MAGIC: Achieving Superior Model Merging via Magnitude Calibration

The proliferation of pre-trained models has given rise to a wide array of specialised, fine-tuned models. Model merging aims to merge the distinct capabilities of these specialised models into a unified model, requiring minimal or even no additional training. A core objective of model merging is to ensure the merged model retains the behavioural characteristics of the specialised models, typically achieved through feature alignment. We identify that features consist of two critical components: direction and magnitude. Prior research has predominantly focused on directional alignment, while the influence of magnitude remains largely neglected, despite its pronounced vulnerability to perturbations introduced by common merging operations (e.g., parameter fusion and sparsification). Such perturbations to magnitude inevitably lead to feature deviations in the merged model from the specialised models, resulting in subsequent performance degradation. To address this, we propose MAGnItude Calibration (MAGIC), a plug-and-play framework that rectifies layer-wise magnitudes in feature and weight spaces, with three variants. Specifically, our Feature Space Calibration (FSC) realigns the merged model's features using a small set of unlabelled data, while Weight Space Calibration (WSC) extends this calibration to the weight space without requiring additional data. Combining these yields Dual Space Calibration (DSC). Comprehensive experiments demonstrate that MAGIC consistently boosts performance across diverse Computer Vision tasks (+4.3% on eight datasets) and NLP tasks (+8.0% on Llama) without additional training. Our code is available at: https://github.com/lyymuwu/MAGIC

Physics of Language Models: Part 4.1, Architecture Design and the Magic of Canon Layers

Understanding architectural differences in language models is challenging, especially at academic-scale pretraining (e.g., 1.3B parameters, 100B tokens), where results are often dominated by noise and randomness. To overcome this, we introduce controlled synthetic pretraining tasks that isolate and evaluate core model capabilities. Within this framework, we discover CANON LAYERS: lightweight architectural components -- named after the musical term "canon" -- that promote horizontal information flow across neighboring tokens. Canon layers compute weighted sums of nearby token representations and integrate seamlessly into Transformers, linear attention, state-space models, or any sequence architecture. We present 12 key results. This includes how Canon layers enhance reasoning depth (e.g., by $2\times$), reasoning breadth, knowledge manipulation, etc. They lift weak architectures like NoPE to match RoPE, and linear attention to rival SOTA linear models like Mamba2/GDN -- validated both through synthetic tasks and real-world academic-scale pretraining. This synthetic playground offers an economical, principled path to isolate core model capabilities often obscured at academic scales. Equipped with infinite high-quality data, it may even PREDICT how future architectures will behave as training pipelines improve -- e.g., through better data curation or RL-based post-training -- unlocking deeper reasoning and hierarchical inference.

MAGIC-Talk: Motion-aware Audio-Driven Talking Face Generation with Customizable Identity Control

Audio-driven talking face generation has gained significant attention for applications in digital media and virtual avatars. While recent methods improve audio-lip synchronization, they often struggle with temporal consistency, identity preservation, and customization, especially in long video generation. To address these issues, we propose MAGIC-Talk, a one-shot diffusion-based framework for customizable and temporally stable talking face generation. MAGIC-Talk consists of ReferenceNet, which preserves identity and enables fine-grained facial editing via text prompts, and AnimateNet, which enhances motion coherence using structured motion priors. Unlike previous methods requiring multiple reference images or fine-tuning, MAGIC-Talk maintains identity from a single image while ensuring smooth transitions across frames. Additionally, a progressive latent fusion strategy is introduced to improve long-form video quality by reducing motion inconsistencies and flickering. Extensive experiments demonstrate that MAGIC-Talk outperforms state-of-the-art methods in visual quality, identity preservation, and synchronization accuracy, offering a robust solution for talking face generation.

What do AI-Generated Images Want?

W.J.T. Mitchell's influential essay 'What do pictures want?' shifts the theoretical focus away from the interpretative act of understanding pictures and from the motivations of the humans who create them to the possibility that the picture itself is an entity with agency and wants. In this article, I reframe Mitchell's question in light of contemporary AI image generation tools to ask: what do AI-generated images want? Drawing from art historical discourse on the nature of abstraction, I argue that AI-generated images want specificity and concreteness because they are fundamentally abstract. Multimodal text-to-image models, which are the primary subject of this article, are based on the premise that text and image are interchangeable or exchangeable tokens and that there is a commensurability between them, at least as represented mathematically in data. The user pipeline that sees textual input become visual output, however, obscures this representational regress and makes it seem like one form transforms into the other -- as if by magic.

Reconquering Bell sampling on qudits: stabilizer learning and testing, quantum pseudorandomness bounds, and more

Bell sampling is a simple yet powerful tool based on measuring two copies of a quantum state in the Bell basis, and has found applications in a plethora of problems related to stabiliser states and measures of magic. However, it was not known how to generalise the procedure from qubits to $d$-level systems -- qudits -- for all dimensions $d > 2$ in a useful way. Indeed, a prior work of the authors (arXiv'24) showed that the natural extension of Bell sampling to arbitrary dimensions fails to provide meaningful information about the quantum states being measured. In this paper, we overcome the difficulties encountered in previous works and develop a useful generalisation of Bell sampling to qudits of all $d\geq 2$. At the heart of our primitive is a new unitary, based on Lagrange's four-square theorem, that maps four copies of any stabiliser state $|\mathcal{S}\rangle$ to four copies of its complex conjugate $|\mathcal{S}^\ast\rangle$ (up to some Pauli operator), which may be of independent interest. We then demonstrate the utility of our new Bell sampling technique by lifting several known results from qubits to qudits for any $d\geq 2$: 1. Learning stabiliser states in $O(n^3)$ time with $O(n)$ samples; 2. Solving the Hidden Stabiliser Group Problem in $\tilde{O}(n^3/\varepsilon)$ time with $\tilde{O}(n/\varepsilon)$ samples; 3. Testing whether $|\psi\rangle$ has stabiliser size at least $d^t$ or is $\varepsilon$-far from all such states in $\tilde{O}(n^3/\varepsilon)$ time with $\tilde{O}(n/\varepsilon)$ samples; 4. Clifford circuits with at most $n/2$ single-qudit non-Clifford gates cannot prepare pseudorandom states; 5. Testing whether $|\psi\rangle$ has stabiliser fidelity at least $1-\varepsilon_1$ or at most $1-\varepsilon_2$ with $O(d^2/\varepsilon_2)$ samples if $\varepsilon_1 = 0$ or $O(d^2/\varepsilon_2^2)$ samples if $\varepsilon_1 = O(d^{-2})$.

Where Do I 'Add the Egg'?: Exploring Agency and Ownership in AI Creative Co-Writing Systems

AI co-writing systems challenge long held ideals about agency and ownership in the creative process, thereby hindering widespread adoption. In order to address this, we investigate conceptions of agency and ownership in AI creative co-writing. Drawing on insights from a review of commercial systems, we developed three co-writing systems with identical functionality but distinct interface metaphors: agentic, tool-like, and magical. Through interviews with professional and non-professional writers (n = 18), we explored how these metaphors influenced participants' sense of control and authorship. Our analysis resulted in a taxonomy of agency and ownership subtypes and underscore how tool-like metaphors shift writers' expected points of control while agentic metaphors foreground conceptual contributions. We argue that interface metaphors not only guide expectations of control but also frame conceptions of authorship. We conclude with recommendations for the design of AI co-writing systems, emphasizing how metaphor shapes user experience and creative practice.

Contrastive Learning with Enhanced Abstract Representations using Grouped Loss of Abstract Semantic Supervision

Humans can recognize an image as an instance of a general concept, beyond simply identifying its objects and their relationships. In this paper, we investigate 1. The extent to which VLMs have this concept abstraction capacity, and 2. Strategies for encoding the sort of higher-concept information in images that would enable the resulting VLM model (CLEAR GLASS model) to have this capability to a greater degree. To this end, we introduce a grouped image-caption dataset (MAGIC), which consists of several groups of image captions and for each group a set of associated images and higher-level conceptual labels. We use a novel contrastive loss technique to induce the model to encode in the representation of each image (caption) in a group the information that is common to all members of the image-caption group. Our main contribution is a grouped contrastive loss function based on text-image contrastive groups (outer contrastive loss) as well as an inner loss which measures the distances between image-caption instances in the group. Our training methodology results in the CLEAR GLASS model having the concept abstraction capacity as an emergent capacity because the model is not exposed to the higher-level concepts associated with each group. Instead, the training forces the model to create for each image-caption group a semantic representation that brings it closer to the semantic representation of the higher-level concepts in the latent semantic space. Our experiments show that this training methodology results in a model which shows improvement in abstract concept recognition compared to SOTA models.

The Better You Learn, The Smarter You Prune: Towards Efficient Vision-language-action Models via Differentiable Token Pruning

We present LightVLA, a simple yet effective differentiable token pruning framework for vision-language-action (VLA) models. While VLA models have shown impressive capability in executing real-world robotic tasks, their deployment on resource-constrained platforms is often bottlenecked by the heavy attention-based computation over large sets of visual tokens. LightVLA addresses this challenge through adaptive, performance-driven pruning of visual tokens: It generates dynamic queries to evaluate visual token importance, and adopts Gumbel softmax to enable differentiable token selection. Through fine-tuning, LightVLA learns to preserve the most informative visual tokens while pruning tokens which do not contribute to task execution, thereby improving efficiency and performance simultaneously. Notably, LightVLA requires no heuristic magic numbers and introduces no additional trainable parameters, making it compatible with modern inference frameworks. Experimental results demonstrate that LightVLA outperforms different VLA models and existing token pruning methods across diverse tasks on the LIBERO benchmark, achieving higher success rates with substantially reduced computational overhead. Specifically, LightVLA reduces FLOPs and latency by 59.1% and 38.2% respectively, with a 2.9% improvement in task success rate. Meanwhile, we also investigate the learnable query-based token pruning method LightVLA* with additional trainable parameters, which also achieves satisfactory performance. Our work reveals that as VLA pursues optimal performance, LightVLA spontaneously learns to prune tokens from a performance-driven perspective. To the best of our knowledge, LightVLA is the first work to apply adaptive visual token pruning to VLA tasks with the collateral goals of efficiency and performance, marking a significant step toward more efficient, powerful and practical real-time robotic systems.

MAGIC-Enhanced Keyword Prompting for Zero-Shot Audio Captioning with CLIP Models

Automated Audio Captioning (AAC) generates captions for audio clips but faces challenges due to limited datasets compared to image captioning. To overcome this, we propose the zero-shot AAC system that leverages pre-trained models, eliminating the need for extensive training. Our approach uses a pre-trained audio CLIP model to extract auditory features and generate a structured prompt, which guides a Large Language Model (LLM) in caption generation. Unlike traditional greedy decoding, our method refines token selection through the audio CLIP model, ensuring alignment with the audio content. Experimental results demonstrate a 35% improvement in NLG mean score (from 4.7 to 7.3) using MAGIC search with the WavCaps model. The performance is heavily influenced by the audio-text matching model and keyword selection, with optimal results achieved using a single keyword prompt, and a 50% performance drop when no keyword list is used.

Fine-Grained Customized Fashion Design with Image-into-Prompt benchmark and dataset from LMM

Generative AI evolves the execution of complex workflows in industry, where the large multimodal model empowers fashion design in the garment industry. Current generation AI models magically transform brainstorming into fancy designs easily, but the fine-grained customization still suffers from text uncertainty without professional background knowledge from end-users. Thus, we propose the Better Understanding Generation (BUG) workflow with LMM to automatically create and fine-grain customize the cloth designs from chat with image-into-prompt. Our framework unleashes users' creative potential beyond words and also lowers the barriers of clothing design/editing without further human involvement. To prove the effectiveness of our model, we propose a new FashionEdit dataset that simulates the real-world clothing design workflow, evaluated from generation similarity, user satisfaction, and quality. The code and dataset: https://github.com/detectiveli/FashionEdit.