arXiv subjects

Hanzhe Hong

Publications and source records attributed to Hanzhe Hong.

6 recordsLinked to original sources

arXiv2026

Teaching Vision-Language Models to Use the Scale They Are Given: Label-Free Equivariance Training for Metric Physical Reasoning

Metric questions about video require vision-language models to use supplied real-world references to convert visual measurements into physical units. Yet we find that current models use this scale information only partially. When every world-space quantity in a prompt is rescaled by a common factor, the video remains equally valid and the correct answer changes by exactly that factor, but model predictions move only part of the way and accuracy remains concentrated near the familiar scale of the depicted objects. Across eight vision-language models, this under-response persists over four orders of magnitude. The same models recover the correct closed-form scaling laws when the identical physics is asked in a scale-free form, indicating that the main deficit lies in metric grounding rather than physical mechanism knowledge. We use this exact scaling relation as supervision without requiring metric annotations. Under a common rescaling of the supplied world-space quantities, the correct metric answer must change by the same factor. EquiSD exploits this constraint by projecting a model's own prediction onto the scale-equivariant family and fine-tuning the model on the resulting targets. It requires no ground-truth answers and only one model query per training video. On held-out simulated videos, EquiSD increases a 3B model's median response slope from 0.66 to 0.94 and improves mean relative accuracy by 9.2 points across scales. The learned relation generalizes to unseen world scales and transfers without adaptation to real QuantiPhy videos, where accuracy increases by 6.4 points. These results show that an exact physical symmetry can provide label-free supervision for improving metric grounding in vision-language models.

cs.CV↗

arXiv2026

Beyond Multimodal Alignment: Certifying Physical Language through Response Substitution and Ordered Execution

World models increasingly treat compact multimodal representations as interfaces between perception and physical interaction, yet existing probes do not establish whether different sensors carry the same executable meaning or whether that meaning survives a new action composition. We introduce an operational capability hierarchy and the Disjoint-Bridge Operator-Substitution Certificate (DBOSC), which asks whether independently trained modality compilers enter a frozen response chart interchangeably on evidence outside their training panels. On Cluster Haptic, audio and acceleration representations of the same unseen surface are 4.5x closer in response space than wrong-surface pairings, with the gap holding for all 19 held-out surfaces; unsealing withheld responses confirms that every branch predicts the physics better than the population chart. We then test ordered execution in a controlled elastoplastic system with complementary modality blind spots. At the pre-registered budget, the prerequisite refuses the stack because the frozen executor cannot advance even an exact chart coordinate through a held-out program. At a converged budget, the same rank-three chart executes those programs (oracle NMSE 0.18), fusion improves on both modalities, and 14 of 16 registered checks pass; the two failures arise because a diagonal restriction of the fused information matrix performs as well as the full one. Clearing the gate is a property of the executor, not the chart: an executor emitting whole programs instead of shared per-step dynamics is 38x worse than an entity-blind predictor on the same chart. A matching non-identifiability result explains why compression and fusion alone cannot determine an unseen composition law. These results separate attribute access, response substitution, fusion closure, and ordered execution into distinct, separately testable achievements.

cs.LG↗

arXiv2026

Collusion with Competitive Marginals: Price-Level Audits Are Blind by Construction

Empirical work on algorithmic collusion asks one question of the data: are prices supracompetitive? We show this can be answered "no" by a conspiracy that is nonetheless profitable. Consider bidding agents that couple only through the joint distribution of their unexplained bid components, leaving every agent's own bid law exactly at the competitive law. Any test whose input is a single agent's price or bid history then has power exactly equal to its false-positive rate, for every coupling strength up to comonotonicity. The published detection methodology is therefore blind to this conduct by construction rather than underpowered, and no sample size repairs it. Three empirical results follow. First, the mechanism appears in real language-model agents: twenty models from nineteen independent developers, three deployment prompts each, show residual correlation of $+0.053$ between two deployments of one model against $+0.0001$ across models, with a 95% interval clustered by developer of $[0.030, 0.078]$, under an auditor that sees every order feature and is fitted out of sample. Second, the coupling falls monotonically as sampling temperature rises ($p=0.002$), turning a deployment parameter into a candidate mitigation. Third, on 24 days of Ethereum block-building auction data covering 77,684 bids from 39 bidders, the honest population of bidder pairs is itself so dependent that a screen held at a 5% false-positive rate must sit above a floor of $+0.50$ to $+0.81$, which is 20 to 32 times the family-wise sampling threshold and does not fall as the audit window grows. Since lawful multi-identity operation and conspiracy are behaviourally indistinguishable here, the tractable regulatory target is not detection but counting: resolving 40 bidding identities into 23 operators raises the Herfindahl index by 247.5%, and adding behavioural clusters from public bid streams reaches 324.5%.

cs.GT↗

arXiv2026

What Can Latent World Models Know? Physical Parameter Identifiability in Multimodal Predictive Representations

A central premise of latent world models is that predicting the future forces a representation to internalize the physics of its environment. Which physical quantities does a trained latent actually contain, and what decides this? We answer with controlled interventions in POKEWORLD, an interactive environment whose visually identical objects hide mass, drag, and contact stiffness. A certificate-gated protocol first certifies each parameter as recoverable from raw observations, then measures whether it enters the latent, so a null result can be attributed to the objective rather than to the environment. The resulting identifiability map has two organizing mechanisms and one frontier. Inputs limit what can be known, while prediction targets decide what is retained. Stiffness enters the latent only when touch is forecast ($R^2=0.50$, compared with $-0.02$ when the same signal is merely fused into the input), and under single-step prediction a vision-only latent discards even perfectly visible object state. Drag marks the frontier. It carries a recoverability certificate of 0.89 yet plateaus near 0.13 under every deterministic prediction objective we test, while a supervised head on the same trunk reaches 0.45. Parameters whose readout is slow and ratio-type under the sensed coordinates fall outside what these objectives acquire. On RH20T, an input-target factorial across scaling curves reproduces both mechanisms across two robots and 4,258 episodes. Every arm missing information or prediction pressure stays flat over a fivefold data range, and only the full multimodal objective forecasts force beyond a persistence baseline, with held-out gains that grow with scale. Objective structure determines which physical parameters a latent acquires, and additional data improves only the parameters it already acquires.

cs.LG↗

arXiv2026

GeoFidelity-Bench: Evaluating Segment-Level Geographic Fidelity in Text-to-Image Street-View Generation

Text-to-image models can generate visually plausible city streets, but whether their outputs correspond to a requested road segment rather than a generic city prior remains unclear. We introduce GeoFidelity-Bench, a reference-panel benchmark for segment-conditioned geographic fidelity in street-view generation. It contains 7,117 curated Mapillary images covering 109 named OpenStreetMap road segments in 25 cities across six continents. For each generated panel, the benchmark ranks the target reference panel against panels from the nearest segment in the same city, other segments in the same city, and segments from other cities, making local discrimination rather than absolute target similarity the primary test. We evaluate six open-weight text-to-image generators under city-only, street-and-neighborhood, and GPS-augmented prompts. Adding street and neighborhood names is associated with an increase of 5.5 percentage points in top-1 retrieval accuracy over city-only prompts, with a 95% confidence interval from 3.4 to 7.7 percentage points. However, the similarity margin between the target and the nearest segment in the same city remains near zero, indicating that local names improve broad local plausibility more than exact segment identity. Prompts that keep the city fixed but use incorrect street or neighborhood names further show that only part of the gain depends on the correct local names, while appending raw GPS coordinates as ordinary text yields no statistically clear additional benefit. Held-out real-image queries successfully recover segment identity, showing that the curated references contain usable segment-level signal. GeoFidelity-Bench thus reveals a persistent gap between city- or neighborhood-plausible street-view generation and faithful generation for a specific road segment.

cs.CV↗

arXiv2026

When Does Visual Token Pruning Improve Calibration? The Role of Evidence Coverage in MLLMs

Visual token pruning is widely used to reduce the inference cost of multimodal large language models (MLLMs), but it is usually evaluated only by accuracy. We study how pruning affects calibration, defined as the agreement between confidence and correctness, and show that the selection rule matters more than the token budget alone. On POPE with LLaVA-1.5, coverage-based pruning from 576 to 128 tokens reduces expected calibration error from 0.041 to 0.016 without a statistically significant accuracy loss. In contrast, attention-based selection preserves confidence while accuracy deteriorates, becoming less calibrated than random pruning at aggressive budgets. Across pruning conditions, kept-set coverage is strongly associated with accuracy (Spearman $\rho=+0.89$) but not with mean confidence ($\rho=-0.03$), producing a strong inverse relation with overconfidence ($\rho=-0.92$); controlled kept-set interventions support this evidence-coverage account. Two boundaries limit it: query-conditioned FastV is more overconfident than its coverage predicts, and on language-prior-dominated ScienceQA, coverage ceases to order calibration. The selector ordering otherwise generalizes to GQA and LLaVA-NeXT, and coverage beats random on Qwen2-VL, although the calibration gain over the unpruned model is task- and model-dependent. We also identify an evaluation pitfall in FastV: zeroing rather than removing pruned tokens can reduce accuracy to chance. Visual token pruning therefore changes confidence quality as well as efficiency, and calibration should be evaluated alongside accuracy when comparing pruning methods.

cs.CV↗