Search arXiv⌕ Search

arXiv subjects

Ming Zhang

Publications and source records attributed to Ming Zhang.

At least 37 records · Page 2Linked to original sources
arXiv2026

Phase transitions and quasinormal modes of Euler-Heisenberg-dilaton AdS black holes

In this paper, we study phase transitions and scalar quasinormal modes(QNMs) of magnetically charged Euler--Heisenberg-dilaton AdS black holes in the extended phase space. We classify the horizon structures for all three signs of the EH coupling and derive the extended first law when the pressure and coupling are varied. For negative coupling, the physical branch exhibits a Van der Waals-like small/large black-hole transition. Using the improved asymptotic iteration method, we find two separated QNM trajectories below the critical pressure, whereas they merge into one continuous curve at the critical point. The GHS limit and the positive couplings considered here do not produce a finite-radius critical point.

hep-th↗
arXiv2026

ExplorationBench: Measuring AI Systems' Exploration in Verifiable Alien Worlds

Scientific discovery begins where known problems end. There, AI systems must engage in exploration: framing hypotheses, designing experiments, and iterating on the results. However, evaluating this ability is difficult: (1) how to verify whether a genuinely new hypothesis holds, and (2) how to determine whether a system has discovered it through exploration or merely recalled related knowledge from pre-training data. To this end, we introduce ExplorationBench, which turns the wicked problem of evaluating scientific exploration into a concrete and tractable framework built on verifiable Alien Worlds: their rules are executable, so every answer can be checked exactly, and they conflict with familiar knowledge, so recall alone cannot solve the tasks. The benchmark contains two sandboxes, AlienCode (31 discovery targets, 70 tasks) and AlienLogic (24 discovery targets, 70 tasks). Each sandbox provides a flawed manual, task-specific environmental feedback, and a dedicated tool-call schema. Systems use these resources to explore the sandbox, then solve held-out tasks. We evaluate 10 AI systems and find that the strongest systems can acquire and apply unfamiliar rules, while performance varies substantially across trajectories and continued exploration can stall or reverse earlier gains. ExplorationBench represents a step towards AI systems that can acquire and apply genuinely new knowledge through exploration in unknown environments.

cs.AI↗
arXiv2026

ReSI: Recursive Safety Improvement toward Resistant and Resilient AI

Recursive self-improvement, the participation of AI systems in improving their own capabilities, is beginning to move from theoretical prospect to practice, posing both challenges and opportunities for safety alignment. Models evolve through frequent updates, and their safety alignment requires continual adaptation to each new checkpoint. Meanwhile, with evolving red-teaming methods exposing new vulnerabilities, safety improvement for each checkpoint needs to mitigate exposed vulnerabilities and generalize to risks not yet revealed. Following R$^2$AI, we term these goals resistance to known threats and resilience to unforeseen risks. Recursive self-improvement, in turn, inspires an approach to both goals: safety alignment could likewise advance through successive rounds of evaluation and update. We therefore introduce ReSI, a recursive safety improvement framework that implements this approach through automated research. In each round, ReSI applies diverse red-teaming methods to identify vulnerabilities in the current target model, develops training recipes, and promotes the update with the largest safety gain among those passing a Pareto gate on capability retention as the next target model. Across four dense and mixture-of-experts models, ReSI matches or exceeds evaluated frontier models on in-distribution and out-of-distribution safety benchmarks, and outperforms alignment baselines on nearly all safety evaluations while largely preserving general capabilities. In particular, ReSI reduces the mean X-Teaming attack success rate across the four models from 86.01% to 31.45%, well below GPT-5.6-Luna's leading frontier result of 56.69%, indicating stronger resilience to attacks unseen during training. These findings support recursive safety improvement as a practical path toward resistant and resilient AI.

cs.CR↗
arXiv2026

Time-domain analysis of scalar perturbations of Kerr--Newman black holes with Chern--Simons and Maxwell couplings

To study the onset of scalarization, we investigate linear scalar perturbations around Kerr--Newman black holes in the Einstein-Chern-Simons-Maxwell-scalar theory with quadratic Chern--Simons (CS) coupling $γ$ ($\hatγ=γ/M^2$) and Maxwell coupling $η$. For this purpose, we employ hyperboloidal Rácz--Tóth evolution and a stationary axisymmetric zero-mode computation. The resulting thresholds are compared with the time-domain response at representative points. For $\hatγ=0$, we find two asymmetric threshold branches in $η$, while for $\hatγ=η$, the two branches have opposite signs and unequal magnitudes. For unequal CS and Maxwell couplings, the two real threshold branches merge as the spin or CS coupling increases.

gr-qc↗
arXiv2026

HyperLogic: A Hard, Forward-Authored Chinese Logical Reasoning Benchmark with Execution-Derived Answers

Existing logic benchmarks primarily measure models' ability to answer reasoning questions directly. Scalable benchmarks often generate text from formal structures, which makes answers easy to compute but fixes the formalization before the problem is written. Forward construction preserves the challenge of finding a faithful formalization, yet makes difficulty and answer reliability harder to control. We introduce HyperLogic, a forward-construction pipeline that separates problem authoring from answer generation. A multi-agent workflow hardens undergraduate-authored Chinese seeds without solving them; two agents from different model families independently translate each finished item into executable finite-domain models; their encodings and solver-derived answers undergo layered, agent-assisted adjudication under human-expert oversight. HyperLogic-Base contains 195 items and 922 sub-questions and separates seven frontier models by 33.0 percentage points in strict item accuracy (44.6-77.6%). HyperLogic-Hard contains 100 items with larger, coupled search spaces, on which no model exceeds 16% accuracy in direct answering. We also use Hard to evaluate agents' ability to formalize and solve problems with tools, comparing a code sandbox alone with one that includes our logic modeling library. The sandbox improves every model by 16.7-40.1 points; adding the library helps five models and hurts two. These results highlight the difficulty of faithful formalization even with tool access.

cs.CL↗
arXiv2026

MuseCritic: Learning Multi-Aspect Song Rewards through Natural-Language Aesthetic Critiques

Long-form song generation models continue to improve in duration, structural coherence, and acoustic complexity, increasing the need for reliable aesthetic rewards aligned with human preferences. However, reward models for complete songs remain limited, and existing evaluators typically predict scores in a single forward pass without readable explanations. To this end, we introduce MuseCritic, a semi-scalar reward model that generates a natural-language critique covering five aesthetic dimensions and uses it as an intermediate representation to predict continuous reward scores. MuseCritic follows a two-stage training pipeline: a teacher model first provides high-quality critiques for supervised fine-tuning, then the fine-tuned model generates its own critiques for reward learning, mitigating training-inference distribution shift. On an in-domain test set of 200 SongEval songs, MuseCritic reduces macro-averaged mean squared error from 0.2875 to 0.2316 and improves macro-averaged LCC, SRCC, and Kendall's tau to 0.9068, 0.8838, and 0.7178, respectively. On the out-of-domain Music Arena benchmark with 733 preference pairs, it achieves 71.35% accuracy and remains competitive with strong music-specific reward models. Using MuseCritic with GRPO also improves Muse-0.6B on all nine aesthetic metrics from SongEval and Audiobox Aesthetics. These results show that critique-conditioned reward modeling reduces scoring error and provides an effective optimization signal for song generation. The project repository is available at https://github.com/WuqnEl/MuseCritic.

cs.SD↗
arXiv2026

Analytic detector responses and entanglement in constant-curvature spacetimes

We derive analytic response functions for Unruh--DeWitt detectors coupled to real, massless, conformally coupled scalar fields, to distinguish the effects of the field spectrum and global geometry from those of the interaction protocol. For Gaussian-switched static pairs in Minkowski, anti-de Sitter, and de Sitter spacetimes, we obtain the leading-order excitation probabilities and two-detector coherences in spacetime dimension $\mathcal D=d+1\ge3$. Time-ordered coherence requires distinct spatial worldlines; the local response and single-excitation coherence also admit coincidence. For equal-redshift pairs, Gaussian switching separates the gap dependence of the time-ordered coherence. In de Sitter space, KMS detailed balance fixes the gap that maximizes the ratio of nonlocal coherence to local excitation, giving a single-gap test for the existence of leading-order entanglement and restricting any entangling gaps to one bounded interval. When entanglement is present, the concurrence itself peaks at a smaller positive gap. In the three-dimensional BTZ black hole, an integrated Legendre series gives the sharply switched response of a radially infalling detector. A uniform image-sum estimate establishes logarithmic growth near the singularity and determines its coefficient from the covering-AdS response. A sufficiently smooth monotone onset removes the finite-time onset glitches but leaves a positive logarithmic coefficient; the accumulated leading-order response remains finite. Independent numerical calculations test the analytic expressions, their convergence properties, and these physical distinctions.

gr-qc↗
arXiv2026

Holographic complexity of de Sitter black holes

We investigate holographic complexity within the Schwarzschild-de Sitter (SdS) black hole spacetime. Two distinct de Sitter holography prescriptions are examined: the static patch scheme restricted to the stretched horizon and the de Sitter/Conformal Field Theory (dS/CFT) correspondence scheme defined at asymptotic future and past infinities. We evaluate the Complexity equals Volume (CV) conjecture and extend the analysis to codimension-zero proposals, specifically Complexity equals Spacetime Volume (CV2.0) and Complexity equals Action (CA), through the Wheeler-DeWitt (WDW) patch we construct. The behaviors of the complexity in the static patch holography at late time and in the dS/CFT at infinite spacelike boundary coordinate are studied, respectively. We find that under both the CV and CV2.0 conjectures, the static patch holographic complexity and the dS/CFT holographic complexity consistently exhibit linear growth. Conversely, regarding the CA conjecture, the holographic complexity growth rates for both the static patch and the dS/CFT correspondence vanish. This behavior is attributed to the finiteness of the (regularized) action within the restricted WDW region. Furthermore, we find that, in the limit in which the boundary coordinate is taken to infinity, the complexity growth rate in the static patch prescription coincides with its counterpart in the dS/CFT prescription. This agreement suggests a deeper connection between the two descriptions and may point toward a unified understanding of bulk dynamics in de Sitter holography.

hep-th↗
arXiv2026

Observational constraints on a damped harmonic oscillator model of dark energy

We constrain a damped harmonic oscillator (DHO) dark-energy equation of state using the full cosmic microwave background (CMB) likelihoods in combination with DESI BAO and three distinct Type Ia supernova compilations: Pantheon+, DES-Dovekie, and Union3. The equation of state obeys a second-order damped oscillator equation in number of $e$-folds, so that its frequency $f$, damping rate $b$, and equilibrium value $w_{\rm m}$ fully specify the late-time dynamics. The model exhibits oscillatory behavior only at low redshifts, around the equilibrium value $w=-1$, with distinct characteristics for the different supernova compilations: an underdamped solution for DES-Dovekie and Union3, and an overdamped solution for Pantheon+. At higher redshifts, the model closely mimics $Λ$CDM and deviates significantly only at $z<0.6$, with the magnitude of the deviation depending on the supernova compilation. We further identify a region of the $(f,b)$ parameter space, corresponding to rapid variation of the equation of state at low redshift, in which the perturbation equations become numerically stiff and cannot be integrated with a canonical dark-energy sound speed i.e., $c_s^2 =1$. We show that reducing the rest-frame sound speed removes this obstruction while leaving the observables unchanged at the $10^{-3}$ level, and therefore treat it as a numerical prescription rather than a physical modification of the model. The model yields $H_0 = 67.53^{+1.22}_{-1.18}$ km/s/Mpc for Pantheon+, $H_0 = 69.08^{+1.23}_{-1.16}$ km/s/Mpc for DES-Dovekie, and $H_0 = 70.67^{+1.89}_{-1.87}$ km/s/Mpc for Union3. The present-day equation-of-state parameter is constrained to $w_0 = -0.521^{+0.891}_{-0.414}$, $-3.01^{+1.14}_{-1.21}$, and $-3.16^{+1.11}_{-1.16}$ for Pantheon+, DES-Dovekie, and Union3, respectively.

astro-ph.CO↗
arXiv2026

Measurement of the Weyl Potential Evolution and $E_G$ Statistic from KiDS-1000, BOSS and 2dFLenS

A recently developed model-independent approach to measuring the Weyl potential has shown some tensions with $Λ$CDM (Lambda Cold Dark Matter) in DES (Dark Energy Survey) Y3 data. We apply this framework to Kilo-Degree Survey (KiDS-1000) weak lensing and BOSS/2dFLenS galaxy clustering using the KiDS Cosmology Analysis Pipeline (KCAP) in two redshift bins. To apply this method to the BOSS spectroscopic wedges, we take the early-time power spectrum from a redshift where General Relativity (GR) is recovered and jointly fit the late-time observables $\hat{b}$, $\hat{f}$, and $\hat{J}$. With Planck18 priors, the CMASS measurement of $\hat{J}$ is $1.50σ$ below the GR prediction, while the LOWZ measurement remains consistent with it. The decrease in the central value of $\hat{J}$ from LOWZ to CMASS provides a mild indication of the previously reported ``Lensing is low'' behavior. Without Planck18 priors, we obtain $E_G(0.38)=0.503\pm0.054$ and $E_G(0.61)=0.444\pm0.054$, both consistent with GR. Furthermore, phenomenological gravity tests also show no significant preference for modified gravity.

astro-ph.CO↗
arXiv2026

AutoScientist-Quant: Self-Evolving Coding Agents for Automatic Research in Quantitative Investment

Large language model agents can discover alphas, yet current methods have three weaknesses. The search cannot adapt during the run, automation usually ends at alpha generation while library selection and model choice stay manual, and alpha discovery can read the test window through loop feedback or code problems. We present AutoScientist-Quant, a self evolving search process that regards quantitative research as one budgeted search problem. A single controller conditions every decision on the remaining budget, choosing at each round whether to improve, combine, pivot, or stop, which node to expand, how many alphas to generate, and how to retrieve past trajectories from the shared memory. The same core then selects from the library and tunes the model, closing the loop from hypothesis to deployable strategy. We also review the evaluation pipeline reused from prior work, fix two lookahead problems, and keep the feedback window disjoint from the held out test window, so every comparison tests true generalization. On CSI universes, the framework attains the best value of nearly every metric in every setting, and these conclusions hold across several backbones and markets.

cs.AI↗
arXiv2026

JFTA-Bench: Evaluate LLM's Ability of Tracking and Analyzing Malfunctions Using Fault Trees

In the maintenance of complex systems, fault trees are used to locate problems and provide targeted solutions. To enable fault trees stored as images to be directly processed by large language models, which can assist in tracking and analyzing malfunctions, we propose a novel textual representation of fault trees. Building on it, we construct a benchmark for multi-turn dialogue systems that emphasizes robust interaction in complex environments, evaluating a model's ability to assist in malfunction localization, which contains $3130$ entries and $40.75$ turns per entry on average. We train an end-to-end model to generate vague information to reflect user behavior and introduce long-range rollback and recovery procedures to simulate user error scenarios, enabling assessment of a model's integrated capabilities in task tracking and error recovery, and Gemini 2.5 pro archives the best performance.

cs.AI↗
arXiv2026

Safin-1: Safety from Within through Memory-Native State Evolution

Long-horizon complex tasks require foundation models to accumulate information, maintain internal states, and adapt over extended interactions. Safety should be an intrinsic property of the model itself, rather than a behavioral constraint relying solely on external safeguards or post-hoc alignment such as supervised fine-tuning. This motivates Safety from Within, where safety-relevant capabilities are represented and invoked through the model's native computation. We present Safin-1, a family of foundation models realizing this principle through memory routing and state evolution. Safin-1 is built on Memory-Anchor Routing across Context History (MARCH), a network architecture that maintains structured memory states and selectively retrieves relevant historical information through content-conditioned routing. It supports test-time adaptation of persistent capability states without repeatedly modifying the backbone, enabling controlled specialization over a shared foundation. We investigate this interface on downstream safety tasks through a Safety State, demonstrating effective state-based adaptation with substantial safety improvements. More broadly, the routed-state interface unifies contextual memory and persistent capability adaptation within the model's native computation, reframing memory from a passive record of prior context into an active substrate for maintaining and evolving model behavior. Evaluations across general capabilities, long-context understanding, retrieval, and efficiency further validate Safin-1. These findings provide a path toward safety as a state-native and adaptively maintainable capability. This work is only an initial architectural exploration of Safety from Within, and substantial further work is needed to realize this broader vision.

cs.LG↗
arXiv2026

Switch Attention: Towards Dynamic and Fine-grained Hybrid Transformers

The attention mechanism has been the core component in modern transformer architectures. However, the computation of standard full attention scales quadratically with the sequence length, serving as a major bottleneck in long-context language modeling. Sliding window attention restricts the context length for better efficiency at the cost of narrower receptive fields. While existing efforts attempt to take the benefits from both sides by building hybrid models, they often resort to static, heuristically designed alternating patterns that limit efficient allocation of computation in various scenarios. In this paper, we propose Switch Attention (SwiAttn), a novel hybrid transformer that enables dynamic and fine-grained routing between full attention and sliding window attention. For each token at each transformer layer, SwiAttn dynamically routes the computation to either a full-attention branch for global information aggregation or a sliding-window branch for efficient local pattern matching. An adaptive regularization objective is designed to encourage the model towards efficiency. Moreover, we adopt continual pretraining to optimize the model, transferring the full attention architecture to the hybrid one. Extensive experiments are conducted on twenty-three benchmark datasets across both regular (4K) and long (32K) context lengths, demonstrating the effectiveness of the proposed method.

cs.CL↗
arXiv2026

AI Can Learn Scientific Taste

Scientific discovery depends on expert judgement and foresight, which we call scientific taste: the ability to judge and propose research ideas with the potential for long-term scientific impact. Scientific taste is largely concentrated among highly experienced researchers, whose expertise is usually limited to a few specialised fields. If AI could learn scientific taste, it could reduce reliance on human experts and accelerate scientific discovery. Whether AI can learn this ability remains an open question. We introduce Reinforcement Learning from Community Feedback (RLCF) to learn judgement and ideation. Scientific Judge learns from community feedback, such as citations. Scientific Thinker learns to propose research ideas with high potential impact. Experiments show that Scientific Judge outperforms strong LLM baselines and that learned judgement generalises to future-year papers, other community metrics, and unseen fields. Furthermore, Scientific Thinker proposes research ideas with higher potential impact than those proposed by baselines. These results suggest that AI can learn scientific taste, marking an important step towards AI systems that could help accelerate scientific discovery.

cs.CL↗
arXiv2026

Partial FC: Training 10 Million Identities on a Single Machine

Training face recognition models with millions of identities is challenging because classifier storage, logit memory, and computation grow linearly with the number of classes, eventually making full softmax impractical even when the backbone itself fits comfortably in memory. We present Partial FC (PFC), a scalable approximation to large-class softmax that preserves every positive class center while activating only a sampled subset of negative centers in each mini-batch. This asymmetric treatment retains every target term while avoiding exhaustive interaction with millions of mostly uninformative negatives. Our distributed implementation partitions the classifier across GPUs and samples within each owned shard, so sampling reduces local matrix multiplication and logit storage while sharding eliminates class-gradient synchronization across workers. Together, these properties reduce GPU-resident classifier memory, computation, and class-dependent communication without feature-based hard-negative retrieval. End-to-end system benchmarks demonstrate efficient scaling to massive class spaces, including 64 million classes on a single eight-GPU machine. Across large-scale face-recognition datasets, moderate sampling rates maintain competitive recognition accuracy while substantially improving training efficiency. Our best PFC configurations achieve 97.2\% TAR on IJB-C at FAR $=10^{-4}$ and 94.0\% TAR on ICCV21-MFR at FAR $=10^{-6}$. Beyond clean training data, PFC is robust to inter-class conflicts, label noise, and long-tailed identity distributions: under 40\% label noise, PFC with conflict filtering raises ICCV21-MFR TAR from 43.9\% to 80.2\%, while on long-tailed data PFC improves TAR from 87.4\% to 92.0\%. These results establish positive-preserving negative sampling as an effective foundation for scalable, accurate, and robust identity classification.

cs.CV↗
arXiv2026

Quasinormal frequencies and greybody factors for axial perturbations of dilaton-Euler-Heisenberg de Sitter black holes

We investigate the quasinormal modes (QNMs) and greybody factors of dilaton-Euler-Heisenberg (dEH) de Sitter (dS) black holes in string-inspired Euler-Heisenberg gravity. Since the axial gravitational and electromagnetic perturbations decouple, we treat them independently. By applying the asymptotic iteration method (AIM) alongside a sixth-order WKB approximation, we compute the quasinormal frequencies and find excellent agreement between the two approaches. We also find that the QNM spectra depend sensitively on the magnetic charge $Q_{\text{m}}$, cosmological constant $Λ$, and nonlinear coupling $ε$, with a notable topological anomaly appearing in the electromagnetic frequency trajectories. Additionally, larger values of $Q_{\text{m}}$ or the multipole number $l$ generally suppress wave transmission, while the electromagnetic sector with $ε=1$ exhibits an anomalous response.

hep-th↗
arXiv2026

Intern-S2-Mobius: Foundation Model with Decoupled Knowledge and Reasoning

We introduce Mobius-v0, an architecture that comprises a globally shared Memory (FFN) that stores knowledge vectors and multiple Reasoners (Self-Attn) that iteratively achieve compositional reasoning. Using hidden states as cache and carrier, reasoners repeatedly query memory for required knowledge-vectors, while the knowledge is transmitted back to reasoning operators. Through this knowledge-reasoning-separation architecture, Mobius achieves better knowledge compression and reasoning efficiency. Built upon Mobius-v0 architecture: 1) Our 7B model trained-from-scratch achieves similar downstream score as a 7B Transformer baseline with 62.6% of baseline's training data. 2) Our Intern-S2-Mobius, continually-pretrained from Qwen3.5-35B, achieves similar downstream score while delivering nearly 4x end-to-end inference speedup.

cs.AI↗