Searcharxiv⌕ Search

arXiv subjects

Zehua Cheng

Publications and source records attributed to Zehua Cheng.

At least 19 recordsLinked to original sources
arXiv2026

Constraint-Anchored Reasoning Traces

Autoregressive multimodal large language models (MLLMs) suffer from error snowballing: a single incorrect inference early in a chainof-thought (CoT) trace corrupts all downstream reasoning. We find that in state-of-the-art open-source MLLMs, once the first error occurs, the reasoning cascades into failure across all remaining steps in 65% of such cases (a metric we term the snowball rate). Existing mitigations-sampling multiple chains, post-hoc self-verification, or full program synthesis-either lack symbolic grounding, catch errors too late, or sacrifice the flexibility of natural language reasoning. We propose Constraint-Anchored Reasoning Traces (CART), a neuro-symbolic framework that trains MLLMs to interleave natural language reasoning steps with symbolic constraint assertions: lightweight, machine-checkable statements about visual content (e.g., count(red_objects) = 3). A dual-pronged Constraint Propagation Module-combining a learned neural grounding head with Boolean Constraint Propagation-continuously verifies these anchors against extracted visual features and checks their mutual logical consistency. When a contradiction is detected, a backtrack controller halts generation and reverts to the last consistent checkpoint, preventing error propagation. A variable-frequency emission mechanism allows the model to adaptively control anchor density, avoiding trace bloat. We construct 218K training instances by augmenting GQA, CLEVR-CoGenT, and VCR with ground-truth constraint annotations derived from scene graphs, and fine-tune open-source MLLMs (LLaVA-NeXT, Qwen2-VL) via LoRA. On five benchmarks, CART reduces the snowball rate from 0.65 to 0.14, improves GQA accuracy by +4.6 percentage points over trainingonly baselines, and achieves 89.1 F1 on POPE-all with at most 18% inference overhead.

cs.AI↗
arXiv2026

Equivariant Neural Belief Propagation

Probabilistic inference over spatially embedded variables requires beliefs that respect $SE(3)$ symmetry, yet existing equivariant networks produce only scalars and vectors -- not the rank-2 precision tensors needed for anisotropic uncertainty, and single-component messages collapse multi-modal energy landscapes to physically meaningless averages. We introduce Equivariant Neural Belief Propagation (ENBP), a factor-graph framework whose messages are equivariant Gaussian mixture models with sufficient statistics that transform exactly under $SE(3)$. Rank-2 precision matrices are synthesised via equivariant outer products, ingested through differentiable spectral decomposition, and kept tractable by a greedy KL-based mixture reduction that provably commutes with $SE(3)$. On GEOM-QM9 and GEOM-Drugs, ENBP achieves 98.9% conformational coverage at 0.090 $\mathring{A}$ error with sub-second latency -- over $100\times$ faster than diffusion baselines at higher accuracy. On multi-body robotic inference, vanilla loopy BP diverges at 15+ agents while ENBP converges with near-zero collision rates and machine-precision equivariance error (${\sim}10^{-7}$ vs.\ $10^{-1}$ for augmented baselines).

cs.LG↗
arXiv2026

Invariant Gradient Alignment for Robust Reasoning Distillation

Large language models (LLMs) suffer from shortcut learning: they systematically fail on out-of-distribution (OOD) inputs whose semantic surface differs from training data, even when the logical structure is identical. This undermines knowledge distillation pipelines that transfer chain-of-thought reasoning to smaller students. We introduce Invariant Gradient Alignment (IGA), a training framework that aligns gradient updates across semantically diverse but logically isomorphic examples via three innovations: (i) Logical Isomer Sets, groups of problems sharing identical logical structure across distinct semantic domains (mathematics, medicine, law, science); (ii) a differentiable \emph{Continuous Gradient Conflict Mask}, that suppresses parameter dimensions with high cross-domain gradient variance while preserving invariant directions; and (iii) a truncated SVD projection of the masked gradient back onto the LoRA low-rank manifold, maintaining parameter efficiency throughout. Theoretically, IGA yields tighter OOD generalization bounds than ERM, scaling with the number of isomer domains, and converges at the standard SGD rate under mild regularity. Empirically, IGA outperforms eight baselines across four benchmarks with accuracy gains up to 14.3 pp over ERM-SFT and a Logical Consistency Score of 0.031 versus 0.142 -- a fourfold improvement in representational invariance.

cs.LG↗
arXiv2026

Imbuing Large Language Models with Bidirectional Logic for Robust Chain Repair

Autoregressive chain-of-thought (CoT) reasoning in large language models (LLMs) is fundamentally forward-directed: each step conditions only on prior tokens. This unidirectional inductive bias renders even capable models susceptible to error snowballing, wherein a single logical or arithmetic mistake in an early step irreversibly corrupts the entire reasoning chain. We introduce Teleological Reasoning Infilling (\TRI{}), a training framework that endows decoder-only transformers with a native \emph{goal-conditioned bridging} capability. The key insight is to reframe erroneous reasoning segments as fill-in-the-middle (FIM) tasks: given a verified prefix premise $P$, a verified downstream milestone $S$, and the original query $Q$, the model must synthesise the logical bridge $M$ that connects $P$ to $S$ rigorously and completely. To achieve this with standard causal architectures, we introduce a Prefix-Suffix-Middle (PSM) sequence rearrangement with three non-overlapping sentinel tokens, enabling $M$ to attend to both $P$ and $S$ without any structural modification to the self-attention mechanism. Training proceeds in two stages: (i) Supervised Fine-Tuning (SFT) on symbolically verified $(P, S, M)$ triples extracted from formal mathematics corpora, and (ii) Direct Preference Optimisation (DPO) with a deterministic symbolic verifier (Lean 4 / Python) as the sole reward oracle, eliminating LLM-judge sycophancy. At inference, TRI operates as a surgical repair module within a dual-system loop: a causal draft model generates an initial trace, the verifier pinpoints failures, and TRI infills only the damaged segment, leaving verified sections intact. Comprehensive experiments on three benchmarks demonstrate that TRI achieves state-of-the-art performance across all tasks, while reducing per-problem token expenditure by 31.2%.

cs.CL↗
arXiv2026

In-Context Graphical Inference

Marginal inference in discrete graphical models forces a choice between exactness and scalability: exact algorithms are intractable for high-treewidth graphs, while iterative approximations (Belief Propagation, variational methods) sacrifice convergence guarantees on frustrated topologies. We argue that this dichotomy stems from a mismatched inductive bias: iterative methods abandon the sequential elimination structure that makes exact inference correct. We introduce In-Context Graphical Inference (ICG-I), an autoregressive Graph Transformer that restores this structure by mimicking Variable Elimination with learned, Tensor- Train-compressed intermediate factors, paired with a Dirichlet output layer and Weighted Conformal Prediction for calibrated, distribution-free coverage guarantees under topological shift. We prove that TT compression errors propagate at most lincarly through the autoregressive chain, that the Dirichlet-Multinomial loss is a proper scoring rule, and that WCP maintains coverage with a quantifiable degradation under estimated density ratios. We conducted intensive experiments to evaluate ICG-I and achieved state-of-the-art performance across all benchmarks. ICG-I reduces MAE from 0.041 (best baseline) to 0.020 on standard instances and achieves 0.048 on N=500 frustrated spin glasses where BP diverges entirely.

cs.LG↗
arXiv2026

CasualSynth: Generating Structurally Sound Synthetic Data

Large Language Models (LLMs) generate realistic synthetic data but offer no guarantee that their outputs respect the causal mechanisms governing the target domain. We introduce CausalSynth, a framework that decouples causal structure generation from semantic realization, yielding synthetic data that is both causally valid and linguistically rich. The framework operates in three phases. First, a Structural Causal Model (SCM) - a tuple of structural equations defined over a directed acyclic graph (DAG) generates causal skeletons, i.e., variable assignments that satisfy the Global Markov Property of the governing DAG, via ancestral sampling. Second, an LLM acts as a constrained \emph{realizer}, a conditional translator that maps each skeleton to a high-dimensional observation such as a clinical note or a transaction log. Third, an Iterative Consistency Verification module detects structural violations through deterministic extraction and feeds targeted corrections back to the LLM, forming a closed-loop refinement process. We identify the Semantic Backdoor problem the systematic tendency of LLMs to override imposed causal facts with pre-training priors -- and prove that our iterative mechanism reduces the resulting selection bias relative to standard rejection sampling. On three causal benchmarks (ASIA, ALARM, and MIMIC-Struct), CausalSynth preserved conditional independencies with false-positive rates near the nominal $\alpha=0.05$ level and achieved realizability rates above 96% with 70B-parameter LLM backbones. The framework additionally supports principled interventional and counterfactual generation through noise retention and graph mutilation.

cs.LG↗
arXiv2026

Differentially Private Motif-Preserving Multi-modal Hashing

Cross-modal hashing enables efficient retrieval by encoding images and text into compact binary codes. State-of-the-art methods rely on semantic similarity graphs derived from user interactions for supervision, yet these graphs encode sensitive behavioral patterns vulnerable to link reconstruction attacks. Existing privacy-preserving approaches fail on graph-structured data: Differentially Private SGD destroys relational motifs by treating samples independently, while graph synthesis methods suffer from unbounded local sensitivity in scale-free networks, hub nodes cause single-edge modifications to alter triangle counts by $\mathcal{O}(N)$, necessitating prohibitive noise injection. We term this phenomenon Hubness Explosion. We propose DMP-MH, a Sanitize-then-Distill framework that decouples privacy from representation learning. Our approach first bounds sensitivity by deterministically clipping node degrees, capping the $L_2$-sensitivity of triangle motifs independently of dataset size. A sanitized synthetic graph is then generated via Noisy Mirror Descent under $(\epsilon,\delta)$-Edge Differential Privacy. Finally, dual-stream hashing networks distill this topology using a holistic structural loss that enforces cross-modal alignment. Evaluated on MIRFlickr-25K and NUS-WIDE under a strict inductive protocol, DMP-MH outperforms private baselines by up to 11.4 mAP points while retaining up to 92.5% of non-private performance.

cs.IR↗
arXiv2026

Identity-Decoupled Anonymization for Visual Evidence in Multi-modal Retrieval-Augmented Generation

Multi-modal retrieval-augmented generation (MRAG) systems retrieve visual evidence from large image corpora to ground the responses of large multi-modal models, yet the retrieved images frequently contain human faces whose identities constitute sensitive personal information. Existing anonymization techniques that destroy the non-identity visual cues that downstream reasoning depends on or fail to provide principled privacy guarantees. We propose Identity-Decoupled MRAG, a framework that interposes a generative anonymization module between retrieval and generation. Our approach consists of three components: (i)a disentangled variational encoder that factorizes each face into an identity code and a spatially-structured attribute code, regularized by a mutual-information penalty and a gradient-based independence term; (ii)a manifold-aware rejection sampler that replaces the identity code with a synthetic one guaranteed to be both distinct from the original and realistic; and (iii)a conditional latent diffusion generator that synthesizes the anonymized face from the replacement identity and the preserved attributes, distilled into a latent consistency model for low-latency deployment. Privacy is enforced through a multi-oracle ensemble of face recognition models with a hinge-based loss that halts optimization once identity similarity drops below the impostor-regime threshold.

cs.CV↗
arXiv2026

CircuitSynth: Reliable Synthetic Data Generation

The generation of high-fidelity synthetic data is a cornerstone of modern machine learning, yet Large Language Models (LLMs) frequently suffer from hallucinations, logical inconsistencies, and mode collapse when tasked with structured generation. Existing approaches, such as prompting or retrieval-augmented generation, lack the mechanisms to balance linguistic expressivity with formal guarantees regarding validity and coverage. To address this, we propose CircuitSynth, a novel neuro-symbolic framework that decouples semantic reasoning from surface realization. By distilling the reasoning capabilities of a Teacher LLM into a Probabilistic Sentential Decision Diagram (PSDD), CircuitSynth creates a tractable semantic prior that structurally enforces hard logical constraints. Furthermore, we introduce a convex optimization mechanism to rigorously satisfy soft distributional goals. Empirical evaluations across diverse benchmarks demonstrate that CircuitSynth achieves 100% Schema Validity even in complex logic puzzles where unconstrained baselines fail (12.4%) while significantly outperforming state-of-the-art methods in rare-combination coverage.

cs.CL↗
arXiv2026

Generic Metrics on $S^{n+1}$ Preclude Linearly Stable Singular Tangent Cones of Area-Minimizing Boundaries

We prove that for a residual (and hence dense) subset $\mathcal{G}$ of Riemannian metrics on $S^{n+1}$ in the $C^{3}$ topology, no area-minimizing integral $n$-current that is a boundary admits a singular tangent cone which is linearly stable in the Euclidean sense. The proof proceeds in three stages. First, we develop a perturbation theorem: given any area-minimizer possessing an isolated singularity whose unique tangent cone $C$ is linearly stable, we construct an explicit $C^{3}$-small metric perturbation that destroys the compatibility conditions required for $C$ to persist as a tangent cone. The construction rests on the Hardt--Simon asymptotic expansion near isolated singularities, the spectral theory of the Jacobi operator on the cross-section of $C$, and a surjectivity argument showing that the map from compactly supported metric variations to forcing terms in the linearised minimal-surface equation on $C$ has dense range. Second, we establish that the set of metrics admitting no area-minimizer with a prescribed cone type as tangent cone is open, using compactness of integral currents and upper-semicontinuity of the density function. Third, we assemble these ingredients via a Baire category argument, intersecting countably many open dense sets to obtain the residual set $\mathcal{G}$. An extension to non-isolated singularities is outlined using Federer--Almgren dimension reduction.

math.DG↗
arXiv2026

Constrained Policy Optimization for Provably Fair Order Matching

Automated matching engines execute millions of orders per session, yet systematic asymmetries in latency, order size, and market access compound into persistent execution disparities that erode participant trust. We formulate provably fair order matching as a Constrained Markov Decision Process and propose CPO-FOAM (Constrained Policy Optimization with Feedback-Optimized Adaptive Margins). An inner loop computes an analytic trust-region step on the Fisher information manifold; a PID-controlled outer loop dynamically tightens safety margins, suppressing the sawtooth oscillations endemic to Lagrangian methods under non-stationary dynamics. Group fairness (demographic parity, equalized odds) enters the CMDP cost vector while individual Lipschitz fairness is enforced deterministically via spectral normalization. We prove BIBO stability and that the integral term drives steady-state violations to zero. On LOBSTER NASDAQ data across six market regimes, CPO-FOAM recovers 95.9% of unconstrained throughput at 2.5% constraint violation frequency; on crypto-asset LOB data under MEV injection it captures 98.4% of the reward envelope at 3.2% CVF. The method scales sub-linearly to M=8 constraints, settles on-chain within one Ethereum block, and yields a 2.1X reward improvement on Safety-Gymnasium, confirming domain-agnostic generalization.

cs.GT↗
arXiv2026

Visual Set Program Synthesizer

A user pointing their phone at a supermarket shelf and asking "Which soda has the least sugar?" poses a difficult challenge for current visual Al assistants. Such queries require not only object recognition, but explicit set-based reasoning such as filtering, comparison, and aggregation. Standard endto-end MLLMs often fail at these tasks because they lack an explicit mechanism for compositional logic. We propose treating visual reasoning as Visual Program Synthesis, where the model first generates a symbolic program that is executed by a separate engine grounded in visual scenes. We also introduce Set-VQA, a new benchmark designed specifically for evaluating set-based visual reasoning. Experiments show that our approach significantly outperforms state-of-the-art baselines on complex reasoning tasks, producing more systematic and transparent behavior while substantially improving answer accuracy. These results demonstrate that program-driven reasoning provides a principled alternative to black-box visual-language inference.

cs.MM↗
arXiv2026

Provable Defense Framework for LLM Jailbreaks via Noise-Augumented Alignment

Large Language Models (LLMs) remain vulnerable to adaptive jailbreaks that easily bypass empirical defenses like GCG. We propose a framework for certifiable robustness that shifts safety guarantees from single-pass inference to the statistical stability of an ensemble. We introduce Certified Semantic Smoothing (CSS) via Stratified Randomized Ablation, a technique that partitions inputs into immutable structural prompts and mutable payloads to derive rigorous lo norm guarantees using the Hypergeometric distribution. To resolve performance degradation on sparse contexts, we employ Noise-Augmented Alignment Tuning (NAAT), which transforms the base model into a semantic denoiser. Extensive experiments on Llama-3 show that our method reduces the Attack Success Rate of gradient-based attacks from 84.2% to 1.2% while maintaining 94.1% benign utility, significantly outperforming character-level baselines which degrade utility to 74.3%. This framework provides a deterministic certificate of safety, ensuring that a model remains robust against all adversarial variants within a provable radius.

cs.CL↗
arXiv2026

A Control Theoretic Approach to Decentralized AI Economy Stabilization via Dynamic Buyback-and-Burn Mechanisms

The democratization of artificial intelligence through decentralized networks represents a paradigm shift in computational provisioning, yet the long-term viability of these ecosystems is critically endangered by the extreme volatility of their native economic layers. Current tokenomic models, which predominantly rely on static or threshold-based buyback heuristics, are ill-equipped to handle complex system dynamics and often function pro-cyclically, exacerbating instability during market downturns. To bridge this gap, we propose the Dynamic-Control Buyback Mechanism (DCBM), a formalized control-theoretic framework that utilizes a Proportional-Integral-Derivative (PID) controller with strict solvency constraints to regulate the token economy as a dynamical system. Extensive agent-based simulations utilizing Jump-Diffusion processes demonstrate that DCBM fundamentally outperforms static baselines, reducing token price volatility by approximately 66% and lowering operator churn from 19.5% to 8.1% in high-volatility regimes. These findings establish that converting tokenomics from static rules into continuous, structurally constrained control loops is a necessary condition for secure and sustainable decentralized intelligence networks.

cs.GT↗
arXiv2025

Scaling Decentralized Learning with FLock

Fine-tuning the large language models (LLMs) are prevented by the deficiency of centralized control and the massive computing and communication overhead on the decentralized schemes. While the typical standard federated learning (FL) supports data privacy, the central server requirement creates a single point of attack and vulnerability to poisoning attacks. Generalizing the result in this direction to 70B-parameter models in the heterogeneous, trustless environments has turned out to be a huge, yet unbroken bottleneck. This paper introduces FLock, a decentralized framework for secure and efficient collaborative LLM fine-tuning. Integrating a blockchain-based trust layer with economic incentives, FLock replaces the central aggregator with a secure, auditable protocol for cooperation among untrusted parties. We present the first empirical validation of fine-tuning a 70B LLM in a secure, multi-domain, decentralized setting. Our experiments show the FLock framework defends against backdoor poisoning attacks that compromise standard FL optimizers and fosters synergistic knowledge transfer. The resulting models show a >68% reduction in adversarial attack success rates. The global model also demonstrates superior cross-domain generalization, outperforming models trained in isolation on their own specialized data.

cs.LG↗
arXiv2024

Affinity-Graph-Guided Contractive Learning for Pretext-Free Medical Image Segmentation with Minimal Annotation

The combination of semi-supervised learning (SemiSL) and contrastive learning (CL) has been successful in medical image segmentation with limited annotations. However, these works often rely on pretext tasks that lack the specificity required for pixel-level segmentation, and still face overfitting issues due to insufficient supervision signals resulting from too few annotations. Therefore, this paper proposes an affinity-graph-guided semi-supervised contrastive learning framework (Semi-AGCL) by establishing additional affinity-graph-based supervision signals between the student and teacher network, to achieve medical image segmentation with minimal annotations without pretext. The framework first designs an average-patch-entropy-driven inter-patch sampling method, which can provide a robust initial feature space without relying on pretext tasks. Furthermore, the framework designs an affinity-graph-guided loss function, which can improve the quality of the learned representation and the model generalization ability by exploiting the inherent structure of the data, thus mitigating overfitting. Our experiments indicate that with merely 10% of the complete annotation set, our model approaches the accuracy of the fully annotated baseline, manifesting a marginal deviation of only 2.52%. Under the stringent conditions where only 5% of the annotations are employed, our model exhibits a significant enhancement in performance surpassing the second best baseline by 23.09% on the dice metric and achieving an improvement of 26.57% on the notably arduous CRAG and ACDC datasets.

cs.CV↗
arXiv2024

SoccerNet 2024 Challenges Results

The SoccerNet 2024 challenges represent the fourth annual video understanding challenges organized by the SoccerNet team. These challenges aim to advance research across multiple themes in football, including broadcast video understanding, field understanding, and player understanding. This year, the challenges encompass four vision-based tasks. (1) Ball Action Spotting, focusing on precisely localizing when and which soccer actions related to the ball occur, (2) Dense Video Captioning, focusing on describing the broadcast with natural language and anchored timestamps, (3) Multi-View Foul Recognition, a novel task focusing on analyzing multiple viewpoints of a potential foul incident to classify whether a foul occurred and assess its severity, (4) Game State Reconstruction, another novel task focusing on reconstructing the game state from broadcast videos onto a 2D top-view map of the field. Detailed information about the tasks, challenges, and leaderboards can be found at https://www.soccer-net.org, with baselines and development kits available at https://github.com/SoccerNet.

cs.CV↗
arXiv2024

Does DetectGPT Fully Utilize Perturbation? Bridging Selective Perturbation to Fine-tuned Contrastive Learning Detector would be Better

The burgeoning generative capabilities of large language models (LLMs) have raised growing concerns about abuse, demanding automatic machine-generated text detectors. DetectGPT, a zero-shot metric-based detector, first introduces perturbation and shows great performance improvement. However, in DetectGPT, the random perturbation strategy could introduce noise, and logit regression depends on the threshold, harming the generalizability and applicability of individual or small-batch inputs. Hence, we propose a novel fine-tuned detector, Pecola, bridging metric-based and fine-tuned methods by contrastive learning on selective perturbation. Selective strategy retains important tokens during perturbation and weights for multi-pair contrastive learning. The experiments show that Pecola outperforms the state-of-the-art (SOTA) by 1.20% in accuracy on average on four public datasets. And we further analyze the effectiveness, robustness, and generalization of the method.

cs.CL↗