arXiv subjects

Kai Yang

Publications and source records attributed to Kai Yang.

At least 19 recordsLinked to original sources

arXiv2026

Turing LLM Post-Training: Calibrated Distillation and Efficient Dynamic-MoE Execution

We study post-training for dynamically routed mixture-of-experts (MoE) models, focusing on integrating independently optimized specialists while keeping consolidation efficient. Our approach combines calibrated multi-teacher on-policy distillation (C-MOPD), which matches each selected teacher on a local Top-$K$ support with teacher-wise loss normalization and clipping, with an execution path that uses post-capacity bounded-width dispatch, final-route replay, and asynchronous rollout--training overlap. On Turing-20B-A2B, the selected model achieves a four-domain normalized integration score of 0.9988 and retains at least 90.16\% of the measured specialist gain in every domain. At matched $K=32$, the full calibration objective scores 0.9962, compared with 0.9724 without calibration and 0.9834 without clipping. At a matched 8K response cap, the model exceeds both shared SFT and Qwen3-8B on all four reasoning-enabled mathematics and code benchmarks (unweighted mean: 85.92, versus 80.33 for Qwen3-8B and 74.21 for SFT), and improves over SFT on ten of twelve broader evaluation metrics. In a PPU-based profile, the combined execution configuration reduces consolidation step time from 398.7 to 139.3 seconds, an observed 2.86$\times$ speedup over serial execution. These results demonstrate strong specialist integration alongside more efficient consolidation, while reflecting the complete training and execution framework rather than any single component.

cs.LG↗

arXiv2026

ThinkJev: One Thought, Every Budget for Calibrated Typed Decisions

Many uses of language models end in a typed decision: a choice among a closed set of options, whose probabilities code can act on. Reading those probabilities from one forward pass is fast but often wrong on decisions that take several steps, such as arithmetic, counting, and date comparison. Letting the model think first fixes most of these errors, but costs seconds per decision instead of milliseconds. To pay for thinking only where it helps, ThinkJev builds a fast and a slow system from one open hybrid-thinking model that can answer directly or after a thought. System 1 reads the model's probabilities for the option letters when it answers directly; System 2 reads the same letters after the model has thought on the same cached prompt, so escalating repeats no work and both systems give probabilities over the same options. Because the letters can be read after any number of thought tokens, one sampled thought also gives the answer under every thinking budget. Without training, ThinkJev lets System 1 answer alone only where a statistical test on labeled dev data guarantees that its accepted answers stay under a target error rate. At a 5% target, Qwen3.5-27B answers 24.5% of 8359 test decisions this way with a realized error of 3.9% (Qwen3.5-4B and 9B: 6.5% and 0.0%). For the rest, stopping each thought as soon as its answer is confident reaches 81.6% accuracy, against 82.7% for always thinking, with 72% of the thought tokens. Choosing when to think from System 1's confidence alone saves less, because the decisions System 1 is least sure of are often not the ones thinking fixes. On the public items of JevBench, an independent benchmark of typed-decision services, Qwen3.5-27B thinks on 29.4% and is not distinguishable in accuracy from the two thinking APIs listed (exploratory). Code and data generators are available at https://github.com/ding-xiaolu/ThinkJev.

cs.AI↗

arXiv2026

TPO: Tri-level Distributionally Robust Learning for OOD Direct Preference Optimization

Direct Preference Optimization (DPO) minimizes empirical risk over observed preference comparisons, implicitly assuming that the training preference distribution matches the deployment preference distribution. In practice, this i.i.d. assumption often fails, as preference distribution shift exhibits a natural hierarchical structure across three levels: group-level shift, prompt-level shift, and preference-level shift. Such multi-level distributional mismatch poses a critical challenge for safe LLM deployment, yet existing robust DPO methods typically either model uncertainty over the joint preference distribution or focus on individual sources of shift, leaving the hierarchical structure of OOD preference shifts underexplored. We propose TPO, a tri-level distributionally robust framework for OOD preference alignment, which jointly models all three levels of distributional uncertainty through a hierarchically structured rectangular ambiguity set. By instantiating it with Wasserstein distance, we derive a tractable and interpretable objective that combines a closed-form preference-level margin penalty, prompt-level adversarial embedding perturbation, and worst-case group reweighting, each governed by an independent radius. We establish finite-sample convergence guarantees and empirically validate the effectiveness of TPO over existing robust DPO methods.

cs.LG↗

arXiv2026

Widely Linear-Enhanced IQ Imbalance Interference Cancellation for AFDM-FTN Systems

Internet of Things (IoT) networks require reliable and spectrally efficient wireless links to support massive device connectivity and high mobility applications. To meet these requirements, this paper investigates a widely linear (WL)-empowered receiver design for affine frequency division multiplexing (AFDM) with faster-than-Nyquist (FTN) signaling under practical receiver in phase and quadrature imbalance (IQI). The considered AFDM-FTN system couples the direct and image branches with FTN induced intersymbol interference (FTN-ISI) and delay-Doppler spreading in the affine domain. To address this coupled interference structure, a unified conjugate augmented input output model is established for IQI impaired AFDM-FTN systems. Based on this model, a WL minimum mean square error receiver with parallel interference cancellation (WL-MMSE-PIC) is developed, where WL-MMSE preprocessing is followed by the proposed hard decision PIC (HD-PIC) and damped soft decision PIC (SD-PIC). The proposed receiver is further analyzed in terms of the local stability of HD-PIC, a genie aided performance reference under ideal residual interference cancellation, and the impact of the AFDM chirp parameter on the post equalization interference distribution. Simulation results show that the proposed WL-MMSE-HD-PIC and WL-MMSE-SD-PIC receivers achieve significant error rate performance gains over the conventional linear-MMSE and WL-MMSE baselines. The results further show that FTN signaling provides a nominal spectral-efficiency gain, while residual-coupling-based chirp-parameter selection reduces the post-equalization residual coupling and improves detection reliability for high-mobility IoT links.

eess.SP↗

arXiv2026

Thinking-Free Policy Initialization Makes Distilled Reasoning Models More Effective and Efficient Reasoners

Reinforcement Learning with Verifiable Reward (RLVR) effectively solves complex tasks but demands extremely long context lengths during training, leading to substantial computational costs. While multi-stage training can partially mitigate this, starting with overly short contexts often causes irreversible performance degradation, ultimately failing to reduce overall training compute significantly. In this paper, we introduce Thinking-Free Policy Initialization (TFPI), a simple yet effective adaptation to RLVR that bridges long Chain-of-Thought (CoT) distillation and standard RLVR. TFPI employs a simple ThinkingFree operation, explicitly discarding the thinking content via a direct append, to reduce token usage during inference. Training with ThinkingFree-adapted inputs improves performance and lowers token consumption, even in the original slow-thinking mode. Extensive experiments across various benchmarks have shown that TFPI accelerates RL convergence, achieves a higher performance ceiling, and yields more token-efficient reasoning models without specialized rewards or complex training designs. With TFPI only, we train a 4B model to reach 89.0% accuracy on AIME24 and 65.5% on LiveCodeBench using less than 4K H20 hours.

cs.LG↗

arXiv2026

Composition-RL: Compose Your Verifiable Prompts for Reinforcement Learning of Large Language Models

Large-scale verifiable prompts underpin the success of Reinforcement Learning with Verifiable Rewards (RLVR), but they contain many uninformative examples and are costly to expand further. Recent studies focus on better exploiting limited training data by prioritizing hard prompts whose rollout pass rate is 0. However, easy prompts with a pass rate of 1 also become increasingly prevalent as training progresses, thereby reducing the effective data size. To mitigate this, we propose Composition-RL, an offline data-augmentation approach that composes existing prompts before RL to address the pass-rate-1 bottleneck without modifying the RL optimizer. The composed dataset is generated and verified once from the fixed source pool and then used for RL training. Extensive experiments across model sizes from 4B to 30B show that Composition-RL consistently improves reasoning capability over RL trained on the original dataset. Performance can be further boosted with a curriculum variant of Composition-RL that gradually increases compositional depth over training. Additionally, Composition-RL enables more effective cross-domain RL by composing prompts drawn from different domains. All our code, data, and models are publicly available to facilitate future research.

cs.CL↗

arXiv2026

Threshold and Below-Threshold Dynamics for the Energy-Critical NLS: The Higher-Kinetic-Energy Regime

We study the focusing energy-critical nonlinear Schrödinger equation below and at the ground-state energy, focusing on solutions in the upper energy cone, characterized by \[ E(u)\le E(W),\qquad \|\nabla u\|_{L_x^2}>\|\nabla W\|_{L_x^2}, \] where $W$ is the ground state. Unlike the lower cone, where compactness of critical solutions plays a central role in understanding the dynamics, such compactness is generally unavailable on the upper branch. Consequently, previous results have often required additional assumptions, such as finite mass or finite variance. In this paper, we remove the finite-mass assumption in the radial setting and give a complete characterization of the upper-cone dynamics in the natural energy space $\dot H^1_{\mathrm{rad}}$, resolving an open problem posed by Duyckaerts and Merle\cite{DM08}. In particular, this establishes finite-time blow-up of the special threshold solution $W^+$ in the negative time direction in dimensions $d=3,4$. For non-radial data in dimensions $d\ge5$, we show that {for solutions in the interior of the upper cone, the limit superior of their $\dot H^1$-norm satisfies an explicit lower bound as time approaches each endpoint of the maximal lifespan,} while forward-global solutions on the upper threshold sheet that remain below this level must coincide with $W^+$ modulo symmetries. We also obtain a complete classification of upper-threshold dynamics for finite-variance data in all dimensions $d\ge3$.

math.AP↗

arXiv2026

TreeGraft: Adaptive Multi-Drafter Grafting for Tree-Based Speculative Decoding

Speculative decoding accelerates large language model inference through a draft-then-verify paradigm. Building on this, tree-structured methods improve inference by organizing proposals into multiple candidate paths, increasing the accepted length. However, existing tree-structured methods use a single drafter for all drafting steps, creating a dilemma: a smaller drafter is fast but yields lower-quality trees, whereas a larger drafter improves tree quality but suffers from high latency. To address this, we propose TreeGraft, a multi-drafter framework in which drafters of different costs jointly construct a shared draft tree. TreeGraft uses the stronger drafter to rescore candidates by updating scores assigned by the weaker drafter, reselect grafting positions, and recover promising paths left unexplored. It also integrates stronger drafter expansions non-destructively, preserving existing branches that may still be accepted by the target model. Together, these designs improve the quality of the shared draft tree. To control the drafting cost, TreeGraft introduces a lightweight scheduler distilled from an offline value system to decide when to call the stronger drafter. Across 10 model pairs and 6 benchmarks, TreeGraft outperforms the better of the two fixed single-drafter endpoint strategies by 15.1% on average, reaching a maximum gain of 26.6%. Our code is available at https://github.com/fjm9933/TreeGraft.

cs.CL↗

arXiv2026

RAN-JEPA: An Action-Conditioned World Model for Energy-efficient Mobile Network Control

Operational mobile networks are rich in observations but poor in interventions: they continuously generate large volumes of passive telemetry, while controlled action--response trajectories remain comparatively scarce and costly to collect. This observation--intervention asymmetry is a central obstacle to building practical world models for RAN control. We address this problem with RAN-JEPA, an action-conditioned world model that separates representation learning from intervention learning. The model first learns a predictive latent representation from observation-only network data, adapts that representation to the target deployment domain, and then learns action-conditioned latent dynamics from controlled sleep/wake trajectories while keeping the encoder fixed. We evaluate RAN-JEPA using public operator performance data for pretraining and a real-equipment four-cell 5G testbed for domain adaptation, post-training, and control. Across four future 15-minute intervals, RAN-JEPA achieves lower action-conditioned prediction error than a GPT-2 world-model baseline and a mechanistic rule, while remaining close to a strong persistence reference. What-if evaluation further shows that the model produces distinct and coherent future trajectories under matched SLEEP and KEEP plans. In closed-loop control, RAN-JEPA consistently supports higher sleep ratios than the two baselines at comparable congestion cost. These results show that observation-first, action-conditioned learning can turn abundant passive telemetry and limited intervention data into a practical predictive model for energy-aware RAN control.

cs.AI↗

arXiv2026

DPOison-VL: Clean Preference Poisoning in Vision-Language Direct Preference Optimization

Direct Preference Optimization (DPO) has become a widely adopted approach for aligning vision-language models (VLMs), yet the security of multimodal preference data remains largely unexplored. Existing preference poisoning attacks either corrupt explicit supervision signals or require visible triggers, making them detectable under human inspection. We introduce DPOison-VL, the first clean preference poisoning attack against VLM-DPO, which manipulates model alignment solely through imperceptible perturbations on a small fraction of training images, while leaving all prompts, responses, and preference annotations semantically intact. We formulate poison construction as a bilevel optimization and derive a gradient alignment surrogate that efficiently steers image perturbations toward the attacker-specified preference. We further introduce a preference displacement score for principled poison selection with provable first-order optimality under a fixed poison budget, and derive a sufficient condition for preference inversion. Extensive experiments across diverse attack tasks, different VLM architectures and scales, and baseline comparisons demonstrate that DPOison-VL achieves up to 98\% ASR with only 2\% poisoned data and exhibits cross-model transferability, while largely preserving clean accuracy and general utility. Our findings reveal a critical vulnerability in VLM-DPO, where imperceptible image perturbations in preference training data can silently redirect alignment.

cs.CR↗

arXiv2026

DivMoE: Fine-Grained MoE Upcycling via Cross-Domain Expert Composition

Mixture-of-Experts (MoE) architectures have become essential for scaling large language models, with recent work demonstrating the benefits of fine-grained expert designs. Training such models from scratch is expensive, and sparse upcycling from pre-trained dense models is an attractive alternative. However, we identify a structural pathology of fine-grained upcycling: when fine-grained experts are derived from a single source model, naive routing collapses and downstream accuracy drops to near-random (e.g., on Qwen3-1.7B, Drop-Upcycling-fine-grained reaches only 23.2% average accuracy across 15 benchmarks, essentially matching from-scratch training at 22.2%, while the same method's coarse-grained variant reaches 50.2%). We propose DivMoE, the first framework achieving fine-grained MoE Upcycling with structurally-balanced routing. DivMoE introduces domain-specialized fine-grained expert initialization, deriving experts from dense models that have undergone domain-adaptive continual pre-training, and diversity-constrained routing, a hard structural constraint guaranteeing that each token activates experts from distinct domain groups. Across two base models and 15 benchmarks, DivMoE consistently outperforms six upcycling baselines (55.6% vs. 51.6% for the strongest baseline on Qwen3-1.7B) and strictly improves over the dense base model on every benchmark after Stage 2 continual pre-training -- closing the regression gap that has plagued prior fine-grained upcycling. After supervised fine-tuning on a public reasoning mixture, our 12B-parameter DivMoE model matches Moonlight-MoE (16B) at 64.5% average accuracy while outperforming a controlled NVIDIA-Upcycling baseline by 6.1 percentage points.

cs.AI↗

arXiv2026

A near-quadratic lower bound for sets with no unique sums

Let $m(p)$ be the least size of a subset of $\F_p$ with at least two elements for which every sum has two distinct representations as unordered pairs, allowing repetition. We prove that, for every prime $p\ge64$, \[ m(p)\ge2^{-80}\left(\frac{\log p}{\log\log p}\right)^2. \] The argument compresses the full integer collision lattice by unit-pivot elimination. A shared random sample and forests of bounded diameter give $O(\sqrt n+n/\log p)$ surviving coordinates of polynomial height for a minimal set of size $n$. A nonzero minor divisible by $p$ then gives the lower bound. We also construct weakly ternary-balanced seeds yielding \[ m(p)\le\frac{(\log p)^2}{2(\log3)^2} +\left(\frac1{4\log3}+o(1)\right) \frac{(\log p)^2}{\log\log p}. \] Consequently $m(p)=(\log p)^{2+o(1)}$ as $p$ tends to infinity through the primes. The constant-factor order of $m(p)$ remains undetermined.

math.CO↗

arXiv2026

FineART: Fine-Grained Annotated Robotic Trajectory Dataset and Vision-Language-Action Model for Bimanual Manipulation

Robots operating in real-world environments must often execute complex, multi-step bimanual tasks over long horizons rather than single, isolated actions. Current manipulation datasets provide limited support for this capability: although single-arm datasets reach hundreds of thousands of trajectories, they typically provide only one high-level instruction per episode, while existing bimanual datasets provide subtask annotations for only part of their recorded hours. We present FineART, a densely annotated bimanual manipulation dataset comprising 40,543 episodes (1,718 hours) and 533,913 subtasks across 151 tasks. We also introduce FineART-VLA, a vision-language-action policy that predicts its own next subtask to guide its actions. Mid-training on FineART's subtask annotations improves FineART-VLA's success at following spatial instructions from 32.0% to 100.0%. With step-by-step human subtask guidance, success on unseen long-horizon tasks increases from 16.0% to 76.0%. Furthermore, FineART-VLA matches baseline performance on a new robot with 10x less fine-tuning data and generalizes zero-shot to unseen tasks. We open-source the full dataset, model weights, and training code.

cs.RO↗

arXiv2026

Beyond Prompt Count: How Data Shapes Transfer in On-Policy Distillation

On-policy distillation (OPD) trains students using teacher feedback on their own sampled responses, yet how prompt choice shapes transfer across teacher-student pairs remains poorly understood. We systematically study prompt quantity, source, and selection across RL- and SFT-continuation pairs and cross-model settings. We find that OPD can be highly prompt-efficient: a few prompts can approach large-pool performance, with four DAPO prompts matching the observed mathematics score of 3,840 DeepMath prompts. However, prompt utility is relational rather than intrinsic: changing only the teacher can reverse the relative effectiveness of mathematics and code prompts. To characterize these transfer differences, we analyze parameter and functional changes across prompt supports and model pairs. Functional alignment with the teacher varies across supports and target tasks; in continuation pairs, teacher-aligned prediction changes can coexist with weak parameter alignment. Continued OPD on effective supports can restore performance after unfavorable transfer. Finally, targeted selection does not consistently outperform uniform random sampling, and filtering out a source that performs poorly alone yields no consistent gain across three paired support draws. Overall, our results distinguish prompt efficiency from prompt interchangeability and show that effective data choice depends on the teacher-student pair and target capability, with random sampling providing a competitive baseline in the studied settings.

cs.AI↗

arXiv2026

A Low-Complexity Message-Passing Receiver for FTN-WAFT under Nonlinear Distortion

This letter proposes a faster-than-Nyquist (FTN) weighted affine Fourier transform (WAFT) transmission scheme, termed FTN-WAFT, which combines the peak-to-average power ratio (PAPR) flexibility of WAFT with FTN time compression to improve spectral efficiency and power amplifier (PA) efficiency over high-mobility channels. However, operating the PA at a reduced input back-off (IBO) inevitably introduces nonlinear distortion, which further couples with FTN-induced inter-symbol interference (FTN-ISI) and doubly selective channel dispersion. To address this practical impairment, a multilayer message-passing (MLMP) receiver aided by PA interference cancellation (PA-IC), termed PA-IC-MLMP, is proposed. The proposed receiver reconstructs the dominant PA distortion from soft symbol estimates, cancels the channel-propagated nonlinear interference, and performs MLMP over the WAFT, FTN, channel, and constellation layers. Simulation results show that the proposed receiver improves the bit error rate (BER) and effective throughput under nonlinear PA constraints while maintaining low complexity.

eess.SP↗

arXiv2026

Learning to Steer, Steering to See: Unveiling the Geometry of RLVR in Large Language Models via Trainable Vectors

Reinforcement learning (RL) has become a key paradigm for enhancing the reasoning of large language models, yet the high dimensionality of parameter updates makes its training dynamics hard to analyze. We study reinforcement learning with verifiable rewards (RLVR) and use vector steering to identify a low-dimensional effective manifold in activation space associated with RL-induced gains. We uncover two geometric properties. (1) Effective Manifold Capacity: the capacity needed to reproduce RL gains can be very small but is not infinitely compressible; at extremely low capacity, intervention dimensionality and input-dependent expressiveness become key constraints, and this requirement varies with injection depth. (2) Control Manifold Separation: effective control directions lie mainly in the low-variance complement of the activation principal subspace. Within a task and base model, the learned geometry stays largely consistent across training configurations, and across tasks geometric alignment correlates with capability transfer. Experiments on 5 LLMs and 6 verifiable-reward tasks support these findings. We then propose Alpha-Stabler, a plug-and-play framework with a Predictor that monitors principal-subspace intrusion for early collapse warnings, and a Controller that removes the principal-subspace component of activation gradients during backpropagation while preserving the orthogonal complement. Alpha-Stabler stabilizes training for 2,000 steps and consistently improves RL gains, offering practical insights for robust post-training. Code: https://github.com/caiyuchen-ustc/On_Policy_Vector_Training

cs.LG↗

arXiv2026

Robustness of mKdV Breathers: A Numerical Perspective

We systematically investigate breather solutions in the modified Korteweg-de Vries (mKdV) equation via numerical simulations. We show that the breather solutions are stable under a variety of perturbations, including amplitude changes, modifications of internal parameters, and perturbations of the nonlinearity of the equation. Our results are not only consistent with the known analytical studies on the stability of breathers in integrable systems, extending them further, but also provide numerical evidence that such breather-type structures can persist and remain stable over the simulated time scales in some non-integrable settings with symmetric potentials.

nlin.PS↗

arXiv2026

Blow-Up Dynamics for the $L^2$ critical case of the $2$D Zakharov-Kuznetsov equation

We study blow-up dynamics for the $L^2$-critical cubic Zakharov--Kuznetsov equation in two dimensions, \[ \partial_tu+\partial_{x_1}(Δu+u^3)=0 \qquad\text{on }\mathbb R^2. \] For a class of localized $H^1$ perturbations of the ground state $Q$, we establish a trichotomy near the soliton manifold: exit from a small $L^2$-tube, global asymptotic stability, or finite-time blow-up. In the stable blow-up regime, the solution concentrates a single bubble and \[ λ(t)\sim \ell_0(T-t)^{1/(3-c)}, \] where $\ell_0>0$ depends on the initial datum and $c\in(1,2)$ is an explicit constant determined by the transverse tail of the first-order approximate profile. Consequently, \[ \|\nabla u(t)\|_{L^2} \sim \frac{\|\nabla Q\|_{L^2}} {\ell_0(T-t)^{1/(3-c)}}. \] After subtraction of the concentrating soliton, the radiation converges strongly in $L^p(\mathbb R^2)$ for every $2\leq p<\infty$ to a common nonzero profile $u^*$, while \[ u^*\notin H^s(\mathbb R^2) \qquad\text{for every }s\geq\frac c2. \] The stable blow-up branch is open in the relative $H^1$ topology of the localized class. Finally, every non-soliton datum in this class with non-positive energy blows up in finite time. Interval-arithmetic computer-assisted proofs certify the numerical inputs to the virial coercivity argument. They also yield a rigorous enclosure of $c$, justifying the polynomial moment of order $21$ imposed on the initial data.

math.AP↗