arXiv subjects

Guinan Su

Publications and source records attributed to Guinan Su.

15 recordsLinked to original sources

arXiv2026

Multi-Stream LLMs: Unblocking Language Models with Parallel Streams of Thoughts, Inputs and Outputs

The continued improvements in language model capability have unlocked their widespread use as drivers of autonomous agents, for example in coding or computer use applications. However, the core of these systems has not changed much since early instruction-tuned models like ChatGPT. Even advanced AI agents function on message exchange formats, successively exchanging messages with users, systems, with itself (i.e. chain-of-thought) and tools in a single stream of computation. This bottleneck to a single stream in chat models leads to a number of limitations: the agent cannot act (generate output) while reading, and in reverse, cannot react to new information while writing. Similarly, the agent cannot act while thinking and cannot think while reading or acting on information. In this work, we show that models can be unblocked by switching from instruction-tuning for sequential message formats to instruction-tuning for multiple, parallel streams of computation, splitting each role into a separate stream. Every forward pass of the language model then simultaneously reads from multiple input streams and generates tokens in multiple output streams, all of which causally depend on earlier timesteps. We argue that this data-driven change remedies a number of usability limitations as outlined above, improves model efficiency through parallelization, improves model security through better separation of concerns and can further improve model monitorability.

cs.LG↗

arXiv2026

Efficient Test-Time Inference via Deterministic Exploration of Truncated Decoding Trees

Self-consistency boosts inference-time performance by sampling multiple reasoning traces in parallel and voting. However, in constrained domains like math and code, this strategy is compute-inefficient because it samples with replacement, repeatedly revisiting the same high-probability prefixes and duplicate completions. We propose Distinct Leaf Enumeration (DLE), a deterministic decoding method that treats truncated sampling as traversal of a pruned decoding tree and systematically enumerates distinct leaves instead of sampling with replacement. This strategy improves inference efficiency in two ways. Algorithmically, it increases coverage of the truncated search space under a fixed budget by exploring previously unvisited high-probability branches. Systemically, it reuses shared prefixes and reduces redundant token generation. Empirically, DLE explores higher-quality reasoning traces than stochastic self-consistency, yielding better performance on math, coding, and general reasoning tasks.

cs.LG↗

arXiv2025

When Fewer Layers Break More Chains: Layer Pruning Harms Test-Time Scaling in LLMs

Layer pruning has emerged as a widely adopted technique for improving the efficiency of large language models (LLMs). Although existing methods demonstrate strong performance retention on general knowledge tasks, their effect on long-chain reasoning, a more brittle yet crucial capability, remains largely unexplored. In this work, we study the impact of layer pruning on long-chain reasoning through the lens of test-time scaling, a key mechanism in modern LLMs that enables strong reasoning capacity by allocating more computation at inference time. With extensive experiments, we demonstrate that pruning even one or two layers can severely impair test-time scaling, with performance collapsing drastically on long reasoning benchmarks even when performance on knowledge-intensive and shallow reasoning tasks remains stable. Furthermore, we find that standard supervised fine-tuning remedies fail to recover test-time scaling once it has deteriorated. Through in-depth analyses, we identify the mechanisms underlying this fragility of test-time scaling and highlight the fundamental risks of applying layer pruning to reasoning-intensive LLMs. These findings call for a rethinking of layer pruning strategies and provide insights for developing methods that preserve the robustness of reasoning. We open-source the codebase in \href{https://github.com/keyu-wang-2002/Layer-Pruning-Harms-Inference-Scaling}{https://github.com/keyu-wang-2002/Layer-Pruning-Harms-Inference-Scaling}.

cs.LG↗

arXiv2025

Rewiring Experts on the Fly:Continuous Rerouting for Better Online Adaptation in Mixture-of-Expert models

Mixture-of-Experts (MoE) models achieve efficient scaling through sparse expert activation, but often suffer from suboptimal routing decisions due to distribution shifts in deployment. While existing test-time adaptation methods could potentially address these issues, they primarily focus on dense models and require access to external data, limiting their practical applicability to MoE architectures. However, we find that, instead of relying on reference data, we can optimize MoE expert selection on-the-fly based only on input context. As such, we propose \textit{a data-free, online test-time framework} that continuously adapts MoE routing decisions during text generation without external supervision or data. Our method cycles between two phases: During the prefill stage, and later in regular intervals, we optimize the routing decisions of the model using self-supervision based on the already generated sequence. Then, we generate text as normal, maintaining the modified router until the next adaption. We implement this through lightweight additive vectors that only update router logits in selected layers, maintaining computational efficiency while preventing over-adaptation. The experimental results show consistent performance gains on challenging reasoning tasks while maintaining robustness to context shifts. For example, our method achieves a 5.5\% improvement on HumanEval with OLMoE. Furthermore, owing to its plug-and-play property, our method naturally complements existing test-time scaling techniques, e.g., achieving 6\% average gains when incorporated with self-consistency on DeepSeek-V2-Lite.

cs.CL↗

arXiv2025

Efficient Parallel Samplers for Recurrent-Depth Models and Their Connection to Diffusion Language Models

Language models with recurrent depth, also referred to as universal or looped when considering transformers, are defined by the capacity to increase their computation through the repetition of layers. Recent efforts in pretraining have demonstrated that these architectures can scale to modern language modeling tasks while exhibiting advantages in reasoning tasks. In this work, we examine the relationship between recurrent-depth models and diffusion language models. Building on their similarities, we develop a new diffusion forcing sampler for these models to accelerate generation. The sampler advances by decoding new tokens at every forward pass of the model, while the latent states of these tokens can be further refined in parallel through recurrence. Theoretically, generation with our sampler is strictly more expressive than the baseline autoregressive generation using the same time budget on modern hardware. Moreover, this sampler, based on principles from diffusion literature, can be directly applied to existing 3.5B recurrent-depth transformers without any tuning, leading to up to a 5x speedup. Consequently, our findings not only provide an efficient mechanism for parallelizing the extra computation in recurrent-depth models at inference, but also suggest that such models can be naturally viewed as strong continuous, though causal, diffusion language models.

cs.LG↗

arXiv2025

Sample Smart, Not Hard: Correctness-First Decoding for Better Reasoning in LLMs

Large Language Models (LLMs) are increasingly applied to complex tasks that require extended reasoning. In such settings, models often benefit from diverse chains-of-thought to arrive at multiple candidate solutions. This requires two competing objectives: to inject enough stochasticity to explore multiple reasoning chains, and to ensure sufficient accuracy and quality in each path. Existing works pursue the first objective by increasing exploration at highly uncertain steps with higher temperature or larger candidate token sets, while others improve reliability by rejecting samples with low confidence post-generation, implying that low confidence correlates with low answer quality. These two lines of thought are in conflict, as they conflate different sources of uncertainty. To resolve this, we argue that the decoding rule should be calibrated by correctness, not confidence alone. We should sample from tokens with higher estimated correctness, and reduce sampling where expected correctness is low. We propose simple strategies that achieve this goal: Greedy-Threshold makes sampling greedy at very low confidence steps. Calibrated-TopK and Calibrated-epsilon set truncation threshold based on estimated rank-wise correctness. Together, our findings challenge prevailing heuristics about decoding under uncertainty and show gains across math and general reasoning benchmarks.

cs.LG↗

arXiv2025

aiXiv: A Next-Generation Open Access Ecosystem for Scientific Discovery Generated by AI Scientists

Recent advances in large language models (LLMs) have enabled AI agents to autonomously generate scientific proposals, conduct experiments, author papers, and perform peer reviews. Yet this flood of AI-generated research content collides with a fragmented and largely closed publication ecosystem. Traditional journals and conferences rely on human peer review, making them difficult to scale and often reluctant to accept AI-generated research content; existing preprint servers (e.g. arXiv) lack rigorous quality-control mechanisms. Consequently, a significant amount of high-quality AI-generated research lacks appropriate venues for dissemination, hindering its potential to advance scientific progress. To address these challenges, we introduce aiXiv, a next-generation open-access platform for human and AI scientists. Its multi-agent architecture allows research proposals and papers to be submitted, reviewed, and iteratively refined by both human and AI scientists. It also provides API and MCP interfaces that enable seamless integration of heterogeneous human and AI scientists, creating a scalable and extensible ecosystem for autonomous scientific discovery. Through extensive experiments, we demonstrate that aiXiv is a reliable and robust platform that significantly enhances the quality of AI-generated research proposals and papers after iterative revising and reviewing on aiXiv. Our work lays the groundwork for a next-generation open-access ecosystem for AI scientists, accelerating the publication and dissemination of high-quality AI-generated research content. Code: https://github.com/aixiv-org aiXiv: https://aixiv.science

cs.AI↗

arXiv2025

MedSAMix: A Training-Free Model Merging Approach for Medical Image Segmentation

Universal medical image segmentation models have emerged as a promising paradigm due to their strong generalizability across diverse tasks, showing great potential for a wide range of clinical applications. This potential has been partly driven by the success of general-purpose vision models such as the Segment Anything Model (SAM), which has inspired the development of various fine-tuned variants for medical segmentation tasks. However, fine-tuned variants like MedSAM are trained on comparatively limited medical imaging data that often suffers from heterogeneity, scarce annotations, and distributional shifts. These challenges limit their ability to generalize across a wide range of medical segmentation tasks. In this regard, we propose MedSAMix, a training-free model merging method that integrates the strengths of both generalist models (e.g., SAM) and specialist models (e.g., MedSAM) for medical image segmentation. In contrast to traditional model merging approaches that rely on manual configuration and often result in suboptimal outcomes, we propose a zero-order optimization method to automatically discover optimal layer-wise merging solutions. Furthermore, for clinical applications, we develop two regimes to meet the demand of domain-specificity and generalizability in different scenarios by single-task optimization and multi-objective optimization respectively. Extensive evaluations on 25 medical segmentation tasks demonstrate that MedSAMix effectively mitigates model bias and consistently improves performance in both domain-specific accuracy and generalization, achieving improvements of 6.67% on specialized tasks and 4.37% on multi-task evaluations.

cs.CV↗

arXiv2025

GPTailor: Large Language Model Pruning Through Layer Cutting and Stitching

Large language models (LLMs) have shown remarkable capabilities in language understanding and generation. However, such impressive capability typically comes with a substantial model size, which presents significant challenges in deployment and inference. While structured pruning of model parameters offers a promising way to reduce computational costs at deployment time, current methods primarily focus on single model pruning. In this work, we develop a novel strategy to compress models by strategically combining or merging layers from finetuned model variants, which preserves the original model's abilities by aggregating capabilities accentuated in different finetunes. We pose the optimal tailoring of these LLMs as a zero-order optimization problem, adopting a search space that supports three different operations: (1) Layer removal, (2) Layer selection from different candidate models, and (3) Layer merging. Our experiments demonstrate that this approach leads to competitive model pruning, for example, for the Llama2-13B model families, our compressed models maintain approximately 97.3\% of the original performance while removing $\sim25\%$ of parameters, significantly outperforming previous state-of-the-art methods. The code is available at https://github.com/Guinan-Su/auto-merge-llm.

cs.CL↗

arXiv2025

Fine, I'll Merge It Myself: A Multi-Fidelity Framework for Automated Model Merging

Reasoning capabilities represent a critical frontier for large language models (LLMs), but developing them requires extensive proprietary datasets and computational resources. One way to efficiently supplement capabilities with is by model merging, which offers a promising alternative by combining multiple models without retraining. However, current merging approaches rely on manually-designed strategies for merging hyperparameters, limiting the exploration of potential model combinations and requiring significant human effort. We propose an Automated Model Merging Framework that enables fine-grained exploration of merging strategies while reducing costs through multi-fidelity approximations. We support both single and multi-objective optimization and introduce two novel search spaces: layerwise fusion (LFS) and depth-wise integration (DIS). Evaluating across a number of benchmarks, we find that the search autonomously finds 1) Merges that further boost single-objective performance, even on tasks the model has already been finetuned on, and 2) Merges that optimize multi-objective frontiers across tasks. Effective merges are found with limited compute, e.g. within less than 500 search steps.

cs.AI↗

arXiv2024

BrainMass: Advancing Brain Network Analysis for Diagnosis with Large-scale Self-Supervised Learning

Foundation models pretrained on large-scale datasets via self-supervised learning demonstrate exceptional versatility across various tasks. Due to the heterogeneity and hard-to-collect medical data, this approach is especially beneficial for medical image analysis and neuroscience research, as it streamlines broad downstream tasks without the need for numerous costly annotations. However, there has been limited investigation into brain network foundation models, limiting their adaptability and generalizability for broad neuroscience studies. In this study, we aim to bridge this gap. In particular, (1) we curated a comprehensive dataset by collating images from 30 datasets, which comprises 70,781 samples of 46,686 participants. Moreover, we introduce pseudo-functional connectivity (pFC) to further generates millions of augmented brain networks by randomly dropping certain timepoints of the BOLD signal. (2) We propose the BrainMass framework for brain network self-supervised learning via mask modeling and feature alignment. BrainMass employs Mask-ROI Modeling (MRM) to bolster intra-network dependencies and regional specificity. Furthermore, Latent Representation Alignment (LRA) module is utilized to regularize augmented brain networks of the same participant with similar topological properties to yield similar latent representations by aligning their latent embeddings. Extensive experiments on eight internal tasks and seven external brain disorder diagnosis tasks show BrainMass's superior performance, highlighting its significant generalizability and adaptability. Nonetheless, BrainMass demonstrates powerful few/zero-shot learning abilities and exhibits meaningful interpretation to various diseases, showcasing its potential use for clinical applications.

cs.CE↗

arXiv2023

DualTalker: A Cross-Modal Dual Learning Approach for Speech-Driven 3D Facial Animation

In recent years, audio-driven 3D facial animation has gained significant attention, particularly in applications such as virtual reality, gaming, and video conferencing. However, accurately modeling the intricate and subtle dynamics of facial expressions remains a challenge. Most existing studies approach the facial animation task as a single regression problem, which often fail to capture the intrinsic inter-modal relationship between speech signals and 3D facial animation and overlook their inherent consistency. Moreover, due to the limited availability of 3D-audio-visual datasets, approaches learning with small-size samples have poor generalizability that decreases the performance. To address these issues, in this study, we propose a cross-modal dual-learning framework, termed DualTalker, aiming at improving data usage efficiency as well as relating cross-modal dependencies. The framework is trained jointly with the primary task (audio-driven facial animation) and its dual task (lip reading) and shares common audio/motion encoder components. Our joint training framework facilitates more efficient data usage by leveraging information from both tasks and explicitly capitalizing on the complementary relationship between facial motion and audio to improve performance. Furthermore, we introduce an auxiliary cross-modal consistency loss to mitigate the potential over-smoothing underlying the cross-modal complementary representations, enhancing the mapping of subtle facial expression dynamics. Through extensive experiments and a perceptual user study conducted on the VOCA and BIWI datasets, we demonstrate that our approach outperforms current state-of-the-art methods both qualitatively and quantitatively. We have made our code and video demonstrations available at https://github.com/sabrina-su/iadf.git.

cs.CV↗

arXiv2023

Prompt Your Mind: Refine Personalized Text Prompts within Your Mind

Large language models (LLMs) have demonstrated remarkable potential in natural language understanding and generation, making them valuable tools for enhancing conversational interactions. However, LLMs encounter challenges such as lacking multi-step reasoning capabilities, and heavy reliance on prompts. In this regard, we introduce a prompt-refinement system named PromptMind, also known as "Prompt Your Mind", to provide an automated solution for generating contextually relevant prompts during conversations. PromptMind enhances the overall interaction between humans and chatbots through an automatic prompt suggestion and an automatic prompt refinement. To assess the effectiveness of PromptMind, we designed three interaction tasks to evaluate emotional support, advice acquisition, and task-oriented interactions during human-chatbot interactions. The results demonstrated that PromptMind reduced mental demands during interactions and fostered enhanced performance and social connections between users and chatbots. In summary, our findings indicate that PromptMind acts as a bridge, facilitating smoother information exchange and enhancing the usability of chatbot interactions.

cs.HC↗

arXiv2021

AlphaGAN: Fully Differentiable Architecture Search for Generative Adversarial Networks

Generative Adversarial Networks (GANs) are formulated as minimax game problems, whereby generators attempt to approach real data distributions by virtue of adversarial learning against discriminators. The intrinsic problem complexity poses the challenge to enhance the performance of generative networks. In this work, we aim to boost model learning from the perspective of network architectures, by incorporating recent progress on automated architecture search into GANs. To this end, we propose a fully differentiable search framework for generative adversarial networks, dubbed alphaGAN. The searching process is formalized as solving a bi-level minimax optimization problem, in which the outer-level objective aims for seeking a suitable network architecture towards pure Nash Equilibrium conditioned on the generator and the discriminator network parameters optimized with a traditional GAN loss in the inner level. The entire optimization performs a first-order method by alternately minimizing the two-level objective in a fully differentiable manner, enabling architecture search to be completed in an enormous search space. Extensive experiments on CIFAR-10 and STL-10 datasets show that our algorithm can obtain high-performing architectures only with 3-GPU hours on a single GPU in the search space comprised of approximate 2 ? 1011 possible configurations. We also provide a comprehensive analysis on the behavior of the searching process and the properties of searched architectures, which would benefit further research on architectures for generative models. Pretrained models and codes are available at https://github.com/yuesongtian/AlphaGAN.

cs.CV↗

arXiv2019

TextNAS: A Neural Architecture Search Space tailored for Text Representation

Learning text representation is crucial for text classification and other language related tasks. There are a diverse set of text representation networks in the literature, and how to find the optimal one is a non-trivial problem. Recently, the emerging Neural Architecture Search (NAS) techniques have demonstrated good potential to solve the problem. Nevertheless, most of the existing works of NAS focus on the search algorithms and pay little attention to the search space. In this paper, we argue that the search space is also an important human prior to the success of NAS in different applications. Thus, we propose a novel search space tailored for text representation. Through automatic search, the discovered network architecture outperforms state-of-the-art models on various public datasets on text classification and natural language inference tasks. Furthermore, some of the design principles found in the automatic network agree well with human intuition.

cs.LG↗