arXiv subjects
Publications and source records attributed to Jason Li.
We consider the maximum cardinality matching problem in general graphs, and give an $n^{2+o(1)}$-time algorithm for the problem. Our work builds on recent work on combinatorial algorithms for maximum flow and extends the techniques to skew-symmetric flow problems which capture general matchings as a special case.
Large language models are typically adapted to a task either by placing knowledge in the input context, which keeps it external and re-encodes it on every call, or by fine-tuning, which internalizes it but demands a costly round of optimization per task. We aim to internalize the knowledge carried by a context directly into model parameters, without per-task optimization and without keeping the context in the input at inference time. To this end, we propose CLIK, a context-conditioned parameter generator that maps a given context to LoRA parameters for internalizing knowledge in a single forward pass through a Transformer encoder with cross-attention, requiring no test-time gradient updates. Once the generated LoRA is inserted into a frozen base model, the model answers downstream queries from the query alone, having internalized the context into its weights rather than re-reading it. Because the adapter depends only on the context, it is generated once and reused for every query on the same document, so the cost of reading that document is amortized across queries. On document question answering, CLIK substantially improves over the no-context base model and outperforms existing hypernetwork-based generators, while remaining behind a base model that reads the context directly, and it carries over to multi-turn dialogue without any multi-turn-specific training. Analysis further shows that the context-dependent part of the generated adapter concentrates in a compact and specific substructure.
Fine-grained multimodal intelligence requires models to both understand a designated region and localize it from language. Existing methods learn these directions with separate objectives and task-specific annotations. We present Generalized CycleGRPO, the first unified reinforcement learning framework coupling region understanding and localization across spatial, temporal, and spatio-temporal domains. Given a region, a multimodal large language model generates candidate descriptions and then localizes each description to reconstruct the input region. Reconstruction overlap provides a verifiable reward for captioning, while individual localization rewards improve the inverse mapping. A common region-space formulation extends this cycle from image masks and boxes to video intervals and timestamped spatial evidence, without reference captions or an external language-model judge during reinforcement learning. In videos, the cycle encourages descriptions that distinguish a target event from similar recurrences. We introduce DOVE (Distinctive Occurrence-level Video-caption Evaluation), a human-verified contrastive benchmark for this temporal distinctiveness. CycleGRPO improves SAMTok's overall gIoU from 57.5% to 68.2% on GroundingSuite and average gIoU from 71.2% to 78.7% on GRES. For temporal tasks, it achieves 51.9% tIoU on Charades and 33.3% Temporal Evidence Coverage on DOVE, compared with 45.6% and 14.8% for the Qwen3-VL initialization. These results establish region-text cycle consistency as a general self-evaluation principle for multimodal understanding and localization. Models, code, and DOVE are available at https://devinxzhang.github.io/Generalized-CycleGRPO-Page/.
Vision-language-action (VLA), video-action (VA), and world action models (WAMs) form the backbone of modern physical AI systems, yet their real-time single-batch inference remains a defining bottleneck. Latency is dominated not by individual kernel computation, but by system-level overhead: thousands of short kernel launches, intermediate data movement across operator boundaries, and redundant recomputation across consecutive decisions. General-purpose graph compilers cannot address these costs, as they optimize individual calls in isolation and lack visibility into cross-call state semantics and global pipeline structure. We present FLASHRT-PAI, a low-latency inference engine built around model-hardware pipeline co-design rather than isolated kernel optimization. FLASHRT-PAI combines three core mechanisms: deep kernel fusion with launch overlap eliminates inter-operator boundaries and launch gaps; an execution contract formalizes the pipeline's stage DAG, update rates, and named buffers to enable cross-decision state reuse and selective refresh; and execution-state capsules enable fast snapshot, restore, and fork of persistent pipeline state. The system also supports drop-in acceleration of unmodified PyTorch models via parity- and latency-gated kernel replacement. Evaluation covers six policies on RTX 5090 and four on Jetson AGX Thor. At matched 16-bit precision, FLASHRT-PAI is 1.3-3.7x faster than official compiled paths. Combining pipeline co-design with reduced precision delivers 3.38-12.66x speedup over official eager implementations. In LIBERO-10, four-bit pi0.5 achieves 94.04% success on Thor and 93.32% on RTX 5090, compared with 92.4% for the measured official FP16 reference. FLASHRT-PAI is open source, distributed via Hugging Face Kernel Hub, and its opt-in pi0.5 kernels have been adopted into the LeRobot framework.
Generated videos often lose geometric consistency under camera motion: structures warp under viewpoint change, parallax becomes inconsistent, and smooth camera movement may collapse into shot-like transitions. We argue that this failure is largely a data problem. In web-scale video corpora, most visible motion comes from dynamic objects captured by static or weakly moving cameras, leaving only sparse supervision for viewpoint-induced appearance changes. To address this, we propose Video Geometry Corrector (VGC), a post-hoc video-to-video model trained on synthetic paired supervision. Starting from real videos with diverse camera motion and pose annotations, we construct VG-Pairs with a synthetic paired-supervision data generation pipeline. Each video clip is treated as a geometry-consistent target and paired with a distorted counterpart synthesized by a pretrained image-to-video inpainting model using trajectory-adaptive keyframes. VGC learns to map the distorted input to the geometry-consistent target without modifying the source generator, and uses optical flow as an auxiliary cue to help disentangle camera motion from object motion. Experiments on pose-conditioned and prompt-driven benchmarks show that VGC improves geometric consistency across multiple source generators, and scaling to a larger backbone further extends the framework toward more general video refinement.
Progress toward general intelligence in the physical world requires scaling robot learning across diverse environments and interactions. However, existing robot policies use shared parameters that remain fixed across interaction contexts. When different operations favor competing parameter adjustments, fitting their mixture can leave individual operations insufficiently adapted. Our atomic-task diagnostics reveal competing update directions in both vision-language and action parameters, with source-improving updates increasing loss on other tasks. This motivates MorphAct, a contextual parameterization framework that dynamically adjusts robot policy parameters as interaction contexts change. Two coupled generators produce context-dependent low-rank updates for the vision-language and action modules. Their coupling follows the policy's computational dependency: vision-language updates shape the representations that condition action-side parameter generation. Both generators are trained jointly through the native action objective, without operation labels or expert-adapter targets. MorphAct improves performance across three VLA backbones on LIBERO and Meta-World, with gains of up to 16.30%. In real-world evaluation, it achieves gains of up to 14% over $π_{0.5}$. Ablations support the benefits of contextual parameter generation and representation coupling.
Reliable video reasoning requires models to produce correct answers and expose supporting visual evidence. Open-o3-Video enables verifiable reasoning by embedding timestamps and object-level bounding boxes into reasoning traces, but is restricted to frames from an initial sparse sampler. We introduce Open-o3-Video-v2, which formulates video reasoning as a unified agentic evidencegrounded process. The model identifies missing information, acquires additional observations, grounds the moments and objects supporting its reasoning, and writes this evidence into the trajectory before continuing or answering. To support this behavior, we construct STGR-v2-CoT135K, including 52K multi-turn agentic grounded reasoning trajectories, and STGR-v2-RL-13K for reinforcement learning. We further propose Clue-Seek GRPO (CS-GRPO), which tracks how tool calls change accumulated evidence, rewards clue coverage and step-wise gains, and penalizes excessive, repetitive, or uninformative retrieval. On V-STaR, Open-o3-Video-v2-4B achieves 40.4% mAM and 56.3% mLGM, outperforming its Qwen3-VL-4B backbone by 8.5% and 12.6%, respectively. It also delivers consistent gains across general video understanding, long-video reasoning, and temporal grounding benchmarks. These results show that active evidence acquisition and verifiable object-level grounding are complementary for reliable video reasoning. The code and models are released at https://msalab-pku.github.io/projects/Open-o3-Video-v2.
We give a randomized reduction from global minimum cut in a directed graph with $n$ vertices and $m$ weighted edges to maximum-flow computations on graphs of total size $\widetilde{O}(m)$. For polynomially bounded integral weights, this yields the first almost-linear $m^{1+o(1)}$-time algorithm [vdBCK+23], improving the previous $m^{1+o(1)}\min\{\sqrt{n},n/m^{1/3}\}$ bound [CLN+22]. Previous work reduced finding the minimum cut to constructing a 1-respecting arborescence, a directed spanning tree with exactly one edge crossing a minimum cut [CLN+22]. They then construct such a tree by sampling from a $(1+ε)$-approximate arborescence packing, which currently requires superlinear time. Our algorithm sidesteps constructing the packing entirely and relies on a novel iterative arborescence sampling procedure. We show that, after $O(\log n)$ rounds of sampling, our final arborescence 1-respects the minimum cut with constant probability.
We introduce NemotronLabs VoiceChat, an open full-duplex speech-to-speech model with native tool-calling capabilities. NemotronLabs VoiceChat combines a streaming speech encoder and decoder-only language model with parallel specialized output streams for agent text and structured function calls, an auxiliary RNN-T branch for incremental user transcription, and a streaming TTS decoder. This design enables the model to listen, transcribe, reason, invoke tools, and speak within a unified streaming architecture while preserving the temporal behavior required for natural conversation. On Full-Duplex-Bench 1.0, NemotronLabs VoiceChat achieves the lowest pause-handling takeover rates among evaluated open-weight systems, 100\% takeover following user interruptions, and a 4.33/5 post-interruption response-quality score. On Full-Duplex-Bench 1.5, it resumes its response after user backchannels in 93\% of cases. NemotronLabs VoiceChat obtains a 55.1 normalized average on VoiceBench and, on Full-Duplex-Bench 3.0 (FDB 3.0), achieves 82.5\% tool-selection F1, while argument accuracy and end-to-end tool execution remain areas for improvement. These results demonstrate that full-duplex interaction, speech recognition and generation, general language capabilities, and external tool use can be integrated in a single open speech-to-speech model without sacrificing real-time conversational behavior.
Advantage-guided reinforcement learning provides a practical way to post-train vision-language-action (VLA) policies using limited robot data. However, its performance depends on several coupled choices, including how critic-derived advantages are constructed, calibrated, and used for policy training. Existing recipes often combine these choices into a single end-to-end procedure, making their individual effects difficult to identify. In this work, we dissect advantage-guided VLA post-training through a controlled empirical study that separates these design choices while accounting for their distinct estimands. We develop stage-specific offline evaluation methods to screen alternative choices efficiently, without requiring extensive real-robot policy evaluations for every possible combination. The staged evaluation identifies a modular recipe that combines temporal-difference advantage construction, group-wise calibration, and continuous advantage weighting. Across four real-world bimanual tasks, the resulting recipe improves mean task progress and success over the SFT initialization by 0.42 and 0.63, respectively. Moreover, the proposed evaluation diagnostics show an overall alignment with downstream real-world performance, supporting their use for interpreting empirical outcomes and selecting advantage-guided post-training designs in practice.
The planar separator theorem by Lipton and Tarjan [FOCS '77, SIAM Journal on Applied Mathematics '79] states that any planar graph with $n$ vertices has a balanced separator of size $O(\sqrt{n})$ that can be found in linear time. This landmark result kicked off decades of research on designing linear or nearly linear-time algorithms on planar graphs. In an attempt to generalize Lipton-Tarjan's theorem to nonplanar graphs, Alon, Seymour, and Thomas [STOC '90, Journal of the AMS '90] showed that any minor-free graph admits a balanced separator of size $O(\sqrt{n})$ that can be found in $O(n^{3/2})$ time. The superlinear running time in their separator theorem is a key bottleneck for generalizing algorithmic results from planar to minor-free graphs. Despite extensive research for more than two decades, finding a balanced separator of size $O(\sqrt{n})$ in (linear) $O(n)$ time for minor-free graphs remains a major open problem. Known algorithms either give a separator of size much larger than $O(\sqrt{n})$ or have superlinear running time, or both. In this paper, we answer the open problem affirmatively. Our algorithm is very simple: it runs a vertex-weighted variant of breadth-first search (BFS) a constant number of times on the input graph. Our key technical contribution is a weighting scheme on the vertices to guide the search for a balanced separator, offering a new connection between the size of a balanced separator and the existence of a clique-minor model. We believe that our weighting scheme may be of independent interest.
We show that the balancing weights technique of Li (2026) actually produces an approximate *pseudo-circulation* of a directed, capacitated graph in $m^{1+o(1)}$ time. Together with standard flow techniques, we obtain an $m^{1+o(1)}$ time maximum flow algorithm that avoids the interior-point method framework of recent almost-linear time algorithms (Chen et al. FOCS 2022, van den Brand et al. FOCS 2024).
The minimum $k$-cut problem asks for the fewest edges whose removal leaves an input graph with at least $k$ connected components. Previously, the best algorithm for simple graphs ran in $O_k(n^{(1-\varepsilon)k+O(1)})$ time~\cite{HL22}, showing that the \(n^k\) barrier can be broken up to a polynomial overhead. We give the first $\widetilde O_k(n^{ck})$-time algorithm for Minimum $k$-Cut on simple graphs for an absolute constant $c<1$. More precisely, the running times are $\widetilde O(n^2)$ for $k=3$, $\widetilde O(n^{55/19})$ for $k=4$, and $\widetilde O(n^{4.112007})$ for $k=5$; for every $k\ge6$, the running time is \[ k^{O(k^2)}n^{1+(6k-6)\frac{k-1.749614}{7k-10}}(\log n)^{O(k^2)}, \] whose exponent is $\frac67k-0.132\ldots+O(1/k)$. The algorithm combines three ingredients. First, for weighted Minimum $k$-Cut we give a randomized \[ k^{O(k^2)}n^{k-2}(m+n)\log^3(n) \] -time algorithm: it perturbs the edge weights so that any minimum $k$-cut has a side with boundary strictly smaller than average. These then cut few edges of some tree in a logarithmic-size sample from a tree packing with high probability. After we enumerate them, we recursively compute $(k-1)$-cuts to complete them to the $k$-cuts of which they were a part. A variant of the perturbation and processing the entire packing support give a deterministic $k^{O(k^2)}n^{k+O(1)}$-time variant. Second, for cut size $s$, we give an improved FPT algorithm using a near-linear-time construction of an $(O(s\log^2 n\log\log n),s)$ edge-unbreakable tree decomposition with $O(s\log^2 n\log\log n)$ adhesion; this also gives a near-linear-time approximation algorithm for Minimum $k$-Cut. Third, we refine the border/island framework of~\cite{HL22}, using rectangular matrix multiplication to recover singleton islands and balancing it against the improved FPT algorithm.
A spectral sparsifier of a weighted graph is a reweighted subgraph whose Laplacian quadratic form approximates that of the original graph. Let $G$ be a positively weighted $n$-vertex, $m$-edge multigraph, let $0<\varepsilon\le1/2$. Assuming $m,\varepsilon^{-1}\le n^{O(1)}$ and the ratio of maximum to minimum weight is polynomially bounded, we deterministically construct a $(1\pm\varepsilon)$-spectral sparsifier with \[ O\!\left(n\varepsilon^{-2}\log^{24+o(1)}n\right) \] edges in \[ m^{1+o(1)}+O\!\left(n^2\varepsilon^{-9/2}\log^{113/2+o(1)}n\right) \] time. The construction has two main ingredients. First, we sparsify an approximately regular expander by partitioning its edges into few matchings and viewing their normalized Laplacians as an isotropic family of positive semidefinite matrices. Rather than sample from this family and apply matrix Chernoff, we select matchings deterministically using a pessimistic estimator. We evaluate the resulting conditional-expectation scores in two ways to produce two algorithms: using dense matrix multiplication and sparsely using polynomial approximations to the inverse square root and matrix exponential. Deterministic expander decomposition, along with replacing vertices by fixed expander graphs to achieve approximate regularity, extends these algorithms to general graphs. Second, a recursive blocking scheme applies the dense algorithm to smaller subgraphs and the sparse algorithm to their union, balancing their costs. Reusing the resulting algorithm as the dense algorithm gives $α_{r+1}=3-1/(α_r-1)$, starting from $α_0=ω$. After $O(\log n)$ levels, the exponent is $2+O(1/\log n)$, yielding $m^{1+o(1)}+\widetilde O_{\varepsilon}(n^2)$ time.
Spoken dialogue is a natural form of human--computer interaction, yet most speech language models remain limited to turn-based operation and lack real-time adaptability, such as user barge-in. Recent duplex speech-to-speech and speech-to-text models reduce latency by replacing multi-stage pipelines, but often compromise speech quality because accurate ASR, interruption handling, and high-fidelity synthesis must be optimized jointly. We propose VoiceChat-TTS, a low-latency, continuous, and streamable text-to-speech model for interactive agents. VoiceChat-TTS is driven directly by LLM text-token streams, supports explicit interruption via control tokens, and produces silence when no textual input is available. The model enables always-on, responsive speech generation while preserving modularity and high speech quality, and it supports mid-utterance interruptions without resetting the KV cache.
We present Xiaomi-Robotics-1, a foundational vision-language-action (VLA) model capable of (1) following diverse language instructions to perform a wide range of mobile manipulation tasks in unseen environments out-of-the-box, and (2) efficiently adapting to novel downstream tasks with minimal fine-tuning data. We propose a two-stage training recipe consisting of pre-training and post-training. During pre-training, we imbue the model with broad and generalizable action-generation capabilities by training on over 100k hours of real-world manipulation trajectories collected via UMI devices. Crucially, we develop a scalable auto-labeling pipeline that annotates trajectory clips with natural languages describing scene state transitions, providing rich and precise conditioning for action learning. During post-training, we aim to align these capabilities with robot embodiments and imperative instructions that humans naturally use to prompt robots. Extensive experiments demonstrate strong scaling behavior. Xiaomi-Robotics-1 consistently improves with increased data scales and model sizes during pre-training. This scaling behavior directly transfers to post-training, where a stronger pre-training model yields better out-of-the-box real-robot performance in unseen environments. Furthermore, Xiaomi-Robotics-1 serves as a strong robot foundation policy that can be efficiently fine-tuned on complex, dexterous tasks with high data efficiency. Across multiple simulation benchmarks, Xiaomi-Robotics-1 outperforms state-of-the-art methods. Notably, it establishes a new state-of-the-art with a 57.4% success rate on RoboCasa365, surpassing the previous best of 46.6%. Furthermore, it achieves an average score of 20.07 on RoboDojo, significantly outperforming the prior state-of-the-art (13.07). Code and model checkpoints will be released. Project page: https://robotics.xiaomi.com/xiaomi-robotics-1.html
We revisit the single-source shortest paths (SSSP) problem on directed graphs with nonnegative real weights and give a deterministic parallel algorithm with $O(n^{1+o(1)}t^2 + m^{1+o(1)})$ work and $\tilde{O}(n/t)$ span, for any $t \in [1, n]$. This matches (up to subpolynomial factors) the tradeoff due to [Shi and Spencer '99] for undirected graphs with nonnegative real weights.
We consider the single-source shortest paths problem on a directed graph with real-valued (possibly negative) edge weights and solve this problem in $m^{1+o(1)}$ time.