Search arXiv⌕ Search

arXiv subjects

Hao Wu

Publications and source records attributed to Hao Wu.

At least 55 records · Page 3Linked to original sources
arXiv2026

Radioactive Feedback: Tracing Models Trained by Label-Only LLM Supervision

Conventional distillation transfers capability through teacher-generated demonstrations. LLM-as-a-Judge knowledge distillation offers a distinct and increasingly practical alternative. In the label-only variant studied here, the student generates every response and improves from scalar scores alone. Prior work shows that this non-demonstrative feedback can improve student capability, but it creates a fundamentally different provenance problem. No teacher-generated answer enters the student's training set, so any evaluator-specific trace must pass through a low-bandwidth reward signal rather than copied text. Existing watermarks and distillation detectors do not cover this score-only channel. We introduce Radioactive Feedback, which encodes a registered pattern of behavioral preferences in the rewards themselves. The evaluator applies bounded pointwise adjustments to quality-qualified responses, allowing repeated optimization to transfer the pattern to a student trained by another party. Experiments on code and mathematics show that the registered preferences transfer through scores alone and across datasets, enabling black-box detection against a fixed reference pool.

cs.AI↗
arXiv2026

RSIPO: Recursive Self-Improving Policy Optimization for Agentic Reinforcement Learning via Skill-Policy Co-Evolution

Recent focus in agentic reinforcement learning (RL) seeks to improve credit assignment in long-horizon tasks. One promising direction combines reinforcement learning with verifiable rewards (RLVR) and on-policy self-distillation (OPSD), integrating trajectory-level advantages with dense token-level supervision. This line of work often conditions a teacher model on privileged context, such as agentic skills, to score tokens generated by the student policy. Despite their progress, they largely overlook the on-policy evolution of agentic skills and the reliability of teacher supervision. Towards this end, we introduce Recursive Self-Improving Policy Optimization (RSIPO), a self-evolving agentic RL framework that enables agentic skills and the policy model to co-evolve throughout the training process. Specifically, RSIPO enables the policy itself to extract skills from completed trajectories and explicitly optimizes its skill-generation capability through trajectory replay. It further employs empirical Bayes shrinkage to dynamically calibrate the student's trust in teacher scores, accounting for the fact that skill-enhanced teacher guidance can sometimes be less reliable. Experiments with Qwen-family models on embodied interaction and search-based agentic tasks demonstrate that RSIPO outperforms all the compared baseline approaches. With Qwen2.5-7B as the base model, RSIPO achieves an average success rate of 98.6\% on ALFWorld, yielding absolute improvements of 2.5\% and 17.4\% over the second-best baseline and standard GRPO, respectively.

cs.CL↗
arXiv2026

On a thermodynamically consistent diffuse interface model for incompressible two-phase flows with unmatched densities: Energy equality and Lyapunov stability

We consider the initial-boundary value problem of a thermodynamically consistent diffuse interface model for incompressible two-phase flows with unmatched densities in a bounded domain $Ω\subset\mathbb{R}^3$. Our first aim is to study the energy equality for global weak solutions by establishing mixed $L_t^qL_x^r$-regularity conditions on the velocity field, its gradient, and its time derivative, under which the global weak solution satisfies the energy equality for all times. The proof is based on the propagation of regularity for weak solutions to the convective Cahn-Hilliard equation with a physically relevant Flory-Huggins-type potential, combined with global mollification and boundary cut-off techniques. Next, we prove the existence and uniqueness of global strong solutions in the general setting with non-constant gradient energy coefficient and non-degenerate mobility, provided that the initial velocity is sufficiently small and the initial phase-field variable is a sufficiently small perturbation of a local minimizer of the free energy. This yields Lyapunov stability for each steady state consisting of a zero velocity together with a local energy minimizer. The proof relies on the energy equality for (local) strong solutions and the Łojasiewicz-Simon approach.

math.AP↗
arXiv2026

Remembering What Matters for Long-Horizon Robotic World Models

Autoregressive robot video world models based on discrete visual tokens remain unreliable over long rollout horizons. Retaining the entire generated history does not necessarily improve prediction quality. We investigate a possible memory-role conflict in Full Context decoding: persistent scene information, recent visual-action states, and stale generated history compete for attention despite their different predictive roles. This perspective motivates ReCAP (Receding Context with Anchor Preservation), a training-free context-scheduling method for long-horizon robot video world models. ReCAP keeps the pretrained tokenizer, world model, and visual decoder frozen. It preserves the initial scene anchor and a contiguous window of recent complete visual-action blocks while evicting older generated history. The schedule preserves within-block structure and temporal order, requires no additional modules or parameter updates, and bounds active-context growth independently of rollout length. We evaluate ReCAP on RT-1, BridgeV2, LIBERO-90, and CALVIN using frozen autoregressive world models. Across evaluated environments and horizons, ReCAP improves visual fidelity, action-conditioned consistency, and robot structure preservation relative to Full Context. On BridgeV2, it reduces LPIPS by 63.3%. In the system benchmark at H=64, it reduces the logical KV footprint by 73.3% and p50 latency by 6.51%. Attention analyses show increased attention to retained recent states, while history interventions associate larger amounts of stale context with higher prediction error. These observations are consistent with the memory-role-conflict hypothesis, although they do not isolate it from context-length and position-handling effects. Our results support inference-time context scheduling as a practical approach to improving long-horizon robot video prediction and computational efficiency without retraining.

cs.RO↗
arXiv2026

Representation Drift in Continual Learning: LoRA versus N-gram Memory

A language model that keeps learning online must not slowly erode its own base capabilities. We ask which part of a modern LLM should absorb continual updates, by measuring how each choice changes the model's internal representations. On a 48-layer mixture-of-experts model with a native pretrained N-gram memory, we train a seven-task chain under three update mechanisms with identical data and budgets: updating only the N-gram memory table, continually training one shared LoRA adapter, and training a fresh LoRA per task that is merged into the backbone. We capture fixed-context representations at every layer and task boundary. Updating only the N-gram memory changes the base model's representations of general math and code inputs by 26%/15% (relative RMS), against 64%/35% for merged and 97%/39% for shared LoRA. This drift does not grow with the number of tasks: it reaches 25% after the first task and stays there for all seven, while both LoRA variants keep accumulating. Its generic behaviour is essentially unchanged (no code problem lost, math NLL -0.01), and every change is confined to inspectable, reversible table rows. The cost appears elsewhere: N-gram memory acquires new tasks more weakly and its stored tasks interfere with one another, losing as much old-task NLL as shared LoRA. We show that this forgetting is governed by drift on the learned tasks' own inputs, measured from the checkpoint at which each was learned, not by generic drift (Spearman rho=0.85; rho=0.71 on Llama-3-8B), and we derive the exact relations behind it: with a frozen head, every pairwise logit margin changes by the projection of drift onto a head-difference direction. These results support a design principle for long-running online learning: write new knowledge only into the memory table and keep the backbone fixed; the remaining problem, interference between stored tasks, then lies inside the memory.

cs.LG↗
arXiv2026

MagCilia: A Compact Magnetociliary Tactile Sensor with 3D Force Sensing for Robotic Contact Perception and Grasping Feedback

Robotic grasping and surface exploration benefit from simultaneous measurement of normal and tangential forces and from surface information obtained through contact. Here, we present a compact magnetociliary tactile sensor (MagCilia) that combines a flexible magnetic-cilia structure with a Hall sensor for 3D force sensing. Quasi-static finite element analysis is used to investigate structural deformation and magnetic responses under multidirectional loading. To reconstruct forces from the coupled magnetic channels, we propose causal history fusion regression (CHFR), which combines current magnetic-field measurements with their recent changes. Five-fold cross-validation grouped by calibration record yields root-mean-square errors of 0.40, 0.57, and 0.69 N for Fx, Fy, and Fz, respectively, with corresponding coefficients of determination of 0.93, 0.90, and 0.92. Robotic experiments demonstrate tangential-force-guided gripper adjustment and multi-axis load monitoring under external perturbations. Frequency-domain features of the reconstructed forces distinguish six surface categories with 99.39% accuracy in three-fold cross-validation grouped by acquisition session. An online robotic demonstration additionally identifies all six tested surfaces. These results demonstrate 3D force reconstruction, grasping feedback, and surface recognition using a single compact tactile unit.

cs.RO↗
arXiv2026

GeoGR: Spatial-Semantic Token Compression and Group Routing for Efficient 3D VLMs

Multi-view representations enable pretrained image and video vision-language models to understand 3D scenes, but their long visual token sequences make inference costly. Redundancy arises both from repeated observations of the same scene regions across views and from processing all visual tokens at every LLM layer, regardless of their relevance to the current computation. We propose GeoGR, a two-stage framework that jointly compresses scene representations and visual computation. At the projector level, GeoSemZip consolidates cross-view observations into 3D voxel tokens and combines semantic saliency with spatial coverage to construct a compact scene representation. Unselected voxel features are aggregated into contextual anchors to retain scene information that would otherwise be discarded. At the LLM level, we identify an effective visual computation window and observe that query-conditioned token relevance remains locally stable but changes over longer layer intervals. Guided by these findings, GroupRoute combines late entry and early exit with recoverable group-wise skipping. Low-relevance tokens bypass complete LLM layers while retaining their hidden states and can rejoin the active group at subsequent anchor layers selected under a computation budget. For compression-aware post-training, we construct a compact dataset of 67K samples by retaining 30% of the original training pool. With this protocol, GeoGR retains 95.2% and 96.8% of the dense post-trained models' performance on LLaVA-OneVision-7B and Video-3D LLM-7B, respectively, while accelerating LLM prefill by $6.8\times$ and $6.1\times$. These results demonstrate the consistent effectiveness of GeoGR across 3D VLMs with and without explicit 3D spatial encoding, highlighting a favorable performance-efficiency trade-off. The code is available at https://github.com/EIT-NLP/GeoGR.

cs.CV↗
arXiv2026

TAM: Task-Aware Memory Distillation for Efficient Spatiotemporal Prediction

Knowledge distillation enables efficient spatiotemporal prediction by transferring knowledge from an accurate teacher to a compact student. However, matching outputs or features independently for each sample leaves cross-sample predictive structure underused. Exploiting this structure requires representations and historical references that reflect the dynamics of each task. We propose TAM, a Task-Aware Memory Distillation framework that organizes a frozen teacher's knowledge into a bounded, retrievable history. Memory entries encode latent features, forecast changes, or flow residuals, while task-specific selection rules identify relevant historical references. The student either matches the teacher's similarity distribution over shared references or regresses observation-conditioned residual prototypes. These objectives complement supervised prediction and conventional distillation. The teacher, memory, and auxiliary adapters are used only during training, leaving student inference unchanged. We evaluate TAM on video prediction, weather forecasting, and traffic flow prediction across multiple teacher-student configurations. Averaged over four paired runs, adding TAM improves SSIM on all six video datasets and reduces MSE on five relative to the corresponding KD baselines. Mean paired MSE reductions reach 1.86% on KittiCaltech, 1.93% on WeatherBench with a gSTA teacher, and 1.01% on TaxiBJ. These results demonstrate the utility of historical teacher supervision across distinct forecasting tasks without additional student inference cost.

cs.CV↗
arXiv2026

Well-posedness of Strong Solutions for a Thermodynamically Consistent Diffuse-interface Model for Incompressible Two-phase Flows with a Soluble Surfactant

We consider a thermodynamically consistent diffuse-interface model for binary mixtures of incompressible viscous Newtonian fluids with a soluble surfactant. This hydrodynamic system generalizes the Abels-Garcke-Grün model for two-phase flows with unmatched densities by incorporating an additional Cahn-Hilliard equation for the surfactant concentration. We study the strong well-posedness of the initial-boundary value problem with physically relevant singular potentials, subject to a no-slip boundary condition for the fluid velocity and homogeneous Neumann boundary conditions for the phase-field variables and the corresponding chemical potentials. First, we establish the existence of local-in-time strong solutions in three dimensions and global-in-time strong solutions in two dimensions. In the two-dimensional case, we further show that, under natural growth assumptions on the singular potentials, strong solutions remain strictly separated from the pure phases, thereby guaranteeing uniqueness. In the three-dimensional case, we establish local-in-time uniqueness of strong solutions, provided that the initial data are strictly separated from the pure phases. Our results extend previous work on strong solutions to the Abels-Garcke-Grün model from the special case of constant mobility to the more general scenario with concentration-dependent mobilities and interactions with a surfactant.

math.AP↗
arXiv2026

Quasi-exponentials and semi-classical limits for the chordal BPZ system

We study the nonlinear Hamilton--Jacobi system arising as the formal semi-classical limit of the chordal Belavin--Polyakov--Zamolodchikov equations with a common eigenvalue parameter $λ$. We determine the number of global real-valued solutions modulo additive constants. When $λ=0$, the number of solutions can be derived using the connection between the solutions and rational functions, due to Eremenko. We focus on the case when $λ\neq 0$ in this article. For $λ>0$, a Schwarzian correspondence relates the algebraic equations for the gradients to quasi-exponentials with prescribed critical points. Enumeration and transversality results of Mukhin, Tarasov, and Varchenko yield smooth real branches, and a Poisson bracket identity establishes their exactness. An energy bound rules out global real-valued solutions for $λ<0$.

math-ph↗
arXiv2026

How Much Evidence Should a Coding Agent's Self-Correction Carry? Adaptive Dirichlet Evidence for Self-Distillation

Execution feedback lets coding agents revise programs and learn from their own corrections. A correction's learning weight should reflect both the transitions supported by its executions and the amount of evidence behind that support. We introduce Effective-Evidence Self-Distillation (EESD), which represents these quantities separately. Normalized execution relevance determines relative transition support and an effective pseudo-count mass; a Dirichlet posterior then produces an uncertainty-penalized weight for KL-anchored correction learning. Under a symmetric prior, changing mass preserves category ordering, and effective mass yields a supervised coefficient bounded by its matched fixed-mass counterpart. Across four model-domain history sweeps, increasing visible observations from one to eight reduces future-outcome NLL by 55.0-59.3%. At eight observations, effective mass achieves lower NLL than fixed mass in all four comparisons. In the primary matched DeepSeek/RunBugRun study, argmax predictions agree on all 3,000 examples, with the largest NLL gain under concentrated relevance. After one correction-learning round, DeepSeek/CodeARC all-tests Pass@1 increases from 15.0% to 20.4%, with a paired 95% source-bootstrap interval of [+2.8, +8.0] percentage points. The twelve-setting downstream evaluation establishes the model-domain scope of this update. These results show how separating evidence support from evidence mass changes probability estimation and correction learning in coding agents.

cs.AI↗
arXiv2026

Gaussian free field with logarithmic singularities and reduced Green's energy

We analyze the level lines of the Gaussian Free Field with logarithmic singularities in polygonal domains. We focus on the (none-zero) singular case and compare it to the singular-free setting. The law of the level lines for the field in singular case is absolutely continuous with respect to the law for the singular-free case, with a Radon-Nikodym derivative related to the reduced Green's energy which is introduced by F. Viklund and D. W. Nyström~[VN24]. These level lines admit a Brownian motion representation. In the singular-free case, such representation involves Poisson point processes of Brownian loops and excursions; while for the singular case, it additionally includes Poisson point processes of Brownian bubbles and bridges. The special cases of one or two singularities yield moment-generating functions of the conformal radius, which can be interpreted as one-point and two-point correlation functions of vertex operators in conformal field theory at the free-boson point. While the Radon-Nikodym derivative is bounded away from the singularities, the behavior of the level lines near the singularities is markedly different from the singular-free case, with different asymptotic probabilities for all level lines to remain close to the singularity.

math.PR↗
arXiv2026

Multiple SLEs for $κ\in (0,8)$: Coulomb gas integrals and pure partition functions

In this article, we give an explicit relationship of SLE partition functions with Coulomb gas formalism of conformal field theory. We first construct a family of SLE$(κ)$ partition functions as Coulomb gas integrals and derive their various properties. In accordance with an interpretation as probabilistic correlations in loop $O(n)$ models, they are always positive when $κ\in (8/3,8)$, while they may have zeroes for $κ\le 8/3$. They also admit a Frobenius series expansion that matches with the algebraic content from CFT. Moreover, we check that at the first level of fusion, they have logarithmic asymptotic behavior when $κ= 8/3$ and $κ= 8$, in accordance with logarithmic minimal models $M(2,1)$ and $M(2,3)$, respectively. Second, we construct $\SLE_κ$ pure partition functions and show that they are real-analytic in $κ\in (0,8)$ and decay to zero as a polynomial of $(8-κ)$ as $κ\to 8$. We explicitly relate the Coulomb gas integrals and pure partition functions together in terms of the meander matrix. As a by-product, our results yield a construction of global non-simple multiple chordal SLE$(κ)$ measures ($κ\in (4,8)$) uniquely determined by their re-sampling property.

math-ph↗
arXiv2026

Classification of commutation relation for multi-chordal SLE

We classify the linear solution spaces of the chordal Belavin--Polyakov--Zamolodchikov (BPZ) equations, which arise in Dubédat's commutation relations for multiple Schramm--Loewner evolutions (SLE). Without imposing growth conditions, we determine the subspaces selected by successive conformal Ward identities and their exact dimensions. For every $κ>0$ and BPZ spectral parameter $λ\in\mathbb{R}$, we determine the spectrum and Jordan structure of the translation operator. When $λ=0$ and $κ\in(0,8)$, we further determine all admissible scaling exponents and the dimensions of the corresponding translation-invariant solution spaces. A triangular change of derivative coordinates yields a first-order system of rational Knizhnik--Zamolodchikov type, allowing us to represent the Ward operators explicitly and reduce the classification to finite-dimensional linear algebra. For $2N$ boundary points and $κ\in(0,8)$, translation-invariant BPZ solutions with $λ=0$ and the homogeneity required by Möbius covariance automatically satisfy the third Ward identity and the usual power-law bound. This recovers, under weaker assumptions, the Catalan dimension formula previously established by Flores and Kleban. At $κ=8$, the third Ward identity imposes an additional constraint when $N\ge2$, but the full Möbius-covariant solution space still has Catalan dimension, establishing completeness of the partition functions constructed from uniform spanning trees. Finally, at $κ=4$ and $λ>0$, we construct an explicit basis of positive BPZ solutions and identify its elements as partition functions for level lines of a Gaussian free field with suitable harmonic means.

math.PR↗
arXiv2026

Uniform Local Asymptotics for Lévy Processes with Subexponential Jumps

This paper is devoted to unifying the uniform local large-deviation asymptotics for a centered Lévy process $X$ with subexponential jumps. Our results assert that for any $θ,δ_0>0$ and $K\geq0$, $$\lim_{t\to\infty}\sup_{x\geqθt}\sup_{|y|\leq Kb(x)}\sup_{δ\in[δ_0,\infty]}\sup_{0<s\leq t}\bigg|\frac{\mathbf P\big(X_s\in(x-y,x-y+δ]\big)}{s\cdot\mathbf P\big(X_1\in(x,x+δ]\big)}-1\bigg|=0,$$ where the natural-scale function $b$ satisfies a polynomial growth condition. This provides a continuous-time and simultaneously uniform analogue of the results of Denisov et al. [Ann. Probab., 2008], while being established under a weaker moment assumption.

math.PR↗
arXiv2026

Local Asymptotics for Entrance Laws of Reflected Lévy Excursions

This paper establishes uniform local large-deviation asymptotics for the entrance law $\underline n_t(dx)$ of excursions of a Lévy process with negative drift and locally regularly varying Lévy measure $ν$, reflected at its running infimum. In the fixed-time regime, we prove that as $x\to\infty$, $$ \underline n_t\big((x,x+δ]\big) \sim \underline n(ζ\wedge t)\, ν\big((x,x+δ]\big), $$ uniformly in $δ\in[δ_0,\infty]$ for every $δ_0>0$. Our principal result concerns the large-time regime and shows that, as $t\to\infty$, $$ \underline n_t\big((x,x+δ]\big)\sim \int_x^{x+δ} V(z) \,ν(βt+dz), $$ uniformly in $x\geq0$ and $δ\in[δ_0,\infty]$, where $V$ denotes the renewal function. Using a fluctuation-theoretic decomposition of the killed semigroup, we further derive corresponding uniform local large-deviation asymptotics for the Lévy process killed upon its first passage below zero.

math.PR↗
arXiv2026

Clipped Affine Policy: Low-Complexity Near-Optimal Online Power Control for Energy Harvesting Communications over Fading Channels

This paper studies online power control for battery-limited point-to-point energy harvesting communications over slow block-fading channels. A linear-policy-based approximation is developed for the relative-value function in the Bellman equation of the power control problem. This approximation leads to two fundamental parameterized clipped affine policies: an optimistic policy derived from a certainty-equivalence-type approximation and a robust policy derived from worst-case analysis. For independent and identically distributed energy arrivals and channel states, two families of power control schemes are developed based on the optimistic clipped affine (OCA) and robust clipped affine (RCA) policies, respectively. The proposed adaptive RCA policy based on reinforcement learning (RCA-RL) is further extended to address four scenarios with contextual information: one-step energy lookahead, one-step channel lookahead, one-step joint energy-channel lookahead, and Markov energy arrivals. Extensive simulation results show that the proposed schemes provide a favorable tradeoff between computational complexity and performance. The adaptive RCA policy based on the maximin optimal linear-policy-slope approximation (RCA-OLA-A) and the RCA-RL scheme achieve the best overall performance, while the RCA policy based on the maximin optimal linear policy (RCA-OL) is the best-performing closed-form policy. In particular, RCA-OLA-A, RCA-RL, and the aforementioned RCA-RL extensions achieve less than 2% performance loss relative to the optimal policy across a range of scenarios, consistently outperforming the considered benchmark approaches, including generic reinforcement learning baselines. The RCA-OL policy also performs well with less than 4% performance loss.

cs.IT↗
arXiv2026

CHILLGuard: Towards Fine-Grained Chinese LLM Safety Guardrail with Scalable Data Construction and Model-aware Preference Alignment

Malicious content generated from large language models (LLMs) could pose severe safety risks and ethical concerns. While existing LLM safety guardrails excel in English or multilingual settings, they lack adaptation to Chinese-specific regulatory policies, cultural context, and linguistic nuances, failing to support fine-grained risk classification for diverse deployment needs. In this paper, we introduce a 5-macro, 31-micro category fine-grained risk taxonomy for Chinese scenarios, and build CHILLGuard: a dedicated Chinese LLM content safety guardrail. To address the critical scarcity of high-quality annotated Chinese safety data, we propose a scalable multi-stage data construction pipeline: we expand multi-source corpus via retrieval-augmented generation, generate implicit harmful samples through prompt engineering rewriting, and refine high-quality data via multi-model voting-based label calibration. Based on this, we build CHILLGuardTrain, a large-scale training set with 405,007 samples, and CHILLGuardTest, a rigorously curated annotated test set with 51,745 samples. We then train CHILLGuard on CHILLGuardTrain under a generator-classifier collaborative framework via Model-aware Direct Preference Optimization. Extensive experiments under multiple settings demonstrate the state-of-the-art performance of CHILLGuard, e.g., a 15.92% relative improvement of F1 score over Qwen3Guard-8B-Strict on our benchmark. We release our resources at https://github.com/cswbyu/CHILLGuard.

cs.CL↗