arXiv subjects

Yifan Yu

Publications and source records attributed to Yifan Yu.

At least 19 recordsLinked to original sources

arXiv2026

Osprey: Target-agnostic Pre-training Makes Stronger Drafters in Speculative Decoding

Speculative decoding is critical for accelerating LLM inference. However, the speedup is fragile: drafters are typically trained against a narrow distribution for a single target model, and their acceptance rate collapses under workload shifts. This is a striking inversion of modern LLM development, where target models are valued precisely for the broad generalization they acquire through large-scale pretraining. We argue that the natural remedy, pretraining, has been hard to apply to drafters because existing recipes are target-specific: the drafter consumes the target's hidden states and is distilled on the target's logits, so pretraining must be repeated for each target. We introduce Osprey, which instead bootstraps drafters from off-the-shelf pretrained small language models, treating broad pretraining as a reusable, target-agnostic asset and reducing per-target work to a lightweight adaptation step. Realizing this requires overcoming two challenges: small LMs are far deeper than a latency-bound drafter can afford, and their pretrained computation must remain intact while the drafter learns to ingest target hidden states and emit tokens in the target's vocabulary. Osprey addresses both by pruning to a shallow backbone, restoring its language-modeling capability with target-agnostic next-token pretraining, and adapting it to each target through vocabulary alignment, zero-initialized QKV expansion, and distillation from the target model's output distribution. Empirically, a single pretrained Osprey backbone transfers across targets and improves mean acceptance length by 16.1% for Qwen3-8B, 21.2% for Llama-3.3-70B-Instruct, and 22.7% for the 229B MiniMax-M2.5 (with 17.5% higher tokens per second), with the largest gains on out-of-domain and multilingual data. Our code is available at https://github.com/LeanModels/Osprey.

cs.CL↗

arXiv2026

Host-guest Crystal Engineering Tailors the Room Temperature Spin Dynamics in Molecular Quantum Devices

Molecular materials that enable coherent control over an electron's spin state at room temperature are promising candidates for quantum technologies, including quantum sensors and ultra-low noise microwave amplifiers, known as masers. Host-guest molecular crystals enable independent control of spin-active guests and their local environments to enhance molecular spin properties and so improve device performance. Using electron paramagnetic resonance and optically-detected magnetic resonance, we demonstrate the ability to tune triplet population, depopulation, and spin-lattice relaxation by modulating host-dependent lattice rigidity and vibrational coupling to significantly reduce the operating requirements for building useful masers. Importantly, the most rigid host, picene, reveals the ability to slow spin-lattice relaxation without lengthening triplet lifetime, though at the cost of strain-induced line width broadening and reduced triplet spin polarisation. We also find that deuteration reduces the triplet resonance line width and vibrationally-mediated triplet depopulation. Consequently, we find that perdeuterated pentacene in perdeuterated p-terphenyl is the most viable candidate for building a continuous wave maser. This work demonstrates host-guest engineering as an important and practical method for tuning the spin-dependent performance of room-temperature molecular quantum technologies.

quant-ph↗

arXiv2026

Harsher on Male? Evaluating LLMs on Gender-Asymmetric Moral Framing Across Diverse Conflict Scenarios

Existing studies on gender bias in LLMs have largely focused on stereotypes, occupational associations, or explicit harmful outputs. In this work, we ask whether LLMs apply consistent response standards to the same negative behavior under matched male-actor and female-actor conditions. We introduce GAMA-Bench, a gender-mirrored benchmark of 1,298 scenarios covering intimate relationship and public social conflicts. It constructs gender-neutral misconduct templates through controlled grids and cross-model review, then compiles them into paired first-person prompts with matched actor-gender and role-reference variations. We further design a structured response-framing protocol to measure how models allocate punishment, empathy, escalation, instruction, and blame. Experiments on 10 representative LLMs reveal a consistent male-disadvantaging asymmetry: male actors receive more punitive, escalatory, and blame-centered framing, whereas female actors receive more therapeutic and empathy-oriented framing for the same misconduct. Further analyses show that this pattern persists across model families, scenario tracks, model scale, and explicit thinking-style reasoning. The official code is available at https://github.com/xufeiqiong/GAMA-Bench.

cs.CL↗

arXiv2026

Interdependent Hitting Times

This paper studies interdependent durations as equilibrium outcomes of a synchronization game, a continuous-time stopping game in which the incentive to stop increases when other players stop. We allow the payoffs to vary with both common shocks and observed and unobserved agent characteristics. The common shocks follow a spectrally negative L\'evy process, a semiparametric process that includes Brownian motion as a special case but may also have jumps. We show that equilibrium outcomes can be represented as interdependent hitting times and use this to establish the game's nonparametric identification from data on stopping times and covariates. We develop maximum simulated likelihood and method of simulated moments estimators and evaluate their finite-sample and computational performance in Monte Carlo experiments. The results provide a tractable framework for identifying and estimating synchronization games from interdependent duration data.

econ.EM↗

arXiv2026

Geometry-Preserving Nudged Elastic Band and Dimer Methods under Anisotropic Force Uncertainty

The nudged elastic band (NEB) and Dimer methods are standard tools for computing minimum-energy paths and index-one saddle points in atomistic transition problems. They are increasingly driven by surrogate or learned force models, whose force errors are often anisotropic and spatially varying near transition states and defect cores, where saddle-search iterations are most sensitive. We introduce uncertainty-aware NEB and Dimer methods (UA-NEB, UA-Dimer) that use covariance as an optimizer-level reliability metric while preserving the mean-potential saddle-search equations: an oblique normal projection for NEB and covariance-weighted rotation and translation for Dimer. Both algorithms fit Robbins--Monro recursions; under a local Lyapunov stability hypothesis, verified explicitly for a canonical UA-NEB setting and stated as a hypothesis for UA-Dimer, the stochastic iterations converge almost surely within the corresponding local stability neighborhood. In the analytic benchmark, UA-NEB reduces mean barrier error by $21\%$ relative to stochastic NEB and UA-Dimer reduces the reflected-gradient residual by $22\%$; in the 127-atom tungsten-vacancy benchmark, full UA-NEB reduces mean barrier error by $56\%$ relative to stochastic NEB and by $23\%$ relative to diagonal covariance weighting. These results show that anisotropic uncertainty is most useful when embedded in the constrained geometry of the optimizer rather than collapsed into a scalar acquisition or trust criterion.

math.NA↗

arXiv2026

EgoProx: Evaluating MLLMs on Egocentric 3D Proximity Reasoning Across a Cognitive Hierarchy

Humans constantly reason about 3D proximity, the relations between their body and surrounding objects, to guide perception and action in daily life. Whether multimodal large language models (MLLMs) can perform such embodied 3D reasoning remains unclear. To this end, we introduce EgoProx, a benchmark for egocentric 3D proximity reasoning. We organize our tasks along a cognitive chain, covering intention, exploration, exploitation, and chain-of-actions reasoning. We also design an agent based data engine that produces diverse and consistent QA pairs at scale. We benchmark prevailing MLLMs on EgoProx and conduct additional analyses with dataset specific and task specific instruction tuning. We observe large cross-domain gains, indicating that current MLLMs contain some spatial knowledge; however, they still struggle to effectively leverage it for spatial reasoning VQA.

cs.CV↗

arXiv2026

Introspective Diffusion Language Models

Diffusion language models promise parallel generation, yet still lag behind autoregressive (AR) models in quality. We stem this gap to a failure of introspective consistency: AR models agree with their own generations, while DLMs often do not. We define the introspective acceptance rate, which measures whether a model accepts its previously generated tokens. This reveals why AR training has a structural advantage: causal masking and logit shifting implicitly enforce introspective consistency. Motivated by this observation, we introduce Introspective Diffusion Language Model (I-DLM), a paradigm that retains diffusion-style parallel decoding while inheriting the introspective consistency of AR training. I-DLM uses a novel introspective strided decoding (ISD) algorithm, which enables the model to verify previously generated tokens while advancing new ones in the same forward pass. From a systems standpoint, we build I-DLM inference engine on AR-inherited optimizations and further customize it with a stationary-batch scheduler. To the best of our knowledge, I-DLM is the first DLM to match the quality of its same-scale AR counterpart while outperforming prior DLMs in both model quality and practical serving efficiency across 15 benchmarks. It reaches 69.6 on AIME-24 and 45.7 on LiveCodeBench-v6, exceeding LLaDA-2.1-mini (16B) by more than 26 and 15 points, respectively. Beyond quality, I-DLM is designed for the growing demand of large-concurrency serving, delivering about 3x higher throughput than prior state-of-the-art DLMs.

cs.AI↗

arXiv2026

From Dyads to Groups: Rethinking Emotional Support with Conversational AI

Advances in artificial intelligence (AI), together with persistent gaps in access to reliable emotional support, have positioned AI as an increasingly prominent source of emotional assistance. However, most AI-based emotional support applications and prior research focus on one-on-one interactions between users and a single AI agent, leaving the potential advantages of alternative support configurations largely unexplored. Drawing on social support and support group theory, this research examines whether AI-based emotional support delivered by a group of AI agents (group AI support) can constitute a more effective support form than single-agent support (single AI support). We propose that group AI support enhances users' perceived support efficacy, that this effect operates by strengthening users' connectedness with the AI system, and that the composition of support types within AI groups further shapes support outcomes. Three experiments provide convergent support for these claims. By identifying when and why group AI emotional support outperforms single AI support, this work advances theoretical understanding of AI-based emotional support and provides actionable guidance for the design of AI support systems.

cs.HC↗

arXiv2026

Variability in Performance of a Machine-Learning Seismicity Catalog: Central Italy, 2016-2017

Machine learning (ML) catalogs contain many more earthquakes than routine catalogs, but their performance in phase picking and earthquake detection has not been fully evaluated. We develop station-level detection probabilities using logistic regression and combine them across a seismic network to compute spatial magnitude-of-completeness fields. We apply this approach to two catalogs from the 2016-2017 Central Italy sequence that were constructed from the same seismic network, one routine and one ML based. At the station level, the ML picker increases detection sensitivity by identifying smaller magnitude events and detecting earthquakes at greater distances. Spatially, the magnitude-of-completeness decreases substantially, with median values shifting from 1.6 to 0.5 for P waves and from 1.7 to 0.5 for S waves. However, the ML catalog also shows greater variability in station-level performance than the routine catalog. These results demonstrate that ML-based improvements in detectability are widespread but spatially non-uniform, highlighting their benefits, their limitations, and the potential for further improvements.

physics.geo-ph↗

arXiv2026

IOTA: Corrective Knowledge-Guided Prompt Learning via Black-White Box Framework

Recently, adapting pre-trained models to downstream tasks has attracted increasing interest. Previous Parameter-Efficient-Tuning (PET) methods regard the pre-trained model as an opaque Black Box model, relying purely on data-driven optimization and underutilizing their inherent prior knowledge. This oversight limits the models' potential for effective downstream task adaptation. To address these issues, we propose a novel black-whIte bOx prompT leArning framework (IOTA), which integrates a data-driven Black Box module with a knowledge-driven White Box module for downstream task adaptation. Specifically, the White Box module derives corrective knowledge by contrasting the wrong predictions with the right cognition. This knowledge is verbalized into interpretable human prompts and leveraged through a corrective knowledge-guided prompt selection strategy to guide the Black Box module toward more accurate predictions. By jointly leveraging knowledge- and data-driven learning signals, IOTA achieves effective downstream task adaptation. Experimental results on 12 image classification benchmarks under few-shot and easy-to-hard adaptation settings demonstrate the effectiveness of corrective knowledge and the superiority of our method over state-of-the-art methods.

cs.CV↗

arXiv2026

Sel-assembled Rhodium Nanoantennas for Single-Protein UV SERS

Surface-enhanced Raman scattering (SERS) provides critical insights into analyte structure, dynamic processes, and intermolecular interactions at the single-molecule level. By exploiting the hotspot formation in the vicinity of plasmonic structures, SERS constitutes an established tool for fundamental biological research, particularly for early-stage disease diagnostics. In this context, the DNA Origami technique, with its high addressability, enables both the assembly of plasmonic nanostructures with nanometric accuracy, and the deterministic placement of a single analyte molecule precisely at the generated hotspot within them. To date, most DNA Origami based nanoantennas rely on gold or silver nanoparticles (NPs), whose plasmonic resonances are confined to the visible spectrum, severely limiting their use in other spectral ranges. To extend the operating range, we have recently established a robust strategy for self-assembling programmable ultraviolet (UV)-plasmonic dimer antennas using rhodium nanocubes. Herein, we leverage this tailored architecture to systematically investigate its performance for single-molecule UV-SERS. We demonstrated how biofabricated Rh-dimers can be used to detect the characteristic SERS signal of a single streptavidin molecule linked at the dimer s gap. Our results are validated through polarization dependent measurements that yield the expected signal modulation depending on the the dimer orientation only for the DNA origami with a protein at the hotspot. This work establishes a highly sensitive and polarization-tunable UV-SERS platform, laying a solid foundation for label-free optical investigation and bio-spectroscopy of individual biomolecules in the UV spectral range.

physics.app-ph↗

arXiv2026

Scaleable LED-pumped Room-temperature Maser using a Multi-blade Optical Injector

Though the performance of room-temperature masers has improved over the last decade, relatively little attention has been paid to the optics used to pump the maser's gain medium. In this work, we investigate a novel multi-blade optical ``injector'' that permits more effective and more scaleable pumping. The reported work encompasses an interdisciplinary mix of conceptualization, simulation, crystal growth, fabrication, and microwave engineering. Our gain medium is pentacene dissolved as a solid solution with para-terphenyl (Pc:PTP) molecular crystal. We accurately determine this pentacene's molecular absorption cross-section as a function of wavelength. Ray-tracing is then used to assess how different designs of waveguide inject light into the Pc:PTP crystal. A multi-blade injector made of high-refractive-index glass (namely Ohara S-TIH6) is predicted to pump it more completely and uniformly than previous designs. Upon hand-fabricating such an injector and Bridgman-growing a crystal of 0.1% Pc:PTP over it, an experimental maser oscillator using this combined injector-crystal assembly is demonstrated. The performance and scaleability of multiblade injection vis-a-vis alternative strategies is analyzed.

physics.optics↗

arXiv2026

DNA-Origami-Assembled Rhodium Nanoantennas for Deep-UV Label-Free Single-Protein Detection

Nanoparticles of plasmonic metals have significantly to the development of spectroscopic techniques, enabling strong confinement of electromagnetic fields at the nanoscale and corresponding signal amplification. However, to date, plasmonic applications have been limited mainly to the visible and near-infrared range, as materials supporting ultraviolet resonances typically exhibit poor chemical stability and lack robust surface functionalisation methods. In this work, we address these limitations by introducing a fully programmable approach to UV plasmonics based on rhodium nanocube dimers assembled using DNA origami templates. We have developed a reliable ligand exchange strategy that allows the functionalisation of rhodium nanocubes with DNA while maintaining their colloidal stability. These DNA-modified nanocubes act as modular building blocks that can be assembled into dimers with 69% efficiency and an average gap size of 10 nm. The DNA origami design also allows for the deterministic placement of a single streptavidin protein in the plasmonic gap, unlike previous methods based on stochastic diffusion. Experiments with single-molecule autofluorescence in UV, supported by numerical simulations, show an increase in brightness of up to 22, a reduction in fluorescence lifetime, and a more than tenfold increase in the total number of detected photons. By creating a robust and versatile platform for the production of UV-resonant plasmonic nanoantennas, this work extends the functionality of plasmonics to the deep UV spectrum and opens up new possibilities for labelling-free single-protein spectroscopy.

physics.optics↗

arXiv2025

Single-LED-pumped, room-temperature, solid-state maser

Through their ability to achieve cryogenic levels of noise performance while operating at room temperature, optically-pumped, solid-state (OPSS) masers show great promise as quantum sensors, oscillators, and amplifiers. We here demonstrate maser oscillation in a microwave cavity containing a crystal of pentacene-doped para-terphenyl (ptc:ptp) pumped by a single, chip-scale LED. Here, unlike previous work, the size of the pump source does not dominate the size of the maser system as a whole. This miniaturization is achieved through invasive optical pumping in the form of a waveguide, the tip of which is embedded into the maser crystal. Using experimental measurements combined with microwave and optical simulations, we find that our approach offers at least a factor-of-2 enhancement in cooperativity over end-on optical excitation. We use our simulations to define a figure of merit for maser pumping efficiency, and conclude that there remains significant headroom to improve the performance of ptc:ptp masers through improved optical design.

quant-ph↗

arXiv2025

XRPO: Pushing the limits of GRPO with Targeted Exploration and Exploitation

Reinforcement learning algorithms such as GRPO have driven recent advances in large language model (LLM) reasoning. While scaling the number of rollouts stabilizes training, existing approaches suffer from limited exploration on challenging prompts and leave informative feedback signals underexploited, due to context-independent rollout allocation across prompts (e.g., generating 16 rollouts per prompt) and relying heavily on sparse rewards. This paper presents XRPO(eXplore - eXploit GRPO), a unified framework that recasts policy optimization through the principled lens of rollout exploration-exploitation. To enhance exploration, XRPO introduces a mathematically grounded rollout allocator that adaptively prioritizes prompts with higher potential for uncertainty reduction. It further addresses stagnation on zero-reward prompts through an in-context seeding strategy that injects curated exemplars, steering the model into more difficult reasoning trajectories. To strengthen exploitation, XRPO develops a group-relative, novelty-aware advantage sharpening mechanism that leverages sequence likelihoods to amplify low-probability yet correct responses, thereby extending the policy's reach beyond sparse rewards. Experiments across diverse math and coding benchmarks on both reasoning and non-reasoning models demonstrate that XRPO outperforms existing advances (e.g., GRPO and GSPO) up to 4% pass@1 and 6% cons@32, while accelerating training convergence by up to 2.7X.

cs.LG↗

arXiv2025

OPPO: Accelerating PPO-based RLHF via Pipeline Overlap

Proximal Policy Optimization (PPO)-based reinforcement learning from human feedback (RLHF) is a widely adopted paradigm for aligning large language models (LLMs) with human preferences. However, its training pipeline suffers from substantial inefficiencies due to sequential multi-model dependencies (e.g., reward model depends on actor outputs) and long-tail response lengths, where a few long responses straggle the stage completion. We present OPPO, a novel, lightweight, and model-agnostic PPO-based RLHF framework that improves training efficiency by overlapping pipeline execution. OPPO introduces two novel techniques: (1) Intra-step overlap, which streams upstream model outputs (e.g., actor model) in right-sized chunks, enabling the downstream model (e.g., reward) to begin prefill while the upstream continues decoding; and (2) Inter-step overlap, which adaptively overcommits a few prompts and defers long generations to future steps, mitigating tail latency without discarding partial work. OPPO integrates easily with existing PPO implementations with a lightweight wrapper. Extensive evaluations show that OPPO accelerates PPO-based RLHF training by $1.8\times$--$2.8\times$ and improves GPU utilization by $1.4\times$--$2.1\times$ without compromising training convergence.

cs.LG↗

arXiv2025

CORRECT: COndensed eRror RECognition via knowledge Transfer in multi-agent systems

Multi-agent systems (MAS) are increasingly capable of tackling complex real-world tasks, yet their reliance on inter-agent coordination, tool use, and long-horizon reasoning makes error recognition particularly challenging. Minor errors can propagate across agents, escalating into task failures while producing long, intertwined execution trajectories that impose significant costs for both human developers and automated systems to debug and analyze. Our key insight is that, despite surface differences in failure trajectories (e.g., logs), MAS errors often recur with similar structural patterns. This paper presents CORRECT, the first lightweight, training-free framework that leverages an online cache of distilled error schemata to recognize and transfer knowledge of failure structures across new requests. This cache-based reuse allows LLMs to perform targeted error localization at inference time, avoiding the need for expensive retraining while adapting to dynamic MAS deployments in subseconds. To support rigorous study in this domain, we also introduce CORRECT-Error, a large-scale dataset of over 2,000 annotated trajectories collected through a novel error-injection pipeline guided by real-world distributions, and further validated through human evaluation to ensure alignment with natural failure patterns. Experiments across seven diverse MAS applications show that CORRECT improves step-level error localization up to 19.8% over existing advances while at near-zero overhead, substantially narrowing the gap between automated and human-level error recognition.

cs.MA↗

arXiv2025

A Conformal Prediction Framework for Uncertainty Quantification in Physics-Informed Neural Networks

Physics-Informed Neural Networks (PINNs) have emerged as a powerful framework for solving PDEs, yet existing uncertainty quantification (UQ) approaches for PINNs generally lack rigorous statistical guarantees. In this work, we bridge this gap by introducing a distribution-free conformal prediction (CP) framework for UQ in PINNs. This framework calibrates prediction intervals by constructing nonconformity scores on a calibration set, thereby yielding distribution-free uncertainty estimates with rigorous finite-sample coverage guarantees for PINNs. To handle spatial heteroskedasticity, we further introduce local conformal quantile estimation, enabling spatially adaptive uncertainty bands while preserving theoretical guarantee. Through systematic evaluations on typical PDEs (damped harmonic oscillator, Poisson, Allen-Cahn, and Helmholtz equations) and comprehensive testing across multiple uncertainty metrics, our results demonstrate that the proposed framework achieves reliable calibration and locally adaptive uncertainty intervals, consistently outperforming heuristic UQ approaches. By bridging PINNs with distribution-free UQ, this work introduces a general framework that not only enhances calibration and reliability, but also opens new avenues for uncertainty-aware modeling of complex PDE systems.

cs.LG↗