arXiv subjects

Liyuan Liu

Publications and source records attributed to Liyuan Liu.

At least 19 recordsLinked to original sources

arXiv2026

Ling and Ring 2.6 Technical Report: Efficient and Instant Agentic Intelligence at Trillion-Parameter Scale

Efficient and scalable agentic intelligence requires models that can deliver both low-latency responses and strong reasoning capabilities while remaining practical to train, serve, and deploy. In this report, we present Ling-2.6 and Ring-2.6, a family of models designed to address this challenge at scale. Ling-2.6 is optimized for instant response generation and high capability per output token, whereas Ring-2.6 is tailored for deeper reasoning and more advanced agentic workflows. Instead of training from scratch, we upgrade the Ling-2.0 base model through architectural migration pre-training and large-scale post-training. This upgrade is guided by a unified co-design of model architecture, optimization objectives, serving systems, and agent training environments, enabling improvements in both model capability and deployment efficiency. At the architectural level, we introduce a hybrid linear attention design that integrates Lightning Attention with MLA, improving the efficiency of long-context training and decoding. To further enhance token efficiency, we optimize capability per output token through Evolutionary Chain-of-Thought, Linguistic Unit Policy Optimization, bidirectional preference alignment, and shortest-correct-response distillation. For agentic capabilities, we propose KPop, a reinforcement learning framework designed to support stable training of Ring-2.6-1T on large-scale environment-grounded data. KPop improves training efficiency through asynchronous scheduling across coding, search, tool use, and workflow execution, enabling scalable learning from complex agent-environment interactions. Together, Ling-2.6 and Ring-2.6 provide a practical pathway toward efficient, scalable, and open agentic systems. We open-source all checkpoints in the 2.6 family to support further research and development in practical agentic intelligence.

cs.CL↗

arXiv2026

Quantifying Multimodal Capabilities: Formal Generalization Guarantees in Pairwise Metric Learning

Multimodal learning leverages the integration of diverse data modalities to enhance performance in complex tasks. Yet, it frequently encounters incomplete or redundant modality data in real-world scenarios. This paper presents a fine-grained theoretical analysis of the generalization properties of multimodal metric learning models, addressing critical gaps in understanding the relationship between modality selection and algorithmic performance. We establish hierarchical relationships between function classes corresponding to different modality subsets and quantify the discrepancy between learned mappings and ground truth. Through rigorous analysis of pairwise complexity within the multimodal learning framework, we derive novel generalization error bounds that reveal the joint impact of modality quantity and granularity on model performance. Our theoretical findings on both upper and lower bounds demonstrate that incorporating fine-grained modality features reduces the complexity of the hypothesis space by enhancing modality complementarity. This work offers both theoretical foundations and practical implications for improving convergence rates and accuracy in multimodal learning systems.

cs.LG↗

arXiv2026

Adaptation of Agentic AI: A Survey of Post-Training, Memory, and Skills

Large language model (LLM) agents are moving beyond prompting alone. ChatGPT marked the rise of general-purpose LLM assistants, DeepSeek showed that on-policy reinforcement learning with verifiable rewards can improve reasoning and tool use, and OpenClaw highlights a newer direction in which agents accumulate persistent memory and reusable skills. Yet the research landscape remains fragmented across post-training, retrieval, memory, and skill systems. This survey studies these developments under a single notion of \emph{adaptation}: improving an agent, its tools, or their interaction after pretraining. We organize the field with a four-paradigm framework spanning agent adaptation and tool adaptation. On the agent side, A1 (tool-execution-signaled) and A2 (agent-output-signaled) improve the agent itself through supervised fine-tuning, preference optimization, and reinforcement learning with verifiable rewards. On the tool side, T1 (agent-agnostic) provides reusable pre-trained modules any agent can call, while T2 (agent-supervised) uses the agent's outputs to train memory systems, skill libraries, or lightweight subagents. Using this framework, we review post-training methods, adaptive memory architectures, and agent skills; compare their trade-offs in cost, flexibility, and generalization; and summarize evaluation practices across deep research, software development, computer use, and drug discovery. We conclude by outlining open problems in agent-tool co-adaptation, continual learning, safety, and efficient deployment.

cs.AI↗

arXiv2026

Test-time Recursive Thinking: Self-Improvement without External Feedback

Modern Large Language Models (LLMs) have shown rapid improvements in reasoning capabilities, driven largely by reinforcement learning (RL) with verifiable rewards. Here, we ask whether these LLMs can self-improve without the need for additional training. We identify two core challenges for such systems: (i) efficiently generating diverse, high-quality candidate solutions, and (ii) reliably selecting correct answers in the absence of ground-truth supervision. To address these challenges, we propose Test-time Recursive Thinking (TRT), an iterative self-improvement framework that conditions generation on rollout-specific strategies, accumulated knowledge, and self-generated verification signals. Using TRT, open-source models reach 100% accuracy on AIME-25/24, and on LiveCodeBench's most difficult problems, closed-source models improve by 10.4-14.8 percentage points without external feedback.

cs.CL↗

arXiv2025

A Digital SRAM-Based Compute-In-Memory Macro for Weight-Stationary Dynamic Matrix Multiplication in Transformer Attention Score Computation

Compute-in-memory (CIM) techniques are widely employed in energy-efficient artificial intelligent (AI) processors. They alleviate power and latency bottlenecks caused by extensive data movements between compute and storage units. To extend these benefits to Transformer, this brief proposes a digital CIM macro to compute attention score. To eliminate dynamic matrix multiplication (MM), we reconstruct the computation as static MM using a combined QK-weight matrix, so that inputs can be directly fed to a single CIM macro to obtain the score results. However, this introduces a new challenge of 2-input static MM. The computation is further decomposed into four groups of bit-serial logical and addition operations. This allows 2-input to directly activate the word line via AND gate, thus realizing 2-input static MM with minimal overhead. A hierarchical zero-value bit skipping mechanism is introduced to prioritize skipping zero-value bits in the 2-input case. This mechanism effectively utilizes data sparsity of 2-input, significantly reducing redundant operations. Implemented in a 65-nm process, the 0.35 mm2 macro delivers 42.27 GOPS at 1.24 mW, yielding 34.1 TOPS/W energy and 120.77 GOPS/mm2 area efficiency. Compared to CPUs and GPUs, it achieves ~25x and ~13x higher efficiency, respectively. Against other Transformer-CIMs, it demonstrates at least 7x energy and 2x area efficiency gains, highlighting its strong potential for edge intelligence.

cs.AR↗

arXiv2025

Decoder-Hybrid-Decoder Architecture for Efficient Reasoning with Long Generation

Recent advances in language modeling have demonstrated the effectiveness of State Space Models (SSMs) for efficient sequence modeling. While hybrid architectures such as Samba and the decoder-decoder architecture, YOCO, have shown promising performance gains over Transformers, prior works have not investigated the efficiency potential of representation sharing between SSM layers. In this paper, we introduce the Gated Memory Unit (GMU), a simple yet effective mechanism for efficient memory sharing across layers. We apply it to create SambaY, a decoder-hybrid-decoder architecture that incorporates GMUs in the cross-decoder to share memory readout states from a Samba-based self-decoder. SambaY significantly enhances decoding efficiency, preserves linear pre-filling time complexity, and boosts long-context performance, all while eliminating the need for explicit positional encoding. Through extensive scaling experiments, we demonstrate that our model exhibits a significantly lower irreducible loss compared to a strong YOCO baseline, indicating superior performance scalability under large-scale compute regimes. Our largest model enhanced with Differential Attention, Phi4-mini-Flash-Reasoning, achieves significantly better performance than Phi4-mini-Reasoning on reasoning tasks such as Math500, AIME24/25, and GPQA Diamond without any reinforcement learning, while delivering up to 10x higher decoding throughput on 2K-length prompts with 32K generation length under the vLLM inference framework. We release our training codebase on open-source data at https://github.com/microsoft/ArchScale.

cs.CL↗

arXiv2025

Reinforcement Learning for Reasoning in Large Language Models with One Training Example

We show that reinforcement learning with verifiable reward using one training example (1-shot RLVR) is effective in incentivizing the math reasoning capabilities of large language models (LLMs). Applying RLVR to the base model Qwen2.5-Math-1.5B, we identify a single example that elevates model performance on MATH500 from 36.0% to 73.6% (8.6% improvement beyond format correction), and improves the average performance across six common mathematical reasoning benchmarks from 17.6% to 35.7% (7.0% non-format gain). This result matches the performance obtained using the 1.2k DeepScaleR subset (MATH500: 73.6%, average: 35.9%), which contains the aforementioned example. Furthermore, RLVR with only two examples even slightly exceeds these results (MATH500: 74.8%, average: 36.6%). Similar substantial improvements are observed across various models (Qwen2.5-Math-7B, Llama3.2-3B-Instruct, DeepSeek-R1-Distill-Qwen-1.5B), RL algorithms (GRPO and PPO), and different math examples. In addition, we identify some interesting phenomena during 1-shot RLVR, including cross-category generalization, increased frequency of self-reflection, and sustained test performance improvement even after the training accuracy has saturated, a phenomenon we term post-saturation generalization. Moreover, we verify that the effectiveness of 1-shot RLVR primarily arises from the policy gradient loss, distinguishing it from the "grokking" phenomenon. We also show the critical role of promoting exploration (e.g., by incorporating entropy loss with an appropriate coefficient) in 1-shot RLVR training. We also further discuss related observations about format correction, label robustness and prompt modification. These findings can inspire future work on RLVR efficiency and encourage a re-examination of recent progress and the underlying mechanisms in RLVR. All resources are open source at https://github.com/ypwang61/One-Shot-RLVR.

cs.LG↗

arXiv2025

Text Generation Beyond Discrete Token Sampling

In standard autoregressive generation, an LLM predicts the next-token distribution, samples a discrete token, and then discards the distribution, passing only the sampled token as new input. To preserve this distribution's rich information, we propose Mixture of Inputs (MoI), a training-free method for autoregressive generation. After generating a token following the standard paradigm, we construct a new input that blends the generated discrete token with the previously discarded token distribution. Specifically, we employ a Bayesian estimation method that treats the token distribution as the prior, the sampled token as the observation, and replaces the conventional one-hot vector with the continuous posterior expectation as the new model input. MoI allows the model to maintain a richer internal representation throughout the generation process, resulting in improved text quality and reasoning capabilities. On mathematical reasoning, code generation, and PhD-level QA tasks, MoI consistently improves performance across multiple models including QwQ-32B, Nemotron-Super-49B, Gemma-3-27B, and DAPO-Qwen-32B, with no additional training and negligible computational overhead.

cs.CL↗

arXiv2025

Routing Mamba: Scaling State Space Models with Mixture-of-Experts Projection

Linear State Space Models (SSMs) offer remarkable performance gains in efficient sequence modeling, with constant inference-time computation and memory complexity. Recent advances, such as Mamba, further enhance SSMs with input-dependent gating and hardware-aware implementations, positioning them as strong alternatives to Transformers for long sequence modeling. However, efficiently scaling the expressive power of SSMs, particularly with Mixture of Experts (MoE), remains challenging, as naive integration attempts often falter or degrade performance. In this work, we introduce Routing Mamba (RoM), a novel approach that scales SSM parameters using sparse mixtures of linear projection experts. By sharing routing decisions between projection layers and lightweight sub-modules within Mamba across experts, RoM leverages synergies among linear projection experts for effective and efficient sparse scaling of Mamba layers. At a scale of 1.3B active parameters (10B total) and 16K training sequence length, RoM achieves language modeling performance equivalent to a dense Mamba model requiring over 2.3x more active parameters, and demonstrates consistent perplexity across context lengths. Experimental results further show RoM effectively scales hybrid language models, yielding a 23% FLOPS saving compared to dense Mamba scaling for similar performance.

cs.LG↗

arXiv2025

Optimal Embedding Learning Rate in LLMs: The Effect of Vocabulary Size

Pretraining large language models is a costly process. To make this process more efficient, several methods have been proposed to optimize model architecture/parametrization and hardware use. On the parametrization side, $μP$ (Maximal Update Parametrization) parametrizes model weights and learning rate (LR) in a way that makes hyperparameters (HPs) transferable with width (embedding dimension): HPs can be tuned for a small model and used for larger models without additional tuning. While $μ$P showed impressive results in practice, recent empirical studies have reported conflicting observations when applied to LLMs. One limitation of the theory behind $μ$P is the fact that input dimension (vocabulary size in LLMs) is considered fixed when taking the width to infinity. This is unrealistic since vocabulary size is generally much larger than width in practice. In this work, we provide a theoretical analysis of the effect of vocabulary size on training dynamics, and subsequently show that as vocabulary size increases, the training dynamics \emph{interpolate between the $μ$P regime and another regime that we call Large Vocab (LV) Regime}, where optimal scaling rules are different from those predicted by $μ$P. Our analysis reveals that in the LV regime, the optimal embedding LR to hidden LR ratio should roughly scale as $Θ(\sqrt{width})$, surprisingly close to the empirical findings previously reported in the literature, and different from the $Θ(width)$ ratio predicted by $μ$P. We conduct several experiments to validate our theory, and pretrain a 1B model from scratch to show the benefit of our suggested scaling rule for the embedding LR.

cs.LG↗

arXiv2025

Training Language Models to Generate Quality Code with Program Analysis Feedback

Code generation with large language models (LLMs), often termed vibe coding, is increasingly adopted in production but fails to ensure code quality, particularly in security (e.g., SQL injection vulnerabilities) and maintainability (e.g., missing type annotations). Existing methods, such as supervised fine-tuning and rule-based post-processing, rely on labor-intensive annotations or brittle heuristics, limiting their scalability and effectiveness. We propose REAL, a reinforcement learning framework that incentivizes LLMs to generate production-quality code using program analysis-guided feedback. Specifically, REAL integrates two automated signals: (1) program analysis detecting security or maintainability defects and (2) unit tests ensuring functional correctness. Unlike prior work, our framework is prompt-agnostic and reference-free, enabling scalable supervision without manual intervention. Experiments across multiple datasets and model scales demonstrate that REAL outperforms state-of-the-art methods in simultaneous assessments of functionality and code quality. Our work bridges the gap between rapid prototyping and production-ready code, enabling LLMs to deliver both speed and quality.

cs.CL↗

arXiv2025

An Introduction to Discrete Variational Autoencoders

Variational Autoencoders (VAEs) are well-established as a principled approach to probabilistic unsupervised learning with neural networks. Typically, an encoder network defines the parameters of a Gaussian distributed latent space from which we can sample and pass realizations to a decoder network. This model is trained to reconstruct its inputs and is optimized through the evidence lower bound. In recent years, discrete latent spaces have grown in popularity, suggesting that they may be a natural choice for many data modalities (e.g. text). In this tutorial, we provide a rigorous, yet practical, introduction to discrete variational autoencoders -- specifically, VAEs in which the latent space is made up of latent variables that follow a categorical distribution. We assume only a basic mathematical background with which we carefully derive each step from first principles. From there, we develop a concrete training recipe and provide an example implementation, hosted at https://github.com/alanjeffares/discreteVAE.

cs.LG↗

arXiv2025

Vector-ICL: In-context Learning with Continuous Vector Representations

Large language models (LLMs) have shown remarkable in-context learning (ICL) capabilities on textual data. We explore whether these capabilities can be extended to continuous vectors from diverse domains, obtained from black-box pretrained encoders. By aligning input data with an LLM's embedding space through lightweight projectors, we observe that LLMs can effectively process and learn from these projected vectors, which we term Vector-ICL. In particular, we find that pretraining projectors with general language modeling objectives enables Vector-ICL, while task-specific finetuning further enhances performance. In our experiments across various tasks and modalities, including text reconstruction, numerical function regression, text classification, summarization, molecule captioning, time-series classification, graph classification, and fMRI decoding, Vector-ICL often surpasses both few-shot ICL and domain-specific model or tuning. We further conduct analyses and case studies, indicating the potential of LLMs to process vector representations beyond traditional token-based paradigms.

cs.CL↗

arXiv2024

MamKPD: A Simple Mamba Baseline for Real-Time 2D Keypoint Detection

Real-time 2D keypoint detection plays an essential role in computer vision. Although CNN-based and Transformer-based methods have achieved breakthrough progress, they often fail to deliver superior performance and real-time speed. This paper introduces MamKPD, the first efficient yet effective mamba-based pose estimation framework for 2D keypoint detection. The conventional Mamba module exhibits limited information interaction between patches. To address this, we propose a lightweight contextual modeling module (CMM) that uses depth-wise convolutions to model inter-patch dependencies and linear layers to distill the pose cues within each patch. Subsequently, by combining Mamba for global modeling across all patches, MamKPD effectively extracts instances' pose information. We conduct extensive experiments on human and animal pose estimation datasets to validate the effectiveness of MamKPD. Our MamKPD-L achieves 77.3% AP on the COCO dataset with 1492 FPS on an NVIDIA GTX 4090 GPU. Moreover, MamKPD achieves state-of-the-art results on the MPII dataset and competitive results on the AP-10K dataset while saving 85% of the parameters compared to ViTPose. Our project page is available at https://mamkpd.github.io/.

cs.CV↗

arXiv2024

STOP: Spatiotemporal Orthogonal Propagation for Weight-Threshold-Leakage Synergistic Training of Deep Spiking Neural Networks

The prevailing of artificial intelligence-of-things calls for higher energy-efficient edge computing paradigms, such as neuromorphic agents leveraging brain-inspired spiking neural network (SNN) models based on spatiotemporally sparse binary spikes. However, the lack of efficient and high-accuracy deep SNN learning algorithms prevents them from practical edge deployments at a strictly bounded cost. In this paper, we propose the spatiotemporal orthogonal propagation (STOP) algorithm to tackle this challenge. Our algorithm enables fully synergistic learning of synaptic weights as well as firing thresholds and leakage factors in spiking neurons to improve SNN accuracy, in a unified temporally-forward trace-based framework to mitigate the huge memory requirement for storing neural states across all time-steps in the forward pass. Characteristically, the spatially-backward neuronal errors and temporally-forward traces propagate orthogonally to and independently of each other, substantially reducing computational complexity. Our STOP algorithm obtained high recognition accuracies of 94.84%, 74.92%, 98.26% and 77.10% on the CIFAR-10, CIFAR-100, DVS-Gesture and DVS-CIFAR10 datasets with adequate deep convolutional SNNs of VGG-11 or ResNet-18 structures. Compared with other deep SNN training algorithms, our method is more plausible for edge intelligent scenarios where resources are limited but high-accuracy in-situ learning is desired.

cs.NE↗

arXiv2024

Model Tells You What to Discard: Adaptive KV Cache Compression for LLMs

In this study, we introduce adaptive KV cache compression, a plug-and-play method that reduces the memory footprint of generative inference for Large Language Models (LLMs). Different from the conventional KV cache that retains key and value vectors for all context tokens, we conduct targeted profiling to discern the intrinsic structure of attention modules. Based on the recognized structure, we then construct the KV cache in an adaptive manner: evicting long-range contexts on attention heads emphasizing local contexts, discarding non-special tokens on attention heads centered on special tokens, and only employing the standard KV cache for attention heads that broadly attend to all tokens. Moreover, with the lightweight attention profiling used to guide the construction of the adaptive KV cache, FastGen can be deployed without resource-intensive fine-tuning or re-training. In our experiments across various asks, FastGen demonstrates substantial reduction on GPU memory consumption with negligible generation quality loss. We will release our code and the compatible CUDA kernel for reproducibility.

cs.CL↗

arXiv2024

LoRC: Low-Rank Compression for LLMs KV Cache with a Progressive Compression Strategy

The Key-Value (KV) cache is a crucial component in serving transformer-based autoregressive large language models (LLMs), enabling faster inference by storing previously computed KV vectors. However, its memory consumption scales linearly with sequence length and batch size, posing a significant bottleneck in LLM deployment. Existing approaches to mitigate this issue include: (1) efficient attention variants integrated in upcycling stages, which requires extensive parameter tuning thus unsuitable for pre-trained LLMs; (2) KV cache compression at test time, primarily through token eviction policies, which often overlook inter-layer dependencies and can be task-specific. This paper introduces an orthogonal approach to KV cache compression. We propose a low-rank approximation of KV weight matrices, allowing for plug-in integration with existing transformer-based LLMs without model retraining. To effectively compress KV cache at the weight level, we adjust for layerwise sensitivity and introduce a progressive compression strategy, which is supported by our theoretical analysis on how compression errors accumulate in deep networks. Our method is designed to function without model tuning in upcycling stages or task-specific profiling in test stages. Extensive experiments with LLaMA models ranging from 8B to 70B parameters across various tasks show that our approach significantly reduces the GPU memory footprint while maintaining performance.

cs.LG↗

arXiv2024

Tell Your Model Where to Attend: Post-hoc Attention Steering for LLMs

In human-written articles, we often leverage the subtleties of text style, such as bold and italics, to guide the attention of readers. These textual emphases are vital for the readers to grasp the conveyed information. When interacting with large language models (LLMs), we have a similar need -- steering the model to pay closer attention to user-specified information, e.g., an instruction. Existing methods, however, are constrained to process plain text and do not support such a mechanism. This motivates us to introduce PASTA -- Post-hoc Attention STeering Approach, a method that allows LLMs to read text with user-specified emphasis marks. To this end, PASTA identifies a small subset of attention heads and applies precise attention reweighting on them, directing the model attention to user-specified parts. Like prompting, PASTA is applied at inference time and does not require changing any model parameters. Experiments demonstrate that PASTA can substantially enhance an LLM's ability to follow user instructions or integrate new knowledge from user inputs, leading to a significant performance improvement on a variety of tasks, e.g., an average accuracy improvement of 22% for LLAMA-7B. Our code is publicly available at https://github.com/QingruZhang/PASTA .

cs.CL↗