arXiv subjects

Guojun Zhu

Publications and source records attributed to Guojun Zhu.

8 recordsLinked to original sources

arXiv2026

Bad Genius: Counterfactual-Guided Harness Evolution Beyond Task-Specific Shortcuts

Reliable agent evaluation is complicated by automatic harness optimization, which repeatedly uses a released benchmark $B_{\mathrm{rel}}$ to guide a Proposer that edits prompts, memory, retrieval, tools, and control code around a fixed target agent. Task holdout varies semantic tasks but leaves the benchmark protocol fixed, so a "bad genius" Proposer can produce a cheating harness whose released-benchmark gain depends on a benchmark-wide shortcut. We introduce Counterfactual Harness Search and Evolution (CHASE), which casts harness evolution as constraint generation over validity-preserving benchmark counterfactuals. After each Proposer update, a Challenger searches for an executable protocol transformation with large gain destruction. A validity firewall checks that task semantics are preserved, while a confirmation set determines whether the counterfactual enters a finite archive. We formalize an exact shortcut-neutralized benchmark $B_0$ and establish statistical guarantees linking finite counterfactual archives to $B_0$ and characterizing sequential Challenger search. We evaluate CHASE on a synthetic benchmark and on OfficeQA, where CHASE retains strong released-benchmark gains while substantially reducing gain destruction under valid protocol changes.

cs.AI↗

arXiv2026

Pattern-Calibrated Multimodal Prediction under Blockwise Missingness

Blockwise missingness in multimodal data is usually treated as an incomplete-input problem. We instead focus on prediction for a prespecified observed-modality pattern, where the observed modality set determines the information on which the prediction rule can condition. A procedure that imputes missing modalities, zero-fills unobserved modalities, or trains a single pooled predictor may borrow information across patterns, but it can also mix pattern-specific prediction rules. We propose Multimodal Overlap-aware Shared-specific Alignment and Inter-pattern Calibration (MOSAIC), a pattern-calibrated framework for borrowing across missingness patterns without collapsing their prediction rules. MOSAIC learns shared and modality-specific representations, uses the available representations that overlap with the target pattern to fit a first-stage predictor, and then estimates the calibration gap from target-pattern data. We establish non-asymptotic bounds that decompose the error into overlap effective sample size, calibration gap, and representation-learning error, clarifying when cross-pattern borrowing improves over local fitting and when the improvement is controlled by rule mismatch or representation-learning error. Simulations examine representation recovery and target-pattern correction, and applications to ICU mortality prediction, emotion recognition, and glaucoma classification show gains when target-pattern samples are limited or pattern-specific rules differ.

stat.ME↗

arXiv2026

Chaos Is a LADDER: Domain Generalization Beyond Invariance via Reweighting

Domain generalization (DG) aims to learn from multiple source domains and generalize to unseen target domains. Most DG methods pursue invariance: they seek a causal representation whose prediction rule is invariant across domains. This principle is effective when the causal mechanism is stable, but becomes restrictive when the domain itself modulates how causal content maps to the response. In this case, directly feeding domain style into the predictor can create misleading shortcuts, since style does not by itself cause the response. Yet the apparent chaos of multiple styles can become a ladder: style can locate the unseen target domain among source domains and guide which domain-dependent prediction rules should be trusted. We propose \emph{Latent Adaptive Domain Disentanglement and Environment Reweighting} (LADDER), a fixed-model DG pipeline that learns causal/style representations, freezes the encoders, fits source-specific classifiers, and uses an unlabeled target-domain covariate set only at inference to compute weights over these fixed classifiers, with no target labels or model-state updates. We establish theoretical guarantees for source reweighting and validate LADDER on simulations, FMoW, and a location-grouped iWildCam protocol, with gains in overall and group-averaged accuracy.

stat.ML↗

arXiv2026

Enhancing Federated Class-Incremental Learning via Spatial-Temporal Statistics Aggregation

Federated Class-Incremental Learning (FCIL) enables Class-Incremental Learning (CIL) from distributed data. Existing FCIL methods typically integrate old knowledge preservation into local client training. However, these methods cannot avoid spatial-temporal client drift caused by data heterogeneity and often incur significant computational and communication overhead, limiting practical deployment. To address these challenges simultaneously, we propose a novel approach, Spatial-Temporal Statistics Aggregation (STSA), which provides a unified framework to aggregate feature statistics both spatially (across clients) and temporally (across stages). The aggregated feature statistics are unaffected by data heterogeneity and can be used to update the classifier in closed form at each stage. Additionally, we introduce STSA-E, a communication-efficient variant with theoretical guarantees, achieving similar performance to STSA-E with much lower communication overhead. Extensive experiments on three widely used FCIL datasets, with varying degrees of data heterogeneity, show that our method outperforms state-of-the-art FCIL methods in terms of performance, flexibility, and both communication and computation efficiency. The code is available at https://github.com/Yuqin-G/STSA.

cs.LG↗

arXiv2013

BEC-BCS Crossover with Feshbach Resonance for a Three-Hyperfine-Species Model

We consider the behavior of an ultracold Fermi gas across a narrow Feshbach resonance, where the occupation of the closed channel may not be negligible. While the corrections to the single-channel formulae associated with the nonzero chemical potential and with particle conservation have been considered in the existing literature, there is a further effect, namely the "inter-channel Pauli exclusion principle" associated with the fact that a single hyperfine species may be common to the two channels. We focus on this effect and show that, as intuitively expected, the resulting corrections are of order $E_F/η$, where $E_F$ is the Fermi energy of the gas in the absence of interactions and $η$ is the Zeeman energy difference between the two channels. We also consider the related corrections to the fermionic excitation spectrum, and briefly discuss the collective modes of the system.

cond-mat.quant-gas↗

arXiv2012

BEC-BCS Crossover with Feshbach Resonance for Three-Hyperfine-Species Model

In a Feshbach resonance, the effective s-wave scattering length grows when one moves toward the resonance point, and eventually diverges at this point. There is one characteristic energy scale, $δ_c$, defined as, in the negative side of the resonance point, the detuning energy at which the weight of the bound state shifts from predominatedly in the open-channel to predominated in the closed-channel. When the many-body energy scale (e.g. the Fermi energy, $E_{F}$) is larger than $δ_c$, the closed-channel weight is significant and has to be included in the many-body theory. Furthermore, when two channels share a hyperfine species, the Pauli exclusion between fermions from two channels also needs to be taken into consideration in the many-body theory. The current thesis addresses the above problem in detail. A set of gap equations and number equations are derived at the mean-field level. The fermionic and bosonic excitation spectra are then derived. Assuming that the uncoupled bound-state of the closed-channel in resonance is much smaller than the inter-particle distance, as well as the s-wave scattering length, $a_s$, we find that the basic equations in the single-channel crossover model are still valid. The correction first comes from the existing of the finite chemical potential and additional counting complication due to the closed-channel. These two corrections need to be included into the mean-field equations, i.e. the gap equations and the number equations, and be solved self-consistently. Then the correction due to the inter-channel Pauli exclusion is in the order of the ratio of the Fermi energy and the Zeeman energy difference between two channels, $E_F/η$, which can be analyzed perturbatively over the previous corrections. Fermionic and bosonic excitation modes are studied.

cond-mat.quant-gas↗

arXiv2011

Competition between BCS-pairing and "moth-eaten effect" in BEC-BCS crossover

We study the change in condensation energy from a single pair of fermionic atoms to a large number of pairs interacting via the reduced BCS potential. We find that the energy-saving due to correlations decreases when the pair number increases because the number of empty states available for pairing gets smaller ("moth-eaten effect"). However, this decrease dominates the 3D kinetic energy increase of the same amount of noninteracting atoms only when the pair number is a sizeable fraction of the number of states available for pairing. As a result, in BEC-BCS crossover of 3D systems, the condensation energy per pair first increases and then decreases with pair number while in 2D, it always is controlled by the "moth-eaten effect" and thus simply decreases.

cond-mat.supr-con↗

arXiv2010

Coboson formalism for Cooper pairs used to derive Richardson's equations

We propose a many-body formalism for Cooper pairs which has similarities to the one we recently developed for composite boson excitons (coboson in short). Its Shiva diagram representation evidences that $N$ Cooper pairs differ from $N$ single pairs through electron exchange only: no direct coupling exists due to the very peculiar form of the BCS potential. As a first application, we here use this formalism to derive Richardson's equations for the exact eigenstates of $N$ Cooper pairs. This gives hints on why the $N(N-1)$ dependence of the $N$-pair ground state energy we recently obtained by solving Richardson's equations analytically in the low density limit, stays valid up to the dense regime, no higher order dependence exists even under large overlap, a surprising result hard to accept at first. We also briefly question the BCS wave function ansatz compared to Richardson's exact form, in the light of our understanding of coboson many-body effects.

cond-mat.supr-con↗