA and AUTOSAR before it can be committed. It is also a narrow domain, written against layered architectures, embedded resource constraints, and organization-specific interface conventions that are scarce in the data general-purpose LLMs are trained on. We present a coding assistant for automotive C/C++ that addresses both demands. First, it exposes three tools to the user over the Model Context Protocol (generate, edit, and verify), allowing developers to run verify on any code on demand rather than relying on the agent to check its own output. Second, rather than using a general-purpose model, we customize the underlying LLM for the automotive domain through prompt engineering and fine-tuning. We construct a dataset from automotive production code of AUMOVIO, where each function is paired with the requirement texts and Unified Modeling Language (UML) behavior it must satisfy. In a two-week pilot on active production projects, developers report verify as the most-used tool and confirm using it on their own code and not only on generated output. Comparing model customization strategies, 75% prefer the faster prompt-engineered configuration. MISRA- and AUTOSAR-inspired pattern checkers indicate that prompt engineering achieves a lower number of violations on C, while fine-tuning shows better conformance with architecture patterns for C++.

cs.SE↗

arXiv2026

Test-Time Augmentation for LLMs: Where to Inject Diversity at Inference Time

Aggregating several LLM responses to the same question can improve accuracy, but every method that does so must decide where the responses should differ. Established test-time scaling methods such as self-consistency vary only the reasoning path, resampling from a fixed input. The input modification offers several further options: the wording of the question can be rephrased, the spelling can be perturbed, and the accompanying images can be transformed. We study Test-Time Augmentation (TTA), which aggregates predictions across transformed versions of the input, and use it to ask where diversity is best injected. Comparing these options across six benchmarks at a matched number of inference calls, we find that where the diversity comes from matters more than how much of it there is. Meaning-preserving paraphrases of the input yield the largest accuracy gains (+1.80 pp over a single call) and reach peak accuracy with the fewest samples. Yet, they are not the most diverse: character-level noise makes the responses disagree far more often (37.8% versus 27.4%) while gaining less accuracy, and perturbing two modalities at once falls below the single-call baseline. Not all diversity is equally useful: perturbations that keep the meaning intact produce disagreement that aggregation can use, while those that corrupt it produce disagreement that is wrong. Overall, our results suggest that for mid-tier LLMs such as Claude Haiku, varying how a question is phrased is a better use of an inference call than varying the reasoning path. Our TTA implementation is available at github.com/aws-samples/sample-genai-reflection-for-bedrock.

cs.LG↗

arXiv2026

Evolving LLM-Generated Features for Interpretable Classification

Large language models (LLMs) are increasingly used as classifiers, yet they operate as opaque systems whose decisions are difficult to interpret, which complicates their use in regulated domains such as credit scoring or medical diagnosis. We propose an evolutionary framework that iteratively discovers natural language feature definitions (rubrics) for interpretable classification. An LLM generates candidate binary features, evaluates each sample against them, and the resulting vectors can be used to train a transparent classifier such as logistic regression. The feature set evolves over multiple iterations guided by classification errors, per-class activation rates, and feature ablation scores. We evaluate across three benchmarks, including a credit risk dataset representative of regulated domains, comparing single-shot LLM rubrics, evolved rubrics, and direct zero-shot LLM classification. Evolved features improve over single-shot rubrics by +2.9 pp on average and outperform zero-shot LLM classification on two of three tasks, while providing fully auditable decision logic. On the credit risk task, the zero-shot LLM performs at chance (50.7%) with a strong bias toward a single class, whereas evolved features achieve balanced, interpretable predictions. Crucially, this failure is invisible in aggregate accuracy and surfaces only under per-class auditing. Analysis reveals that evolution is most effective when label boundaries cannot be inferred from category names alone or when the LLM lacks reliable domain-specific reasoning.

cs.LG↗

arXiv2026

Are we Merging the Right Models? Impact of Expert Training Duration on Model Merging for LLMs

Multi-task model merging combines separately trained expert models into a single model that handles all tasks without co-training. Standard practice merges experts at their optimal validation loss. We challenge this convention by systematically studying how training duration of domain experts affects the quality of the merged model. We fine-tune experts on five domains (Math, Code, Instruction Following, Multilingual, and Safety) across three model sizes (Qwen 3.5 0.8B, 2B, and 4B), saving checkpoints from 25% to 500% of the optimal training steps and evaluating five merging methods at each duration. Our findings reveal a striking method-dependent pattern: simple averaging degrades sharply with overfitting, while sparsification-based methods achieve their best performance well past the validation optimum. We formalize this through bias-variance decomposition analysis, drawing a parallel to random forests where averaging benefits from high-variance individual learners. These results suggest that training duration and merging method should be chosen jointly rather than independently.

cs.LG↗

arXiv2026

SWE-InfraBench: Evaluating Language Models on Cloud Infrastructure Code

Building infrastructure-as-code (IaC) in cloud computing is a critical task, underpinning the reliability, scalability, and security of modern software systems. Despite the remarkable progress of large language models (LLMs) in software engineering -- demonstrated across many dedicated benchmarks -- their capabilities in developing IaC remain underexplored. Unlike existing IaC benchmarks that predominantly center on declarative paradigms such as Terraform and involve generating entire codebases from scratch, our benchmark reflects the incremental code edits common in enterprise development with imperative tools like the AWS CDK. We present SWE-InfraBench, a diverse evaluation dataset sourced from dozens of real-world IaC codebases that challenge LLMs to perform realistic code modifications in AWS CDK repositories. Each example requires models to implement changes to existing codebases based on natural language instructions, with success determined by passing provided test cases. These tasks demand sophisticated reasoning about cloud resource dependencies and implementation patterns beyond conventional code generation challenges. Our evaluation results reveal significant limitations in current LLMs showing that even state-of-the-art systems struggle with many tasks -- the best model, Sonnet 3.7, succeeds in only 34\% of cases, while specialized reasoning models like DeepSeek R1 achieve just 24% success. The SWE-InfraBench dataset is available at: https://www.kaggle.com/datasets/64e59070fd51c0278560b01eb5dc4f3c447d5268cdabe5a350d2969e4413fea5

cs.SE↗

arXiv2025

Finding the Sweet Spot: Trading Quality, Cost, and Speed During Inference-Time LLM Reflection

As Large Language Models (LLMs) continue to evolve, practitioners face increasing options for enhancing inference-time performance without model retraining, including budget tuning and multi-step techniques like self-reflection. While these methods improve output quality, they create complex trade-offs among accuracy, cost, and latency that remain poorly understood across different domains. This paper systematically compares self-reflection and budget tuning across mathematical reasoning and translation tasks. We evaluate prominent LLMs, including Anthropic Claude, Amazon Nova, and Mistral families, along with other models under varying reflection depths and compute budgets to derive Pareto optimal performance frontiers. Our analysis reveals substantial domain dependent variation in self-reflection effectiveness, with performance gains up to 220\% in mathematical reasoning. We further investigate how reflection round depth and feedback mechanism quality influence performance across model families. To validate our findings in a real-world setting, we deploy a self-reflection enhanced marketing content localisation system at Lounge by Zalando, where it shows market-dependent effectiveness, reinforcing the importance of domain specific evaluation when deploying these techniques. Our results provide actionable guidance for selecting optimal inference strategies given specific domains and resource constraints. We open source our self-reflection implementation for reproducibility at https://github.com/aws-samples/sample-genai-reflection-for-bedrock.

stat.ML↗

arXiv2024

Fighting Sampling Bias: A Framework for Training and Evaluating Credit Scoring Models

Scoring models support decision-making in financial institutions. Their estimation and evaluation are based on the data of previously accepted applicants with known repayment behavior. This creates sampling bias: the available labeled data offers a partial picture of the distribution of candidate borrowers, which the model is supposed to score. The paper addresses the adverse effect of sampling bias on model training and evaluation. To improve scorecard training, we propose bias-aware self-learning - a reject inference framework that augments the biased training data by inferring labels for selected rejected applications. For scorecard evaluation, we propose a Bayesian framework that extends standard accuracy measures to the biased setting and provides a reliable estimate of future scorecard performance. Extensive experiments on synthetic and real-world data confirm the superiority of our propositions over various benchmarks in predictive performance and profitability. By sensitivity analysis, we also identify boundary conditions affecting their performance. Notably, we leverage real-world data from a randomized controlled trial to assess the novel methodologies on holdout data that represent the true borrower population. Our findings confirm that reject inference is a difficult problem with modest potential to improve scorecard performance. Addressing sampling bias during scorecard evaluation is a much more promising route to improve scoring practices. For example, our results suggest a profit improvement of about eight percent, when using Bayesian evaluation to decide on acceptance rates.

stat.ML↗

arXiv2024

Probabilistic Demand Forecasting with Graph Neural Networks

Demand forecasting is a prominent business use case that allows retailers to optimize inventory planning, logistics, and core business decisions. One of the key challenges in demand forecasting is accounting for relationships and interactions between articles. Most modern forecasting approaches provide independent article-level predictions that do not consider the impact of related articles. Recent research has attempted addressing this challenge using Graph Neural Networks (GNNs) and showed promising results. This paper builds on previous research on GNNs and makes two contributions. First, we integrate a GNN encoder into a state-of-the-art DeepAR model. The combined model produces probabilistic forecasts, which are crucial for decision-making under uncertainty. Second, we propose to build graphs using article attribute similarity, which avoids reliance on a pre-defined graph structure. Experiments on three real-world datasets show that the proposed approach consistently outperforms non-graph benchmarks. We also show that our approach produces article embeddings that encode article similarity and demand dynamics and are useful for other downstream business tasks beyond forecasting.

cs.LG↗

arXiv2022

Fairness in Credit Scoring: Assessment, Implementation and Profit Implications

The rise of algorithmic decision-making has spawned much research on fair machine learning (ML). Financial institutions use ML for building risk scorecards that support a range of credit-related decisions. Yet, the literature on fair ML in credit scoring is scarce. The paper makes three contributions. First, we revisit statistical fairness criteria and examine their adequacy for credit scoring. Second, we catalog algorithmic options for incorporating fairness goals in the ML model development pipeline. Last, we empirically compare different fairness processors in a profit-oriented credit scoring context using real-world data. The empirical results substantiate the evaluation of fairness measures, identify suitable options to implement fair credit scoring, and clarify the profit-fairness trade-off in lending decisions. We find that multiple fairness criteria can be approximately satisfied at once and recommend separation as a proper criterion for measuring the fairness of a scorecard. We also find fair in-processors to deliver a good balance between profit and fairness and show that algorithmic discrimination can be reduced to a reasonable level at a relatively low cost. The codes corresponding to the paper are available on GitHub.

stat.ML↗

arXiv2019

Shallow Self-Learning for Reject Inference in Credit Scoring

Credit scoring models support loan approval decisions in the financial services industry. Lenders train these models on data from previously granted credit applications, where the borrowers' repayment behavior has been observed. This approach creates sample bias. The scoring model (i.e., classifier) is trained on accepted cases only. Applying the resulting model to screen credit applications from the population of all borrowers degrades model performance. Reject inference comprises techniques to overcome sampling bias through assigning labels to rejected cases. The paper makes two contributions. First, we propose a self-learning framework for reject inference. The framework is geared toward real-world credit scoring requirements through considering distinct training regimes for iterative labeling and model training. Second, we introduce a new measure to assess the effectiveness of reject inference strategies. Our measure leverages domain knowledge to avoid artificial labeling of rejected cases during strategy evaluation. We demonstrate this approach to offer a robust and operational assessment of reject inference strategies. Experiments on a real-world credit scoring data set confirm the superiority of the adjusted self-learning framework over regular self-learning and previous reject inference strategies. We also find strong evidence in favor of the proposed evaluation measure assessing reject inference strategies more reliably, raising the performance of the eventual credit scoring model.

stat.ML↗