Searcharxiv⌕ Search

arXiv subjects

Lavinia Ghita

Publications and source records attributed to Lavinia Ghita.

5 recordsLinked to original sources

Do Pretraining Optimizer Gains Transfer to KL Distillation?

Matrix-structured optimizers can improve language-model pretraining, but it is unclear whether these gains transfer to short-horizon KL distillation from an instruction-tuned initialization. We compare AdamW and diagonal AdaGrad with two tractable Shampoo reductions, SOAP, and All-2D Muon across four dense Qwen student-teacher pairs under a common forward-KL objective and fixed training budget. Across scales, none of the evaluated matrix-structured recipes consistently improves the quality versus trainer-time trade-off over the elementwise references: AdamW generally attains the lowest validation KL, while AdaGrad can be slightly faster at larger scale. SOAP often tracks AdamW in KL but incurs substantial trainer-time overhead, and the two Shampoo reductions reverse ordering across model pairs, showing that conclusions depend on the feasible recipe. These results suggest that pretraining optimizer gains do not transfer automatically to KL distillation and should be re-established for the target objective and systems regime.

cs.LG↗

Reasoning Supervision and Capability Retention in LoRA Domain Adaptation

Fine-tuning foundation models on financial data has a wide range of applications, including identifying event types from news, monitoring risk and compliance language, and answering questions over filings and earnings calls. In this setting, the main challenge is to improve task performance without degrading the generalization capabilities of the model. In this paper, we present an extensive set of ablations for fine-tuning Qwen3-8B with low-rank adaptation (LoRA). We compare three training regimes: (1) response-only LoRA, where the supervised response is the label alone; (2) single-adapter reasoning, where reasoning traces from a teacher model are included alongside the labels in the training loss; and (3) dual-adapter composite training, where separate LoRA adapters are trained on traces and labels. Response-only LoRA improves in-domain performance and leaves general capabilities near those of the frozen model. The in-domain change is largest for calibration and argmax accuracy. Under single-adapter reasoning, the in-domain metrics remain behind those of the response-only regime, while MMLU-Pro can rise slightly and the other retention benchmarks stay within ordinary variation. Dual-adapter composite training remains on the same side of this comparison and yields no further gain. Finally, we compress the teacher traces by summarizing them before training, and show that shorter summaries move calibration toward the response-only regime without reaching it.

cs.AI↗

Effective Teacher Support for Reasoning Distillation

Forward-KL language-model distillation normally computes dense teacher and student scores over the full vocabulary at every supervised position, making training and stored teacher targets expensive. We instead distill from the teacher's highest-probability vocabulary fraction, renormalizing the retained mass as we sweep support from 100% to 0.01%. Our main study distills Qwen3 4B -> 1.7B. We confirm the broader support-reduction finding on larger Qwen3 8B -> 4B and Nemotron Nano 12B -> 9B pairs and under answer-only supervision. We evaluate generation primarily on MMLU-Pro and secondarily on GPQA-Diamond. Across the Qwen3 settings, retaining 0.01% of the vocabulary causes no observed MMLU-Pro degradation relative to full-vocabulary distillation, although disagreement with the complete teacher distribution increases. Nemotron Nano matches full-vocabulary MMLU-Pro at 0.1%, but smaller fractions fail to emit the closing reasoning tag, showing that the reliable cutoff depends on the model family. At 0.01% support, caching the retained next-token scores from the teacher forward pass cuts training step time by 38%, while restricting student output computation to those tokens cuts peak memory by 24%; neither causes observed MMLU-Pro degradation.

cs.AI↗

Scaling Laws for Task-Specific LLM Distillation

Large Language Models (LLMs) achieve strong performance across a growing range of domains, yet their scale poses deployment challenges in applications where latency and cost constraints are critical. This paper derives empirical scaling laws for domain-specific LLM compression, quantifying how in-domain and general knowledge performance scale with dataset size, compression ratio, supervision format, and iterative pruning schedule. Using quantitative finance as our application domain, we compare logit-based and LoRA-based distillation under iterative structural pruning, introducing a blended chain-of-thought supervision loss that stabilizes KL-divergence distillation over reasoning traces. In-domain task quality degrades predictably under compression while general-knowledge benchmarks collapse well before the same point; supervision format is the key driver of this tradeoff, with chain-of-thought supervision actively recovering general knowledge that pruning erases. We release the headline dataset FinHeadlineMix, scaling law results, and practical recommendations to provide a reusable framework for domain-specific compression decisions.

cs.AI↗

Low-Rank Dependence Decomposition via Accelerated Symmetric Non-negative Matrix Factorization

Symmetric non-negative matrix factorization (SymNMF) recovers latent group structure from a dependence matrix, but its dense, quadratic-memory objective has confined prior work to moderate sizes. We present a large-scale GPU study of seven algorithm families (over 30 configurations) on absolute Pearson correlation and tail pairwise dependence matrices from Extreme Value Theory, two proxies for empirical risk-factor estimation on large portfolios. A trace-identity reformulation eliminates all $n \times n$ intermediates, so a single GPU reaches $n \approx 10^5$ and multi-node distribution scales to $n = 10^6$ and beyond. Under a two-phase protocol, eleven methods converge at moderate scale; six remain efficient enough at $n = 10^5$ (five AdaGrad-family plus ADMM), and five AdaGrad-family methods still converge at $n = 10^6$: AdaGrad, RMSprop, and three we introduce (Piecewise AdaGrad, Row-Stochastic SVRG, Block-SVRG AdaptGrow). At $n = 10^6$ the fastest solver tracks the matrix spectrum: Block-SVRG AdaptGrow wins on the flat, ill-conditioned tail-dependence spectrum, where its lower per-iteration cost decides a long factorization, and full-batch AdaGrad wins on the dominant-low-rank correlation spectrum, where the run is short. We also benchmark spherical K-means as a hard-label baseline: cheaper when angular cluster structure is present, yet provably degenerate once the matrix collapses toward a single common factor, where the soft factorization remains necessary.

cs.LG↗