Searcharxiv⌕ Search

arXiv subjects

Delu Zeng

Publications and source records attributed to Delu Zeng.

At least 19 recordsLinked to original sources

A Think Block Is Conditioning, Not a Plan

Generative models increasingly write a chain of thought before drawing, and report large gains from doing so. We ask what that text contributes, by substitution rather than ablation: holding the generator, sampler and seeds fixed, we replace the think block with controlled alternatives and measure the image. Across a unified model with a think mode, a reinforcement-learned reasoning generator, and a planner driving a frozen diffusion model, the same picture emerges. Content-free prose of matched length is worth nothing or less ($-0.05$ to $-0.15$). One sentence from an unrelated 8B model, naming the subject and a few identifying details, matches the generator's own plan in three to six times fewer conditioning tokens. The value sits where the referent is in doubt: writing down a name is worth $+0.33$ when it is right and $+0.05$ when it is wrong, and where the generator already succeeds the whole effect is $-0.07 \pm 0.07$. Cutting a plan at the text encoder's window leaves 272 of 286 images byte-identical, and a plan rewritten to fit that window ties a bare identification ($+0.003 \pm 0.044$), so a plan contains nothing the identification does not already supply. Field-level interventions find only colour executed reliably, and a control writing the same sentences into the user prompt shows that limit belongs to the generator rather than to the planning channel. Two judges and two annotators replicate the central contrasts. A think block earns its keep as a specification of what to draw, not as a plan that is carried out.

cs.CV↗

Trajectory Density Ratios under Weak Support: A Survey across the Sciences

Many problems across the sciences compare two stochastic processes through the sequences they generate. Reinforcement learning asks what a new policy would return from the trajectories of an old one; molecular simulation asks how much more a ligand binds from samples of the unbound state; climate attribution asks how a heat wave shifted under warming; and language-model post-training asks what a deployed agent will do from logged interactions. In every case, one quantity would answer all these questions at once: the ratio of the two probability laws over their generated sequences, the trajectory density ratio. This ratio exists only under strong conditions. When one process reaches states the other cannot, the ratio does not exist; when their overlap is thin, its variance grows without control with the sequence length; and when they concentrate on different lower-dimensional sets, the ratio is undefined altogether. Weak support of this kind is the norm for sequence data, not a sampling accident, and its price is paid in identifiability rather than in noise alone. This survey organizes the field around a three-level diagnosis of weak support and a map of the nine retreats that methods take when the full ratio is out of reach, from direct estimation and reweighting through staged decomposition, stationarization, and partial-identification bounds to smoothing, transport couplings, operator spectra, and direct estimation of a single functional. Each retreat keeps part of what the ratio offers and pays for the rest, and each application domain has standardized the retreat its data force. The survey closes with open problems: testable diagnostics at every level, partial identification at the trajectory level, a rate theory of time-amplified error, and benchmarks with ground truth.

stat.ML↗

One-Step Score-Based Density Ratio Estimation

Density ratio estimation (DRE) is a useful tool for quantifying discrepancies between probability distributions, but existing approaches often involve a trade-off between estimation quality and computational efficiency. Classical direct DRE methods are usually efficient at inference time, yet their performance can seriously deteriorate when the discrepancy between distributions is large. In contrast, score-based DRE methods often yield more accurate estimates in such settings, but they typically require considerable repeated function evaluations and numerical integration. We propose One-step Score-based Density Ratio Estimation (OS-DRE), a partly analytic and solver-free framework designed to combine these complementary advantages. OS-DRE decomposes the time score into spatial and temporal components, representing the latter with an analytic radial basis function (RBF) frame. This formulation converts the otherwise intractable temporal integral into a closed-form weighted sum, thereby removing the need for numerical solvers and enabling DRE with only one function evaluation. We further analyze approximation conditions for the analytic frame, and establish approximation error bounds for both finitely and infinitely smooth temporal kernels, grounding the framework in existing approximation theory. Experiments across density estimation, continual Kullback-Leibler and mutual information estimation, and near out-of-distribution detection demonstrate that OS-DRE offers a favorable balance between estimation quality and inference efficiency.

stat.ML↗

RubricOpt: Self-Revision of Semantic Rubrics for Preference Optimization

A model's failures can reveal gaps in the behavioral criteria guiding its training. We introduce RubricOpt, a framework that lets a policy revise these criteria within preference optimization. The policy diagnoses its own rubric-free outputs, retains requirements already covered by the rubric, and edits clauses that leave recurring failures unaddressed. A shared semantic representation connects task demands, preference-data contrasts, and rubric clauses, while a deterministic gate checks evidence references and fixed semantic anchors. The rubric then guides Direct Preference Optimization (DPO) during training, with preference pairs, labels, and the objective held fixed, and is removed at evaluation. Our analysis relates the local effect on rubric-free success to semantic alignment and a residual beyond the measured factors. Experiments on capability and open-ended generation benchmarks show that rubric conditioning can improve outcomes over ordinary DPO, with gains and regressions varying across models and datasets. A matched continuation isolates the contribution of revision from additional training and finds mixed incremental effects for the tested edit, whose feedback overlaps the evaluation set. RubricOpt thus brings the behavioral specification itself into the policy's feedback loop.

cs.AI↗

Stochastic Pruning of the DPO Computation Graph: 1.4x Faster Training and Lower Peak Memory

Direct preference optimization (DPO) incurs training costs in output scoring, sequence execution, and backpropagation, so reducing supervision alone does not necessarily save computation. We introduce a unified framework for stochastic pruning of the DPO computation graph that independently samples response positions for scoring, response prefixes for execution, and preference branches for backpropagation. By making these selections before the corresponding operations, the framework omits computation at each stage of a training step. Our analysis characterizes the objective changes induced by position and sequence pruning and uses inverse-probability weighting to obtain a conditionally unbiased single-branch gradient estimator for the sampled objective. Controlled measurements show that position and sequence pruning primarily reduce peak memory, with most output-memory savings attributable to an implementation that retains full response supervision. Branch pruning yields a $1.426\times$ step-time speedup over a concatenated DPO baseline and reduces training time by about 30\% relative to stock DPO across two model families, with slightly lower mean accuracy and benchmark-dependent changes in open-ended scores. These findings relate the resource benefits of stochastic graph pruning to the computation omitted and the resulting changes in the training objective.

cs.LG↗

Hindsight-Privileged RLVR: Single-Rollout Token Credit Without Group Sampling

Reinforcement learning with verifiable rewards (RLVR) trains language models from checked task outcomes, but assigning a terminal reward to individual reasoning tokens or actions remains difficult. Group-relative methods such as Group Relative Policy Optimization (GRPO) sample several attempts per prompt, but lose their reward-based learning signal when all attempts receive the same outcome. We propose Hindsight-Privileged RLVR (HP-RLVR), a single-rollout method that turns a completed trajectory's own future into token-level update weights. The generating policy snapshot scores each eligible token under its causal history and under a hindsight context containing the realized future and verifier feedback. A normalized Jensen--Shannon divergence between the two views weights the reward relative to a historical baseline, yielding bounded token credit for both successful and failed attempts, with no reference solution or learned critic. In paired Qwen3-8B runs over a shared $100$-step schedule, HP-RLVR trains $1.21\times$, $1.46\times$, and $2.30\times$ faster than GRPO on mathematics, code generation, and ALFWorld, using one eighth the rollouts per prompt. Accuracy is comparable overall, with HP-RLVR ahead on mathematics at all three sizes and GRPO ahead on code and ALFWorld. In ALFWorld, $43.2\%$ (4B) and $49.0\%$ (8B) of GRPO's gradient-norm records are zero, versus none for HP-RLVR.

cs.LG↗

Neural Bridge Processes

Learning stochastic functions from partially observed context-target pairs requires models that are expressive, uncertainty-aware, and strongly conditioned on inputs. Neural Diffusion Processes (NDPs) improve expressivity with denoising diffusion, but their forward process is input-independent; inputs only enter the reverse denoiser, so the noisy training states themselves do not encode the conditioning inputs. We propose Neural Bridge Processes (NBPs), which replace the unconditional forward kernel with an input-anchored bridge trajectory. When input and output dimensions differ, NBP learns an output-space anchor $a_ψ(x)=P_ψ(x)$, allowing coordinates or other inputs to guide the generative path without changing the denoising backbone. We show theoretically that process-level anchoring induces pathwise input distinguishability, injects information about x into noisy states, and creates a direct gradient pathway unavailable to NDPs. Experiments on synthetic regression, EEG, CylinderFlow, and image regression show consistent improvements. Additional ablations show that the gains come from the full bridge construction with learned alignment, and that the same input-anchored path principle transfers to Flow Matching Neural Processes. These results suggest that bridge-anchored generative paths provide a general mechanism for strengthening conditional stochastic function modeling.

cs.LG↗

Consist-Retinex: One-Step Noise-Emphasized Consistency Training Accelerates High-Quality Retinex Enhancement

Retinex-based low-light image enhancement benefits from separating reflectance and illumination, yet recent generative approaches often rely on iterative sampling and are difficult to deploy under strict latency budgets. Consistency models offer a natural route to one-step restoration, but direct adaptation to Retinex-factorized enhancement is unstable: one-step inference is evaluated at the high-noise endpoint, whereas standard training schedules provide little supervision there, and temporal self-consistency alone does not determine the correct conditional target. We propose Consist-Retinex, which first uses a Retinex Transformer Decomposition Network (TDN) to obtain paired reflectance and illumination maps, then trains two conditional consistency models with a Retinex-aware dual objective and adaptive noise-emphasized fixed-point sampling. The dual objective combines trajectory consistency with paired ground-truth component alignment, while the sampling rule concentrates supervision near the inference endpoint without discarding full-range noise coverage. We further provide an endpoint error bound, an anchoring-propagation result, and a high-noise sample-allocation analysis that explain why endpoint supervision and temporal consistency are complementary for one-step Retinex enhancement. Experiments on paired and unpaired low-light benchmarks show that Consist-Retinex obtains the best VE-LOL-L scores among the compared methods under one-step inference and remains competitive on LOL, with substantially reduced sampling and consistency-stage training cost in the reported setup.

cs.CV↗

Intervention, Not Shared Latents: Blocking Visual Shortcuts in Audio-Video Generation

Joint audio--video (AV) generators are trained on data in which \emph{what an event looks like} and \emph{what it sounds like} are spuriously correlated. We present a \emph{controlled causal study} of the resulting failure mode. In an AV structural causal model where the audio is, by construction, independent of the video's nuisance appearance, models that let audio read video directly---through cross-attention or a shared latent---learn a \emph{visual shortcut}: they predict sound from appearance rather than the causal event and, when the appearance--event correlation is broken at test time, synthesize the wrong event's sound. Crucially, the popular remedy of routing both modalities through a \emph{shared common-cause latent} does \emph{not} fix this---a bottleneck, an unsupervised shared/private factorization, and a faithful shared-prior model all grab the appearance proxy and fail like the direct model. Blocking the shortcut instead requires an \emph{intervention on the nuisance}: under the stated assumptions we prove that counterfactual invariance is necessary and sufficient to identify the causal predictor, and we verify the mechanism from feature-vector SCMs to procedural pixel video, real images with spectrogram audio, moving real digits, and a conditional generator. On a \emph{real, pretrained} V2A generator (MMAudio), an input-intervention test shows the model is far from invariant to sound-irrelevant edits, though a generic-noise control reveals it is broadly input-brittle rather than specifically colour-shortcutting---clean isolation of the shortcut needs the controlled confounds our synthetic studies provide. We characterize \emph{when} the shortcut occurs, compare the objective against supervised counterfactual augmentation, and isolate the \emph{unknown-nuisance} regime---where the intervention cannot be applied---as the central open problem.

cs.LG↗

Partially Observed Sparse Graphs: The Unknown Sampling Rate is a Tail Index

A large graph is often available only in part: a crawl stopped by its budget, a panel, a partial dump. When the sampled fraction $s$ is known by design the total edge count follows from $\hat e=e_s/s^2$ and no model is needed. We treat the case where $s$ is unknown and the population size is known. Our main result is a reduction: under a sparse exchangeable (graphex) model the expected non-isolated fraction obeys $n_s/n_1\to s^{1+σ}$, so the sampling rate becomes estimable once the tail index $σ$ is, and substituting it back gives $e_s(n_1/n_s)^{2/(1+σ)}$ -- the same estimator, with the design quantity inferred. Estimating global edge cardinality in a sparse graph is therefore, in expectation, tail-index estimation, and the quadratic graphon estimator is the case $σ=0$: it fails by an identity rather than by a fit ($260\%$ median error against $27\%$). We bound the finite-size error of the substitution and show the reduction is \emph{modular} in the tail-index estimator --- filled with a published closed-form one it reaches $21.7\%$ over $13$ networks and $39$ sampling budgets with no fitting at all. Fitting a full graphex additionally returns the degree distribution at any size and a generative object, in a representation where sparsity is a coordinate and the interpolation path is dictated rather than chosen. Two limits are exact: rank-one graphexes have transitivity fixed by the degree profile, so high-clustering graphs lie outside the class; and under snowball or random-walk crawls every method here fails, the design-based oracle worst of all ($7.8\%$ to $588\%$).

cs.LG↗

Local-Time Riemannian Score Matching on the Quantum Pure-State Manifold

Score-based diffusion can be defined intrinsically on the manifold of quantum pure states, $\mathbb{CP}^{d-1}$ with the Fubini--Study metric, but no closed-form transition density is available, so the score must be supervised by a local-time teacher taken from the Euclidean limit of the diffusion in normal coordinates. This paper is about what makes that teacher work, and where it stops working. Three training choices turn out not to be incidental: the increment must be divided by the diffusion clock rather than by the elapsed time, since the published expression assumes unit diffusion and a non-unit schedule introduces a time-change mismatch varying by a factor of $400$ across the horizon; the logarithm and exponential maps should be the closed-form Fubini--Study ones, which is the largest single effect we measure; and the global phase must be randomised, because horizontal projection alone does not make a score network descend to the quotient. With these choices the model improves on the published Riemannian local-time baseline in every cell of an eight-benchmark, four-metric comparison over ten seeds, significantly on five of eight after Holm correction, and beats an ambient Euclidean baseline by an order of magnitude everywhere. We then bound what the approximation costs by replacing it with the exact heat kernel of $\mathbb{CP}^{d-1}$, computable up to complex dimension seven, where the local-time teacher loses a factor of $1.2$ to $2.7$.

stat.ML↗

Diagnosing Under-Development of Irreversible Processes in Video Generation

Many physical attributes are \emph{irreversible}: ice melts but does not re-freeze, paper chars but does not un-burn. Do video generators respect this? We show the question is hard to measure, and that what can be measured reliably is \emph{development} rather than reversal. Metrics of local reversal are null-degenerate: a per-clip violation rate scores $0.50$ on pure noise, and a variance-normalized reversal residual sits at its noise ceiling. What survives null-testing is a two-part protocol: progress (a directional attribute correlation) and a stasis rate. Under this protocol, generated video separates cleanly from real footage, and the gap is human-validated. Across seven text-to-video models, real reference footage advances ($ρ{=}{+}0.40$, $35\%$ static) while every generator shows near-zero progress and $92$--$100\%$ stasis; nine annotators rate real footage far above generated ($2.75$ vs.\ $0.99$ on a $0$--$4$ scale). The reliable finding is \emph{under-development}: generators barely advance irreversible attributes rather than reversing them. As a complementary mechanism, we show that post-hoc readout guidance is gameable, whereas enforcing monotonicity by construction in a disentangled attribute latent removes the gameable readout, validated in controlled and semi-synthetic settings.

cs.CV↗

Towards Disentangled Preference Optimization Dynamics: Suppress the Loser, Preserve the Winner

Preference optimization is widely used to align large language models (LLMs) with human preferences. However, many margin-based methods also suppress the chosen response when they try to suppress the rejected one, and there is no general way to prevent this across different objectives. We address this issue with a unified incentive-score decomposition of preference optimization, revealing that different objectives share the same local update directions and differ only in their scalar weights. This decomposition provides a common framework for analyzing objectives that were previously studied in separate settings. Building on this decomposition, by analyzing the dynamics of the chosen/rejected likelihoods, we identify the disentanglement band (DB), a simple, testable condition that tells us when training can follow the desired path: suppress the loser while preserving the winner, possibly after an early stage. Using the DB, we propose reward calibration (RC), a plug-and-play method that adaptively rebalances the updates for chosen and rejected responses to satisfy the DB, without redesigning the base objective. Empirical results show that RC leads to more disentangled dynamics, with better downstream performance observed across several settings. Our code is available at https://github.com/IceyWuu/DisentangledPreferenceOptimization.

cs.LG↗

Look Again Before You Abstain:Budgeted Conformal Evidence Acquisition for Reliable Vision-Language Model

Large vision-language models (LVLMs) hallucinate: they assert visual details that the image does not support. A principled remedy is selective prediction with a distribution-free guarantee-verify each claim and abstain when the claim is not grounded, so that the hallucination rate among asserted claims is provably bounded. We show, however, that this guarantee is bought at a brutal price: to keep the hallucination rate below $5\%$ on a balanced object-existence benchmark, a state-of-the-art conformal filter must abstain on more than $80\%$ of claims. We argue that abstention is wasteful when more visual evidence is cheaply available, and introduce Budgeted Conformal Evidence Acquisition (BCEA), which replaces the binary answer/abstain decision with a three-way choice: answer, abstain, or acquire additional visual evidence by re-examining the image (zooming, cropping, or applying a claim-specific intervention) under a bounded

cs.CV↗

DiFA: Inference-Time Forward-Process Alignment for Diffusion Models

The prevailing inference framework for diffusion models formulates generation fundamentally as a problem of numerical integration. This perspective casts the model as an exact estimator, neglecting the inherent statistical uncertainty of the denoising process. In this work, we propose Forward-Process Aligned Diffusion prediction (\textbf{DiFA}), a training-free framework that reframes inference-time data prediction refinement as a sequential state estimation problem. Rather than reusing past outputs solely for numerical integration, DiFA treats iterative data predictions along the reverse trajectory as correlated observations to build a forward-aligned temporal consensus. Inspired by Kalman filtering, this consensus aggregates historical predictions according to structural consistency and noise-level compatibility. To counteract the over-smoothing tendency of temporal consensus, we introduce a deviation guidance mechanism to adaptively preserve residual details. Empirically, DiFA yields significant improvements on CIFAR-10 and ImageNet across the evaluated metrics, including FID, IS, and FD-DINOv2, demonstrating that aligning inference with the forward statistical structure substantially improves generative fidelity.

cs.LG↗

When Can You Debias an LLM Judge? Identifiability Limits, a Test, and Designs for Top-k Ranking

Large language models (LLMs) are increasingly used as cheap, scalable judges that compare candidate outputs pairwise. Because such judges prefer verbose or well-formatted answers, the natural fix is to add bias covariates to a Bradley--Terry model and estimate the bias away. We show this cannot work as advertised: the quality/bias split is \emph{not identified} by pairwise comparisons, and the failure is exact -- across $48$ real judge-pools the profile likelihood over the coefficient is flat to $\mathbf{0.0000}$ \textbf{nats}, and scaling the comparisons $26\times$ buys none. A ``debiased'' score is selected by the prior, not recovered from data. Our contribution is accordingly not a better estimator but a characterization of \emph{when prior-based correction is justified}, plus designs that supply the missing information when it is not. The assumption the prior encodes -- quality is a priori uncorrelated with the covariate -- pays only while $\mathrm{corr}(θ,x)$ stays below a crossing point (configuration-dependent, $0.22$--$0.60$), which is what makes the same model help on LLMBar and hurt on SummEval and Nectar. We give two escapes: a \textbf{trusted-anchor gate} that decides per (judge, covariate, task) (no false enables in $6{,}000$ decisions at $K\ge10$ anchors, a rate our sample bounds at $\le6\%$), and a \textbf{paired rendering design}. Across fifteen real LLM judges bias is heterogeneous and capability-dependent: correction improves \topk{} recall by $0.20$--$0.32$ on five biased-but-competent cheap judges and is a no-op on frontier ones (Spearman $ρ{=}{-}0.84$ between competence and gain over the $14$ competent judges, $p{<}10^{-3}$), concentrating the benefit where at-scale evaluation happens.

cs.LG↗

AQKA: Active Quantum Kernel Acquisition Under a Shot Budget

Estimating an $N \times N$ quantum kernel from circuit fidelities requires $Θ(N^2 S)$ measurement shots, the dominant bottleneck for deployment on near-term hardware. Existing budget-saving methods (Nyström-QKE, ShoFaR, kernel-target alignment) sub-sample \emph{which} entries to measure but allocate shots \emph{uniformly} within their chosen subset, ignoring how much each entry drives the downstream classifier. We close this gap with two contributions. \textbf{First, a complete regime decomposition} for shot-budgeted quantum kernel learning: a principled menu of when each allocator wins. Our method, \emph{AQKA}, dominates the budget-limited regime ($B \lesssim 16 n_{\mathrm{pairs}}$) on sparse-sensitivity KRR, with the gap \emph{growing} from $+8$ to $+25$ pts over uniform as $N$ scales $225{\to}1000$ and reaching $+26$--$32$ pts on an \texttt{ibm\_pittsburgh} (156-qubit Heron) hardware kernel; Nyström-QKE wins at saturating budgets on planted-sparse via low-rank reconstruction; ShoFaR is competitive only at extreme low budgets. \textbf{Second, a closed-form pair-level acquisition theory}: $s_{ij}^{\star} \propto |g_{ij}|\sqrt{K_{ij}(1-K_{ij})}$ with explicit gradient $g_{ij}$ for KRR (Lemma~1, $|β_iα_j+β_jα_i|\sqrt{K_{ij}(1-K_{ij})}$) and SVM via the envelope theorem ($|η_i^*η_j^*|\sqrt{K_{ij}(1-K_{ij})}$); a \emph{corrected} sparsity-aware Cauchy--Schwarz rate $ρ\le 2m/N$ matching empirics (vs.\ the naive $m^2/N^2$); an explicit-constant plug-in regret bound (Theorem~2); and a tighter SVM ceiling $ρ^{\mathrm{SVM}} \le m_{\mathrm{sv}}^2/N^2$. We close with the first multi-seed live online adaptive shot allocation on quantum hardware: $+17.0 \pm 4.8$ pts at $N{=}20$ on \texttt{ibm\_aachen} ($3.5σ$, 5 seeds), with the advantage holding at $N{=}30$ at higher budget on \texttt{ibm\_berlin} ($+14.0 \pm 8.5$ pts, 5 seeds).

cs.LG↗

Fixed-Protocol Amortized MPS Tomography with Conformalized Predictive Uncertainty

Quantum state tomography is sample-starved, and the states one prepares live on a narrow, learnable manifold. A $k{=}0$ prior-only control shows that on concentrated families a prior estimate is already near-optimal, so ``high fidelity at few measurements'' can be family memorization rather than tomography; genuine measurement-efficiency needs a model that conditions on the measurements and demonstrably uses them. On a shared matrix-product-state (MPS) core parameterization we study two routes. Approach~A learns a generative prior over MPS cores with measurement-guided posterior inference (gold-standard-validated, but whose few-measurement accuracy the control shows is largely the prior). Approach~B, our main proposal, is a \emph{fixed-protocol amortized} MPS estimator trained once with a gauge-invariant fidelity loss; we deliberately do not rest it on a permutation-invariant set encoder (a plain MLP matches it). The decisive lever is the measurement design: motivated by the fact that local reduced density matrices determine a $χ$-MPS, conditioning on an \emph{informative local} Pauli set rather than random strings turns a modest, memorization-prone estimator into a high-fidelity one ($\approx\!0.95$, up to $+0.59$ over prior-only, decisively passing a shuffled-measurement control). A dropout ensemble, conformally recalibrated, gives $\approx\!90\%$-coverage intervals -- including for observables never measured, where a shot-based interval does not exist. Quality holds as the system grows (fidelity $0.90$ at $n{=}10$, gain \emph{growing} in $n$; $0.88$ at bond dimension $χ{=}4$), the parameterization is polynomial (native contraction to $20$ qubits), and we close the loop on IBM hardware ($5$ states at $0.97$ from hardware-measured Paulis).

quant-ph↗