arXiv AI

Everything in Moderation: Per-Domain Coverage Optima and Alignment-Resistant Domain Gaps in Multi-Domain Mid-Training

The study investigates how the composition of data during the mid‑training phase of language models affects performance across multiple domains. Experiments with Qwen3‑8B‑Base on five distinct KOR‑Bench domains show that moderate coverage (10%‑40%) yields the best per‑domain results, and that alignment passes cannot fully close the performance gaps created by mid‑training data choices. Additionally, zero coverage in mid‑training severely degrades accuracy, while a carefully tuned allocation can provide the largest overall pipeline improvement.

Hugging Face Trending Papers
Sep 8

Everything in Moderation: Per-Domain Coverage Optima and Alignment-Resistant Domain Gaps in Multi-Domain Mid-Training

The paper investigates how per-domain data composition during the mid‑training phase (between pre‑training and alignment) affects model performance. Experiments with Qwen3‑8B‑Base across five KOR‑Bench domains show that a moderate coverage band (10%‑40%) yields the best performance for each domain, and that alignment passes cannot fully close the gaps created by suboptimal mid‑training allocations. Additionally, zero coverage during mid‑training severely degrades accuracy, while a carefully tuned allocation can provide the largest overall pipeline improvement.

arXiv AI
3d ago

How many LLM-labelled rows does a decision encoder need, and how wrong can the labels be?

The study investigates how many large‑language‑model (LLM)‑generated training rows a decision encoder, specifically Laya (ModernBERT‑large), requires to achieve high accuracy on a choice‑making task. Using 19 fine‑tunes on nested subsets of a 2,818‑row pool, accuracy improves from 69.4 % before fine‑tuning to 86.9 % with all rows, with diminishing returns as more data is added. The experiment also examines the impact of label noise, finding that 10 % symmetric noise has negligible effect, while 20 % and 40 % noise degrade performance by 3.4 and 8.7 points respectively, and that a single run with 10 % confusable‑pair noise reduces confidence. whyItMatters":"The results quantify the data efficiency and robustness of Laya to noisy LLM‑generated labels, informing how much synthetic training data is needed for reliable decision‑encoding models."

By Mohammed Nayeem Khan, Aliyas Ahmad Misri
arXiv Machine Learning
Nov 5

The Alignment Tax Is Graded: Task-Selective Capability Loss Under Length-Normalized DPO with SFT Retention

The study empirically investigates how length‑normalized Direct Preference Optimization (DPO) with a Structured Fine‑Tuning (SFT) retention term (RPO objective) affects downstream capabilities of a Qwen3.5‑4B model fine‑tuned with sparse LoRA on structured financial‑filing tasks. Results show that the alignment tax is graded and task‑selective: extraction performance is largely preserved, JSON extraction degrades modestly, and router‑based tasks suffer severe loss, with most degradation attributed to suppression of rejected completions. The analysis uses 8 seeds, artifact‑aware scoring, and TOST equivalence testing to quantify these effects across different reward margins and precision settings.

By Ye Yuan
arXiv AI
Oct 16

Rethinking Calibration for High-Sparsity LLM Pruning: From Single Sources to Capability Coverage

The paper argues that post‑training pruning of large language models should use a calibration set that covers multiple capability domains rather than a single source. A balanced four‑way mixture of sources consistently outperforms any single source across sparsity levels, model scales, and held‑out benchmarks. When aligned corpora are unavailable, the IGSP method approximates this mixture using only the base model and an evaluation taxonomy, achieving significant gains over prior self‑generation approaches.

By Hu Xu, Zhaolong Xing, Congcong Liu, Jiaxing Wang, Zhida Jiang, Junshi Huang, Zhen Chen, Jianfeng Xu
arXiv Machine Learning
Oct 16

Selection-Safe Conformal Early Exit: Distribution-Free Coverage under Adaptive Depth

The paper introduces a selection‑safe conformal early‑exit method that calibrates a single split‑conformal quantile using a terminal‑path score derived from prefix maxima. This approach guarantees coverage for any data‑dependent stopping policy that preserves terminal‑set containment, without requiring an α/L correction. Empirical results on eight encoder backbones and CIFAR‑100 show that pathwise selection achieves higher macro coverage (0.901 vs. 0.814) with smaller set sizes and lower latency fractions compared to naive layer‑wise selection, and remains robust to deployment‑time budget changes.

By Adria Binte Habib, Ibne Farabi Shihab
arXiv AI
Aug 25

Right-Sizing LLM-Agent Decomposition in VAT Determination: A Pilot Controlled Sweep

The study evaluates how to best split tasks among large‑language‑model agents for cross‑border VAT determination, comparing one broad agent to configurations ranging from one to five narrow agents. Across 4,400 runs—including token‑matched and failure‑injection scenarios—the intermediate configurations achieved the highest accuracy but did not surpass the fine‑endpoint benchmark, leaving the optimal decomposition hypothesis unconfirmed. The pilot provides a preregistered heuristic for right‑sizing decomposition, along with an oracle, dataset, and analysis pipeline.

By Pedro Santos
arXiv AI
Sep 25

Wiring Beats Blending: Structure-Aware Compensation for Transformer Downscaling

The paper investigates converting a large pretrained transformer (1.4 B parameters) into a smaller sibling (410 M) by studying representation alignment and parameter projection. It finds that dense weight projection destroys structure, and that a low‑budget, structure‑aware compensation—separating least‑squares function alignment from variance‑preserving rescaling—yields significant gains on token‑efficient training, outperforming subcloning and standard distillation pipelines at matched budgets.

By Ravi Satya Durga Prasad Yenugula
arXiv Machine Learning
Aug 20

Learned, Then Lost: A Measured Single-Example Counterfactual in Pre-training

The study measured the impact of a single training example on a GPT‑2 model by running 24 counterfactual experiments. 32 models were trained from scratch on OpenWebText, and at a specific training step a single batch row was replaced with a 194‑token passage under three conditions (fluent prose, fabricated subject, random characters) or left unchanged. Results showed that the passage was learned from one exposure and decayed, with measurable differences in cross‑entropy up to 50 steps after injection but no lasting effect at the final step.

By Zachary Speck, Asa Shepard