LoRA, PEFT, instruction tuning and domain adaptation — adapting a pretrained model without paying to train one.
Using large language models (LLMs) to simulate diverse human populations has the potential to transform many aspects of computational social science, yet many evaluations score the average response ra...
As large language models (LLMs) scale rapidly, dense full-parameter adaptation becomes increasingly expensive, motivating sparse and modular architectures such as Mixture-of-Experts (MoE) models. This...
The paper introduces SCOPE, a method that post‑trains computer‑use agents to balance task completion with safety by conditioning actions on environmental risk. It combines supervised fine‑tuning on three trajectory types—capability demonstrations, safe continuations, and explicit refusals—followed by reinforcement learning to improve performance. Experiments starting from Qwen3.5‑9B show that SCOPE‑RL achieves high task success and attack‑avoidance rates, outperforming other agents on OSWorld and OS‑BLIND benchmarks.
The paper introduces MMSAFE, a multi-layer framework designed to identify safety-degrading data in multilingual large language models. It shows that safety signals are distributed across multiple layers and only partially shared across languages, unlike the single-layer assumption used in monolingual settings. Experiments demonstrate that MMSAFE reduces harmful-response rates by 60% compared to random filtering and outperforms the best single-layer baseline across various models, languages, and safety benchmarks.
The paper introduces FedPFT, a federated learning framework that tackles the feature‑classifier mismatch problem by using personalized prompts processed through a shared self‑attention transformation module. Unlike prior methods that either degrade the feature extractor or address the mismatch only after training, FedPFT aligns local features with the global classifier during training, improving aggregation and model performance. Experiments demonstrate that FedPFT surpasses state‑of‑the‑art methods by up to 5.07%, and gains up to 7.08% when combined with collaborative contrastive learning.
The paper introduces Chart‑RVR, a reinforcement‑learning framework that trains chart‑reasoning agents to produce monitorable, verifiable outputs. It decomposes reasoning into three auditable stages—Structure, Evidence, and Derivation—allowing stakeholders to trace how the model reads the chart, extracts data, and computes the answer. Experiments on six benchmarks show that Chart‑RVR matches or exceeds state‑of‑the‑art accuracy while delivering rationales that are far more verifiable and evidence‑grounded than existing methods.
The paper introduces Bidirectional Reciprocal Learning (BRL), a parameter‑efficient fine‑tuning framework for referring image segmentation that operates on frozen vision foundation models. BRL employs two lightweight adapters—Reciprocal Attention Adapter (RAA) for token‑level cross‑modal attention and Reciprocal Gate Adapter (RGA) for channel‑level gating—to enable hierarchical, bidirectional information flow between vision and language. Experiments on RefCOCO, RefCOCO+, and RefCOCOg show that BRL outperforms existing methods while updating fewer than 0.5% of backbone parameters.
The paper introduces SADA-Merging, a new data‑free model merging framework that addresses limitations of existing subspace‑based methods. It allocates subspace capacity based on each task’s spectral complexity, adapts spectral preservation to task‑specific plasticity, and applies depth‑dependent anchoring to mitigate projection distortion. Experiments show that SADA‑Merging consistently outperforms current data‑free merging techniques across various task scales and adaptation settings.
VideoGen-Agent is a multimodal agent that uses multitask agentic reinforcement learning to coordinate external tools for video generation. It learns to augment, generate, and verify videos through multi‑turn interactions, guided by prompts and intermediate observations. On the new VABench benchmark, the agent improves base text‑to‑video performance by 19.1 points, and further upgrades to generation tools raise the score to 86.1, with human raters favoring the upgraded configuration in 84.3% of comparisons.
The paper introduces GradCIR, a method for training composed image retrieval (CIR) systems on graded relevance rather than binary relevance. It uses a vision‑language model to generate queries and 4‑level relevance labels, an iterative feedback loop to mine hard negatives, and a hierarchy‑aware angular objective to directly optimize graded labels. Experiments on a Walmart catalog and FashionIQ show significant NDCG improvements and the system is deployed in Walmart’s live visual‑search traffic.
SafeTune is a source‑available library that consolidates four safety‑intervention paradigms—post‑hoc weight recovery, safety‑constrained fine‑tuning, gradient‑based unlearning, and inference‑time steering—into a single, configuration‑driven workflow. It offers shared interpretability, evaluation, and deployment tools, and its modular registry allows easy addition of new methods, benchmarks, judges, models, and fine‑tuning domains. The authors demonstrate SafeTune with controlled comparisons and case studies in finance and medical deployments, showing how it characterizes safety drift, evaluates interventions on refusal‑behavior and capability metrics, and supports calibrated or layered mitigation.
The paper investigates how post‑training modifies the weights of Large Language Models relative to their pretrained state. By expressing weight updates in the pretrained matrix’s singular value decomposition, the authors separate changes into three geometric components: diagonal (reshaping singular values), off‑diagonal (rotating input‑output coupling), and null‑space (routing outside the original SVD core). Experiments on a math evaluation suite show that removing the diagonal component largely preserves post‑training gains, indicating that improvements stem mainly from reconfiguring and extending pretrained pathways rather than altering singular values.
CAMFT is a Conflict‑Aware Mergeable Fine‑Tuning method designed to make task adaptation efficient and merge‑aware for large language models. Unlike existing approaches that only resolve parameter conflicts after fine‑tuning, CAMFT shapes mergeability during training by guiding each task to update sparse coordinates with lower cross‑task conflict. Experiments show that CAMFT outperforms standard fine‑tuning baselines in multi‑task merging scenarios.
arXiv:2609.24209v1 Announce Type: new Abstract: The Platonic Representation Hypothesis (PRH) claims that independently trained models converge on a shared statistical model of reality, yet recent wor...
arXiv:2609.22215v1 Announce Type: cross Abstract: Knowledge distillation can transmit unintended behavioral traits from a teacher model to a student through training data that appear semantically unr...
arXiv:2609.23376v1 Announce Type: cross Abstract: Open-vocabulary audio-visual event localization (OV-AVEL) grounds a text-queried event in time from video, audio, and language. The supervision sourc...
arXiv:2609.23377v1 Announce Type: cross Abstract: Repository-level software engineering (SWE) comprises heterogeneous task categories, whose progress under pooled agentic reinforcement learning can b...
arXiv:2609.23753v1 Announce Type: cross Abstract: Generative world models aim to predict future states conditioned on actions, where action controllability is fundamental for reliable dynamics modeli...
arXiv:2609.24302v1 Announce Type: cross Abstract: Virtual screening aims to prioritize active compounds from large chemical libraries within a limited experimental budget. When applying Boltz-2 to vi...
arXiv:2609.24517v1 Announce Type: cross Abstract: Model merging aims to combine multiple fine-tuned models derived from a common pretrained model into a single multi-task model without additional joi...