Training & Fine-tuning A

Audio LLMs can condition speech recognition on textual prompts such as domain descriptions or entity lists, but whether they truly use this context is unclear. IndicContextEval is a benchmark evaluating context utilisation in audio large language models across eight Indic languages.

Identifying the lowest-energy surface-adsorbate configuration is critical for modeling heterogeneous catalysis, but exhaustive ab initio exploration is prohibitive. AdsMind is a physics-grounded multi-agent system that self-corrects to efficiently discover adsorption configurations.

This paper develops local certificates for population-risk increments around a current model. For a local candidate set, the certificate provides a two-sided confidence bound on the change in population risk, giving theoretical guarantees on the risk impact of local model updates.

Team science holds leadership is contingent—helpful only under specific conditions, and unneeded by capable autonomous teams. Asking the analogous question for multi-agent LLMs, this work frames leadership as coordination control, characterizing its behavioral signatures and the recovery-advantage boundary.

Sequence modeling is increasingly popular in recommendation and ranking for its ability to model users' historical behaviors and infer intentions. This work proposes JourneyFormer, which encodes the Airbnb guest journey with sequence modeling to better understand behavior and improve recommendations.

Instruction-based image editing enables complex edits from natural language, but in product-centric scenarios preserving product features and branding is hard. ProductConsistency uses supervised fine-tuning and reinforcement learning to improve product identity preservation during instruction-based editing.

Widespread parameter-efficient fine-tuning yields ecosystems where one backbone pairs with many task-specialized adapters. ARIADNE provides agnostic routing for inference-time dynamic adapter selection, choosing the right adapter per input without model-specific assumptions.

In vibe coding, an emerging AI-driven paradigm, an LLM generates an entire program from a natural-language prompt—but where does the variability that traditional software engineering manages go? This work uses regeneration to move from vibe coding toward software product lines.

Reinforcement learning post-training of Diffusion Transformers is prohibitively expensive, needing thousands of high-end GPUs. Spotlight synergizes seed exploration with cheap, preemptible spot GPUs to substantially reduce the cost of DiT RL post-training.

Reinforcement learning with verifiable rewards has become standard for reasoning models. GraphPO introduces a graph-based policy optimization method that exploits structure across reasoning steps to improve reasoning performance.

Context engineering has become a primary lever for improving AI systems. SAGE is a stochastic prompt optimization method that uses agent-guided exploration to automatically discover effective prompts and improve task performance.

Large language models memorize and reproduce sequences from their training data. This work edits output vectors to mitigate such memorization, reducing the risk of leaking copyrighted or private content.

Preferred Networks continues developing its large language model PLaMo and shares how to use the latest PLaMo-3.0-Prime-β in real development work. Beyond training large models, it covers the many surrounding tasks involved in building high-performance LLMs in practice.

GitLab announced Project Switch, a next-generation Git-compatible source code management service aimed at AI agents, at its GitLab Transcend event in London. Reports cite up to 50x speed and roughly half the token usage; figures reflect the announcement and remain unverified.

The paper proposes a self-evolving zero-shot object-goal navigation framework that builds an agentic rule memory by extracting actionable knowledge from past trajectories and uses a retrieval strategy to enable continuous test-time improvement.

Darshana Graph is a corpus of over 125,000 text records spanning classical Hindu, Buddhist and Jain philosophical traditions, drawn from public-domain and openly licensed translations. It supports comparative Indian philosophy through stylometric and exploratory graph analyses.

On-policy self-distillation (OPSD) helps post-training of LLMs but is unexplored for diffusion LLMs (dLLMs). Existing OPSD methods are autoregressive-centric, injecting privileged information via left-to-right prefix conditioning; this work studies self-distillation suited to dLLMs.

The work builds a dataset of multi-source cybersecurity logs labeled with MITRE ATT&CK and evaluates small language models (SLMs) on it. Summary is title-based and neutral; details and figures are as presented by the source and not independently verified.

The paper proposes WEQA, a framework for question answering over wearable health sensor data using query-adaptive agentic reasoning, arguing that diverse sensor modalities and user intents cannot be handled by a fixed reasoning workflow.

Structured state space models such as S4 and S4D capture long-range dependencies but are hard to deploy on constrained devices. S4oP introduces operator-level pruning to enable efficient deployment of SSMs on time- and resource-constrained hardware.

The paper presents EAGG, an embodiment-aligned grasp generator that represents each end-effector with a topology-aware graph and embodiment-specific conditioning, aiming to generalize grasp generation across objects and diverse robot embodiments.

Post-training pipelines combining supervised fine-tuning with reinforcement learning are key to turning LLMs into robust reasoners. The paper studies compositional generalization in LM reasoning by converting reasoning traces into reusable modules.

Vision-language-action models combine vision-language backbones with expressive generative action heads trained via flow matching on large robotic datasets. Despite strong performance, the paper studies uncertainty quantification for these flow-based VLA models.

Cross-lingual transfer is well studied under supervised fine-tuning, where data and linguistic similarity drive quality. As the field shifts to few-shot in-context learning, this paper examines source-language effects and shows English is not always the best teacher.

Catastrophic forgetting in continual adaptation is usually viewed via parameter drift or replay, which do not reveal which output directions are vulnerable. The paper gives a function-space account in the NTK regime, showing new-task training drifts old-task predictions low-rank through the cross-task kernel.

Conditional independence testing is fundamental to statistics and causal inference. The paper proposes a fast nonparametric conditional independence test based on two-stage regression, aiming to improve computational efficiency and power.

The paper fine-tunes LLMs for passive estimation of depression severity from AI mental-health dialogue, exploring how conversational signals can indicate severity. Figures and efficacy are as reported by the source and not independently verified.

Kolmogorov-Arnold Networks replace linear weights with learnable univariate functions but their high computational cost hampers practical research. KANLib provides a modular, extensible and fast implementation of KANs to ease experimentation.

LLM agents in interactive environments are sensitive to prompts, yet prompt engineering stays manual and task-specific. The paper decomposes the observation-to-action pipeline and proposes an environment-grounded automated prompt optimization framework for LLM game agents.

The study examines the extent to which self-supervised speech models exhibit perceptual compensation for tonal context, analyzing how context effects seen in human speech perception are reflected in the models' learned representations.