Training & Fine-tuning A

The paper presents Probe-and-Refine, a method for tuning the repository guidance (such as AGENTS.md files) that LLM-based coding agents rely on. It targets the higher-level operational knowledge—file layout, test workflows, and error-prone patterns—that is not contained in the code itself.

Style-content dual-reference generation aims to synthesize an image that preserves structure while adopting a reference style. FreeStyle leverages community LoRA mining to give free control over style and content.

It is unclear whether LLMs that score well on vulnerability benchmarks truly reason about security or merely pattern-match. This work diagnoses the limits of fine-tuning LLMs for vulnerability detection in systems software.

The paper proposes aligning large language models using implicit user signals—such as mouse and eye movements—instead of explicit human feedback. It addresses the limitation that users rarely provide explicit ratings, which makes high-quality preference data scarce for reward modeling.

The paper studies how to train spatially grounded vision-language models for radiology without manual spatial annotations. It introduces RefRad2D, a large-scale bilingual (German/English) dataset of 1.2M CT and MR image-text pairs derived from clinical practice, with VQA and spatial grounding subsets.

The paper proposes evolutionary two-stage hyperparameter optimization strategies for physics-informed neural networks (PINNs). It targets PINNs' unstable convergence, training plateaus, and strong sensitivity to architectural and optimization hyperparameters arising from their highly non-convex training.

Data videos combine dynamic charts, voice narration, and synchronized animation to convey insights. DataMagic automatically transforms tabular data into such data-insight videos.

FP4 training promises large memory and compute savings for LLM pretraining but suffers from shrinkage bias. This paper analyzes its geometric origin and systemic impact and proposes a UFP4 recipe to address it.

Large language models show promise for code compilation tasks but struggle with runtime performance tuning. AutoPass uses evidence-guided LLM agents to perform compiler performance tuning.

Explicit skill libraries make computer-using agents easier to inspect, but building them is costly. This work automates SKILL.md generation by mining agents' interaction trajectories.

The paper runs a four-arm head-to-head comparison of fine-tuning, retrieval, and their combination for producing correct statutory citations on the Ontario Residential Tenancies Act and its core regulation. It targets the practical need of tenants, landlords, and help-desk staff to be pointed at the governing provision.

Leveraging multimodal large language models through contrastive learning has become mainstream for retrieval. ELVA explores a ranking-driven approach to universal multimodal retrieval.

Vision-Language-Action models pre-trained on massive video-robot datasets have transformed robot control. This work shows that finetuning them requires fewer layers than previously assumed.

Academic paper search is a core step in research, and LLM-based search agents are emerging. ScholarQuest provides a taxonomy-guided benchmark for agentic academic paper search in open literature environments.

The paper investigates whether emergent misalignment—induced by fine-tuning language models on insecure code—corresponds to a causally actionable, shared direction in activation space. Across four instruction-tuned model families, it studies using such directions to detect and mitigate the misalignment.

Using vision foundation models for camera-to-LiDAR knowledge distillation is promising. HilDA advances self-supervised LiDAR pre-training through hierarchical distillation with diffusion.

The paper presents IHUBERT, a monolingual Persian pretrained language model trained from scratch on a RoBERTa-base encoder. It uses vector-based semantic deduplication and domain-balanced pretraining to address the scarcity of large, high-quality Persian corpora and limited evaluation.

The paper pre-trains the Tsetlin Machine using semantic clusters drawn from language models to improve interpretability in text classification. It aims to combine the transparency of the Tsetlin Machine's clause-based reasoning with the semantic information that models like BERT capture but do not expose transparently.

The paper studies sequential Direct Preference Optimization (DPO) across different preference settings, examining how applying multiple alignment objectives one after another affects earlier ones. It looks beyond uniform forgetting to understand how later training stages interfere with previously learned preferences.

Hugging Face examines whether any approach can beat LoRA, the most popular fine-tuning technique. It compares alternative parameter-efficient methods on performance and cost, probing experimentally whether they can rival LoRA and offering guidance for practitioners.

Passive long-video models 'watch it all,' processing frames uniformly so cost grows with duration regardless of query difficulty. This work treats perception as reasoning, with native active perception that selectively attends to relevant frames for efficient omni-modal understanding.

Preference-based RL learns reward models from pairwise behavior comparisons, bypassing explicit reward design, but existing methods often rely on passive data collection. UBP2 introduces uncertainty-balanced preference planning to actively select comparisons and learn efficiently from fewer preferences.

As LLMs are adapted to specialized domains and languages, the effectiveness of adaptation strategies remains unclear. This empirical study on French medical question answering analyzes the trade-offs of various domain-adaptation methods, clarifying gains and losses in performance and generality.

Existing 3D scene understanding in VLMs relies on complex, model-specific geometry encoders or large training budgets for spatial reasoning. OneCanvas instead uses panoramic reprojection, letting VLMs reason about 3D scenes efficiently without dedicated geometry encoders or heavy training.

AI agents promise to accelerate drug discovery by compressing interpretation and decision loops, but deployment needs trusted evaluation on realistic tasks. TxBench-PP is a benchmark analyzing AI agent performance on small-molecule preclinical pharmacology, assessing their practical reliability.

Reinforcement learning with verifiable rewards, such as GRPO, dominates post-training for complex LLM reasoning but often suffers policy entropy collapse. STARE introduces surprisal-guided token-level advantage reweighting to stabilize policy entropy and preserve exploration during training.

The authors propose MAST (Mechanism-Aligned Selective Targeting), a mechanism-guided method for unlearning RLVR-induced reasoning with substantially less collateral damage than standard full-parameter updates, removing targeted reasoning while preserving other capabilities.

Personal memory in a language model involves two problems: content and reasoning skill, which the brain keeps apart—a sparse local hippocampal engram per episode and slow neocortical skill. Inspired by this, the work internalizes per-user memory as local parametric edits to the model.

The authors introduce Dango, a 1.8B-parameter language model designed for controlled studies of L1-to-L2 (Japanese-to-English) transfer in second language acquisition. By training strictly on L1 only, Dango enables controlled experiments on transfer phenomena that prior SLA model studies could not.

To achieve deeper safety alignment for LLMs, recent work pushes safety interventions earlier into pretraining, mainly by filtering unsafe data or rewriting it into safe forms. Going beyond safe data, this work embeds regular safety reflection during pretraining to instill more fundamental alignment.