Searcharxiv⌕ Search

arXiv subjects

Kewei Tu

Publications and source records attributed to Kewei Tu.

At least 19 recordsLinked to original sources

TPHC: Improving Hyper-Connections with Tensor-Parallel Grouped Updates

Hyper-Connections (HC) expand the residual stream of a Transformer into $N$ parallel streams, and manifold-constrained HC (mHC) stabilizes them at scale. However, mHC benefits little from more than four streams, partly because each sublayer writes a single output to all streams, limiting how differently they can be updated. In this work, we propose TPHC, which integrates tensor-parallel (TP) shards with HC streams to promote greater independence across streams. Specifically, we route a separate mixture of the residual streams to each TP shard and keep the individual shard outputs separate during write-back, so each stream can receive a different mixture of those outputs. This design makes the residual-stream updates more expressive without adding parameters or FLOPs to the attention or FFN sublayers. With native support for stream sharding under TP, TPHC requires only 9% more training time than mHC on an 8B model when both use four streams. Increasing the number of streams from four to sixteen raises its training time by only 19%. On a 1.1B model trained on 100B tokens, TPHC reduces validation loss by 0.7-1.0% and raises average zero-shot accuracy across six commonsense reasoning benchmarks by 0.3-0.7 percentage points relative to mHC at matched stream counts $N\in\{4,8,16\}$. Our code is available at https://github.com/wuyou2002/TPHC.

cs.CL↗

A Scaled-Up Empirical Study of Syntactic Language Models

Syntactic language models (SLMs) that generate sentences along with their syntactic tree structures have shown promise in both syntactic generalization and downstream tasks. Prior studies, however, have mostly evaluated smaller models trained on less pretraining data and tested on narrower task suites. In addition, the design space of tree linearization and attention masks remains underexplored. We conduct a large-scale empirical study of 12 controlled configurations---one terminal baseline and 11 SLM variants spanning different linearizations and masks---at scales up to 1B parameters and 100B training tokens, evaluated across 11 downstream tasks and two syntactic benchmarks. The experimental results show that SLMs with standard causal attention match or exceed structural-mask variants across downstream and syntactic benchmarks. They require no architectural modifications and remain compatible with FlashAttention. We also observe that when scaling up to the 1B-parameter scale, the top-performing SLM variant no longer shows a clear advantage over a compute-matched non-structural baseline, suggesting that SLMs may become less competitive with scale. Our code is available at https://github.com/qlwpc/TG-Interpolation.

cs.CL↗

GRKV: Global Regression for Training-Free KV Cache Compression in Long-Context LLMs

Large language models (LLMs) with extended context lengths rely on the key-value (KV) cache to support attention over prior tokens. However, maintaining the KV cache incurs substantial memory overhead, motivating KV-cache compression methods that enforce a fixed budget through eviction and merging. Modern eviction methods increasingly adopt span-based retention because preserving contiguous spans is empirically effective and better preserves semantic coherence. Yet, when combined with post-eviction merging, span-based retention concentrates merges onto a small set of span-boundary carrier tokens, producing a highly imbalanced merge pattern that exacerbates over-merging and increases information loss. To address this imbalance, we propose GRKV (Global Regression for KV Cache), a training-free KV-cache merging method that directly minimizes the discrepancy between compressed-cache and full-cache attention outputs. GRKV uses ridge-regression-based merge steps to distribute information from evicted tokens across retained tokens, while regularizing the updates to prevent over-smoothing. Across the LongBench and RULER long-context benchmarks, GRKV is the only merging method that improves overall performance with minimal overhead. Our code is available at https://github.com/pjunjie/GRKV.

cs.CL↗

S$^4$R: Selective Sampling, Subspaces, and Sparse Reconstruction for Compressed Long-Context KV Caching

The growth of context window lengths in Large Language Models (LLMs) significantly enhances their long-context capabilities but incurs prohibitive memory costs due to the Key-Value (KV) cache. Although low-rank compression of KV cache is a promising remedy, existing methods face a dilemma: offline approaches depend on external calibration data, whereas online approaches incur substantial compute for full-prompt decomposition and reconstruction. In this paper, we propose S$^4$R, which builds low-rank subspaces from selectively sampled tokens and computes attention over a sparsely reconstructed KV representation. S$^4$R uses prompt-aware initialization to build initial key/value bases from a representative prompt subset, trading off calibration-data dependence against prefilling cost. Because fully reconstructing the cache at every decoding step is prohibitively expensive and hurts throughput, we further adopt sparse reconstruction to retain only informative positions during decoding. Extensive experiments on LongBench and RULER with Llama and Qwen model families show that S$^4$R achieves up to 5$\times$ KV compression with near full-cache accuracy, combining the efficiency of fixed compression with the adaptability of prompt-dependent methods.

cs.CL↗

Dependency-Aware Trajectory Refinement for Efficient Multi-Turn Agent Fine-Tuning

Multi-turn agent trajectories often contain redundant rounds (failed tool calls, parallel sub-queries, verification-only steps) that inflate both training and inference cost. We propose viewing each trajectory as a \emph{round-level dependency DAG} that exposes which rounds are globally load-bearing for the final answer, and fine-tune agents on trajectories refined through this DAG. Given an LLM-annotated DAG, these edits are deterministic and interpretable, with optional rephrasing. Models trained on these refined trajectories consistently outperform those trained on the original trajectories at lower inference cost. Specifically, across four multi-modal QA benchmarks, our refinements improve downstream accuracy by up to $1.7$\,pp over vanilla SFT (and $5.7$\,pp over an LLM-deletion baseline) while reducing per-sample inference messages by up to approximately $40\%$ and inference tokens by up to approximately $48\%$, translating to substantial savings in compute and serving cost. Code is available.

cs.CL↗

DA-DLM: Explicitly Modeling Token Dependencies in Diffusion Language Models

Diffusion Language Models (DLMs) generate text by iteratively denoising a masked sequence, independently predicting multiple tokens at each step. This conditional independence discards inter-token dependencies and degrades coherence-an issue that parallels the multi-modality problem in Non-Autoregressive Translation (NAT). Drawing on the Directed Acyclic Transformer (DAT), which tackles this problem in NAT via a Directed Acyclic Graph (DAG), we propose DA-DLM, a model that adapts DAG-based dependency modeling to DLMs' iterative setting through a position-oriented DAG design. The position-oriented DAG binds node groups to fixed output positions so that tokens fixed in earlier steps anchor neighboring predictions via learned transitions, and evolves with denoising to focus on remaining uncertainty as anchors accumulate. On language modeling, open-ended generation, and summarization, DA-DLM consistently outperforms Block Diffusion, especially under fewer denoising steps, and matches autoregressive models while preserving the parallel generation advantage. Our code is publicly available at https://github.com/jipy0222/DA-DLM.

cs.CL↗

Evolution-Aware MSA Reasoning for Subsampling via Factor Graphs

Multiple Sequence Alignments (MSAs) provide protein language models with explicit evolutionary context, but their large depth makes subsampling unavoidable under limited token budgets. Existing strategies, including random selection, identity-based filtering, and diversity-driven sampling, are effective heuristics, yet provide limited control over the evolutionary signals retained in the subset. In this work, we recast MSA subsampling as an explicit optimization problem, where key evolutionary measures, including query identity and diversity, are treated as controllable objectives. Building on this view, we introduce AP-REASONER, an Affinity-Propagation-based factor-graph approach. With evolution-aware unary factors, exemplar-consistency factors, and two control knobs, AP-REASONER performs factor-graph reasoning through message passing to infer a fixed-budget MSA subset. Experiments on long-range contact prediction and conformational ensemble prediction show that AP-REASONER outperforms baseline subsamplers on structure-sensitive downstream tasks and enables controllable recovery of alternative protein conformations. These results highlight the value of modeling MSA subsampling as a controllable optimization problem, where factor-graph reasoning offers an effective alternative to heuristic selection.

cs.LG↗

Hierarchical Sparse Attention Done Right: Toward Infinite Context Modeling

Scaling modern large language models (LLMs) to long contexts is limited by the quadratic computation cost, and poor length extrapolation of dense attention. Chunk-wise sparse attention offers a promising alternative, but all existing methods fall short of full attention because of their inaccurate chunk selection. We propose Hierarchical Landmark Sparse (HiLS) Attention, a chunk-wise sparse attention mechanism that learns chunk selection end-to-end under the language-modeling (LM) loss. HiLS factorizes attention hierarchically: each query performs attention independently with each retrieved chunk to extract chunk-specific information, and the resulting outputs are fused according to chunk retrieval scores. By incorporating retrieval scores into the forward attention computation, HiLS optimizes them directly with the LM loss, enabling end-to-end retrieval learning and native sparse training. Experimental results show that HiLS-Attention achieves performance comparable to, and in some cases better than, full attention at in-domain context lengths. Meanwhile, HiLS-Attention extrapolates more than $64\times$ the training context length with 90% retrieval accuracy, far beyond full attention. Moreover, existing full-attention models can be converted to HiLS-Attention with lightweight continued pretraining, preserving in-domain performance while acquiring ultra-long-context extrapolation. Together with its sparse KV access and computation, HiLS-Attention breaks the usual efficiency-performance trade-off, enabling long-context LLMs that are both more efficient and more effective on general long-context tasks than their full-attention counterparts.

cs.CL↗

Does Traversal Order Matter? A Systematic Study of Tree Traversal Methods in Transformer Grammars

Transformer Grammars (TGs) enhance language modeling by incorporating syntactic tree structures. Despite the potentially significant impact on model performance of how syntactic trees are linearized in TGs, existing studies rely solely on Depth-First Traversal (DFT) for linearization. In this paper, we expand the traversal design space by exploring Breadth-First Traversal (BFT) and a novel hybrid traversal strategy, Production-Rule Traversal (PRT), which combines the structural lookahead of BFT with the early lexical generation of DFT. We integrate these traversal methods with varying tree configurations and masking strategies, and empirically evaluate their performance on language modeling, syntactic generalization and summarization. We reveal the inherent trade-offs between nested composition and global lookahead, providing actionable recommendations for designing task-aware Transformer Grammars.

cs.CL↗

GiLT: Augmenting Transformer Language Models with Dependency Graphs

Augmenting Transformers with linguistic structures effectively enhances the syntactic generalization performance of language models. Previous work in this direction focuses on syntactic tree structures of languages, in particular constituency tree structures. We propose Graph-Infused Layers Transformer Language Model (GiLT) which leverages dependency graphs for augmenting Transformer language models. Unlike most previous work, GiLT does not insert extra structural tokens in language modeling; instead, it injects structural information into language modeling by modulating attention weights in the Transformer with features extracted from the dependency graph that is incrementally constructed along with token prediction. In our experiments, GiLT with semantic dependency graphs achieves better syntactic generalization while maintaining competitive perplexity in comparison with Transformer language model baselines. In addition, GiLT can be finetuned from a pretrained language model to achieve improved downstream task performance. Our code is released at https://github.com/cookie-pie-oops/GiLT-LM.

cs.CL↗

Exploring the Potential of Probabilistic Transformer for Time Series Modeling: A Report on the ST-PT Framework

The Probabilistic Transformer (PT) establishes that the Transformer's self-attention plus its feed-forward block is mathematically equivalent to Mean-Field Variational Inference (MFVI) on a Conditional Random Field (CRF). Under this equivalence the Transformer ceases to be a black-box neural network and becomes a programmable factor graph: graph topology, factor potentials, and the message-passing schedule are all explicit and inspectable primitives that can be engineered. PT was originally developed for natural language and in this report we investigate its potential for time series. We first lift PT into the Spatial-Temporal Probabilistic Transformer (ST-PT) to repair PT's missing channel axis and weak per-step semantics, and adopt ST-PT as a shared cornerstone backbone. We then identify three distinct properties that PT/ST-PT offers as a factor-graph model and derive three Research Questions, one per property, that probe how each property can be exploited in time series: RQ1. The graph topology and potentials are direct programmable primitives. Can this be used to inject symbolic time-series priors into ST-PT through structural graph modifications, especially under data scarcity and noise? RQ2. The CRF's factor matrices are the operator's potentials. Can an external condition program these factor matrices on a per-sample basis, so that conditional generation becomes structural rather than feature-level modulation of a fixed one? RQ3. Each MFVI iteration is a Bayesian posterior update on the factor graph. Can this turn the latent transition of latent-space AutoRegressive (AR) forecasting from an opaque MLP into a principled posterior update, and can a CRF teacher distill its latents into the AR student to counter cumulative error? We give one empirical study per question. Together, these three studies position ST-PT as a programmable framework for time-series modeling.

cs.LG↗

Scaling Probabilistic Transformer via Efficient Cross-Scale Hyperparameter Transfer

Probabilistic Transformer (PT), a white-box probabilistic model for contextual word representation, has demonstrated substantial similarity to standard Transformers in both computational structure and downstream task performance on small models and small to medium sized datasets. However, PT is less robust to hyperparameter choices than standard Transformers, making it harder to scale efficiently. In this work, we follow Maximal Update Parametrization (muP) to rescale PT's parameters, so that hyperparameters optimized on small models can be transferred to larger models without additional tuning. With this approach, we successfully scale PT to models with up to 0.4B parameters. Experiments show that PT consistently outperforms standard transformer under the same parameter budget on Masked Language Modeling (MLM) tasks. We hope this work will contribute to the practical deployment of probabilistic models at substantially larger scales in the future.

cs.CL↗

YOCO++: Enhancing YOCO with KV Residual Connections for Efficient LLM Inference

Cross-layer key-value (KV) compression has been found to be effective in efficient inference of large language models (LLMs). Although they reduce the memory consumption of the KV cache, such methods usually introduce non-negligible performance degradation. In this work, we aim to enhance the performance of YOCO, a cross-layer KV compression method that shares the KVs of the middle layer with the top-half layers. We propose YOCO++, an enhanced YOCO that incorporates a weighted residual connection between the KVs of each bottom-half layer and the bottom layer. Compared to YOCO, YOCO++ increases model capacity while maintaining the same training and inference efficiency. Our experiments show that YOCO++ achieves state-of-the-art performance among the cross-layer KV compression methods at a 50% KV cache compression rate, outperforming the standard Transformer.

cs.CL↗

Parallel Continuous Chain-of-Thought with Jacobi Iteration

Continuous chain-of-thought has been shown to be effective in saving reasoning tokens for large language models. By reasoning with continuous latent thought tokens, continuous CoT is able to perform implicit reasoning in a compact manner. However, the sequential dependencies between latent thought tokens spoil parallel training, leading to long training time. In this paper, we propose Parallel Continuous Chain-of-Thought (PCCoT), which performs Jacobi iteration on the latent thought tokens, updating them iteratively in parallel instead of sequentially and thus improving both training and inference efficiency of continuous CoT. Experiments demonstrate that by choosing the proper number of iterations, we are able to achieve comparable or even better performance while saving nearly 50% of the training and inference time. Moreover, PCCoT shows better stability and robustness in the training process. Our code is available at https://github.com/whyNLP/PCCoT.

cs.CL↗

Repurposing Synthetic Data for Fine-grained Search Agent Supervision

LLM-based search agents are increasingly trained on entity-centric synthetic data to solve complex, knowledge-intensive tasks. However, prevailing training methods like Group Relative Policy Optimization (GRPO) discard this rich entity information, relying instead on sparse, outcome-based rewards. This critical limitation renders them unable to distinguish informative "near-miss" samples-those with substantially correct reasoning but a flawed final answer-from complete failures, thus discarding valuable learning signals. We address this by leveraging the very entities discarded during training. Our empirical analysis reveals a strong positive correlation between the number of ground-truth entities identified during an agent's reasoning process and final answer accuracy. Building on this insight, we introduce Entity-aware Group Relative Policy Optimization (E-GRPO), a novel framework that formulates a dense entity-aware reward function. E-GRPO assigns partial rewards to incorrect samples proportional to their entity match rate, enabling the model to effectively learn from these "near-misses". Experiments on diverse question-answering (QA) and deep research benchmarks show that E-GRPO consistently and significantly outperforms the GRPO baseline. Furthermore, our analysis reveals that E-GRPO not only achieves superior accuracy but also induces more efficient reasoning policies that require fewer tool calls, demonstrating a more effective and sample-efficient approach to aligning search agents.

cs.CL↗

Efficient Multimodal Planning Agent for Visual Question-Answering

Visual Question-Answering (VQA) is a challenging multimodal task that requires integrating visual and textual information to generate accurate responses. While multimodal Retrieval-Augmented Generation (mRAG) has shown promise in enhancing VQA systems by providing more evidence on both image and text sides, the default procedure that addresses VQA queries, especially the knowledge-intensive ones, often relies on multi-stage pipelines of mRAG with inherent dependencies. To mitigate the inefficiency limitations while maintaining VQA task performance, this paper proposes a method that trains a multimodal planning agent, dynamically decomposing the mRAG pipeline to solve the VQA task. Our method optimizes the trade-off between efficiency and effectiveness by training the agent to intelligently determine the necessity of each mRAG step. In our experiments, the agent can help reduce redundant computations, cutting search time by over 60\% compared to existing methods and decreasing costly tool calls. Meanwhile, experiments demonstrate that our method outperforms all baselines, including a Deep Research agent and a carefully designed prompt-based method, on average over six various datasets. Code will be released.

cs.CL↗

Flash Multi-Head Feed-Forward Network

We explore Multi-Head FFN (MH-FFN) as a replacement of FFN in the Transformer architecture, motivated by the structural similarity between single-head attention and FFN. While multi-head mechanisms enhance expressivity in attention, naively applying them to FFNs faces two challenges: memory consumption scaling with the head count, and an imbalanced ratio between the growing intermediate size and the fixed head dimension as models scale, which degrades scalability and expressive power. To address these challenges, we propose Flash Multi-Head FFN (FlashMHF), with two key innovations: an I/O-aware fused kernel computing outputs online in SRAM akin to FlashAttention, and a design using dynamically weighted parallel sub-networks to maintain a balanced ratio between intermediate and head dimensions. Validated on models from 128M to 1.3B parameters, FlashMHF consistently improves perplexity and downstream task accuracy over SwiGLU FFNs, while reducing peak memory usage by 3-5x and accelerating inference by up to 1.08x. Our work establishes the multi-head design as a superior architectural principle for FFNs, presenting FlashMHF as a powerful, efficient, and scalable alternative to FFNs in Transformers.

cs.LG↗

Hardware-aligned Hierarchical Sparse Attention for Efficient Long-term Memory Access

A key advantage of Recurrent Neural Networks (RNNs) over Transformers is their linear computational and space complexity enables faster training and inference for long sequences. However, RNNs are fundamentally unable to randomly access historical context, and simply integrating attention mechanisms may undermine their efficiency advantages. To overcome this limitation, we propose Hierarchical Sparse Attention (HSA), a novel attention mechanism that enhances RNNs with long-range random access flexibility while preserving their merits in efficiency and length generalization. HSA divides inputs into chunks, selects the top-$k$ chunks and hierarchically aggregates information. The core innovation lies in learning token-to-chunk relevance based on fine-grained token-level information inside each chunk. This approach enhances the precision of chunk selection across both in-domain and out-of-domain context lengths. To make HSA efficient, we further introduce a hardware-aligned kernel design. By combining HSA with Mamba, we introduce RAMba, which achieves perfect accuracy in passkey retrieval across 64 million contexts despite pre-training on only 4K-length contexts, and significant improvements on various downstream tasks, with nearly constant memory footprint. These results show RAMba's huge potential in long-context modeling.

cs.CL↗