Steinke and Zakynthinou(2020) introduces the Conditional Mutual Information (CMI) framework of analyzing the information complexity of learning algorithms based on algorithm-dependent information-theoretic quantities. We study one of these quantities, the evaluated Conditional Mutual Information (eCMI). It has been an interesting question whether the optimal PAC guarantee for VC classes can be recovered from the algorithm-dependent analyses via CMI. And we show that it is possible to recover this guarantee by constructing a learning algorithm whose eCMI is of order O(d) in the realizable case, where d is the VC-dimension of the concept class. Specially, our algorithm is a randomized Majority-of-5 base learners with optimal in-expectation generalization guarantee.
Integrating recall and pre-ranking in e-commerce search requires candidate generation to account for relevance, personalization, and business value before final ranking. To this end, we present VARG, a generative retrieval system for Tmall App search that directly admits generated item candidates to the existing final ranker. VARG-ID constructs semantic prefixes using RQ-VAE, enhances search relevance through bidirectional query-item contrastive learning, and combines these prefixes with a value-ordered third token to provide fine-grained item addresses and a business-value prior. Three-stage supervised fine-tuning progressively learns item-to-identifier mappings, query-semantic retrieval, and personalized retrieval. Personalized model training combines value-aware and hierarchy-aligned supervision with expanded user context, and uses local ordinal supervision (LO-SFT) to learn the local within-cluster ordering encoded by the third token. Prefix-GRPO combines gated rewards based on output legality, user behavior, ranker advantage, and search relevance with prefix-aware token weighting to align candidate generation with business value and ranking objectives. Coordinated daily product and model updates preserve existing item addresses while incorporating new products and behavioral feedback. Offline experiments on tens of millions of products validate identifier stability and demonstrate gains in retrieval quality and head-level value recall from SFT strategies and Prefix-GRPO over their respective baselines. In a 14-day online A/B test covering 20% of search traffic, VARG directly admits generated candidates to the final ranker and improves GMV by 1.45%, per-user IPV by 0.22%, and PCTR by 0.31%. Online shopping-guide query evaluations further show that VARG maintains competitive relevance with a smaller candidate quota.
Eugenio Cruz-Ossa, Sebastian Perez-Salazar, Victor Verdugo
原文摘要
Competition complexity formalizes a compelling intuition: rather than refining the mechanism, how much additional competition is sufficient for a simple mechanism to compete with an optimal one? We begin the study of this question in multi-unit pricing for welfare maximization using prophet inequalities. An online decision-maker observes $m \geq k$ nonnegative values drawn independently from a known distribution, may select up to $k$ of them, and aims to maximize the expected sum of selected values. The benchmark is a prophet who observes a sequence of length $n \geq k$ and selects the $k$ largest values. We focus on the widely adopted class of single-threshold algorithms and fully characterize their $(1-\varepsilon)$-competition complexity. Notably, our results reveal a sharp competition-induced phase transition: in the absence of competition, single-threshold algorithms are fundamentally limited to a $1-1/\sqrt{2kπ}$ fraction of the prophet value, whereas even a $1\%$ multiplicative increase beyond $n$ observations suffices to achieve a $1-\exp(-Θ(k))$ fraction. Another notable result happens when $k=1$: we show that the $(1-\varepsilon)$-competition complexity is exactly $\ln(1/\varepsilon)$, fully resolving an open question by Brustle et al. [Math. Oper. Res. 2024]. Our analysis is based on infinite-dimensional linear programming and duality arguments.
We analyze a mathematical property of the popular card game SET. We present a revised geometric and computer-assisted proof that the maximum number of sets for 12 cards is 14 using finite field geometry, combinatorics, and graph theory. We present several computer algorithms for finding or approximating the maximum number of sets. We show a complete computer proof that iterates over all possible board configurations. We use this method to compute the maximum number of sets with 4 properties for a small number of cards, but it is generally too inefficient. With this method, we compute the maximum number of sets for 3 properties for all possible numbers of cards. Finally, we compare this to an existing approximation algorithm for finding maximal sets.
Zhihao Zhang, Gengwei Zhang, Tianlong Chen, Xiaoming Liu
原文摘要
Monocular 3D object detection spans two regimes: closed-set detectors operating within a fixed category vocabulary, and open-vocabulary detectors that localize arbitrary categories by leveraging depth foundation models for 3D geometry. We find that current depth foundation models, despite their strong zero-shot generalization, lack the object-level precision 3D detection demands: substituting a state-of-the-art depth foundation model for a strong detector's predicted depth degrades accuracy. Rather than pushing detectors or depth models to be more accurate end-to-end, we treat object-level depth refinement as a stand-alone task and present RefineAny3D, a vision-language model that corrects depth without ever predicting a numerical value. Our key insight is that depth error has a direct visual signature in image space: when projected onto the image, a correctly placed box tightly encloses the object, while a too-far box projects too small and a too-close box projects too large. Depth refinement thus reduces to a visual alignment problem rather than a metric regression problem, which we instantiate by extending the VLM's vocabulary with action tokens that replace numerical depth output with categorical decisions, and by supervising the model on a large-scale chain-of-thought dataset that grounds each decision in explicit visual evidence. Applied as a single post-hoc step, RefineAny3D delivers consistent gains across closed-set detectors, open-vocabulary detectors, and 3D auto-labeling tools, and generalizes to novel categories, scenes, and cameras without retraining. Code is available at https://github.com/alanzhangcs/RefineAny3D.
Image super-resolution (SR) with large generative models has recently achieved remarkable perceptual quality, yet maintaining fidelity to the LR observation remains challenging. In particular, we observe that diffusion transformers (DiTs) built on latent representations suffer from a critical limitation: the compression bottleneck of the VAE weakens fine-grained spatial information, leading to hallucinated details that are weakly grounded in the input image. In this work, we revisit generative SR from a representation perspective and propose a pixel-grounded super-resolution (PGSR) framework that preserves LR-observed pixel evidence before VAE compression and reuses it throughout restoration. Instead of relying solely on the compressed latent condition, PGSR extracts pre-VAE pixel evidence from the upsampled LR image and reuses it at two stages. First, Condition-Side Trajectory Guidance fuses LR-derived pixel evidence with the latent LR condition to guide the latent restoration trajectory. Second, Decoder-Side Pixel Grounding injects multi-scale pixel features into the frozen VAE decoder to ground the final rendering with LR-observed cues. To efficiently adapt large pretrained DiT models, we keep the latent autoencoder and main flow-matching backbone frozen, and train only lightweight restoration modules. We further study an efficient local-window attention variant for improved high-resolution efficiency and scalability. Extensive experiments demonstrate that PGSR improves the realism--fidelity trade-off and produces more faithful, visually convincing results than existing latent generative SR approaches.
Shane E. Loeffler, Zan Ahmad, Shiyi Chen, Syed Yusuf Ali, Carolyna Yamamoto, Kensuke Sakata, Dan M. Popescu, Alana Yee 等 11 位作者
原文摘要
High-fidelity simulations of time-dependent nonlinear partial differential equations (PDEs) are expensive to repeat across families of domains and parameter fields, a cost that recurs in design optimization, uncertainty quantification, and computational medicine. Most neural-operator surrogates require each domain to be mapped onto a common grid or reference shape, and few predict the full time course. We introduce Graph Fourier Neural Kernels (G-FuNK), a neural operator for second-order semilinear PDEs with a diffusive principal part that transfers across geometries and anisotropic diffusion tensors without coordinate maps or diffeomorphisms. A weighted graph and its Laplacian supply a spectral basis that encodes geometry and anisotropy. Shared layers act in this basis, at a cost linear in mesh size after one partial eigendecomposition per geometry, and a neural ODE integrator propagates the learned generator to full spatiotemporal trajectories on unseen domains. Across anisotropic heat flow, stiff reaction-diffusion, FitzHugh-Nagumo, and patient-specific left-atrial electrophysiology on CT-derived surfaces, G-FuNK reproduces held-out trajectories with relative $L^2$ error between 0.046 and 0.160. Its accuracy is comparable to that of a geometry-aware Fourier neural operator on regular domains, and its predictions are unchanged when a rectangular domain is rotated. On left atria, whose venous and valvular openings prevent mapping onto a reference shape, it predicts an 85-ms trajectory in under half a second, against about 5 minutes for finite elements, with a mean activation-time error of 1.73 ms. Accuracy depends on graph construction and is highest when the learned generator is integrated at its training time step.
While many advances in time series models focus exclusively on numerical data, research on multimodal time series, particularly those involving contextual textual information, remains in its infancy. With recent progress in large language models and time series learning, we revisit the integration of paired texts with time series through the Platonic Representation Hypothesis, which posits that representations of different modalities converge to shared spaces. In this context, we identify that time-series-paired texts may naturally exhibit periodic properties that closely mirror those of the original time series. Building on this insight, we propose a novel framework, Texts as Time Series (TaTS), which considers the time-series-paired texts to be auxiliary variables of the time series. TaTS can be plugged into any existing numerical-only time series models and effectively enable them to handle time series data with paired texts. Through extensive experiments on both multimodal time series forecasting and imputation tasks across benchmark datasets with various existing time series models, we demonstrate that TaTS can enhance multimodal predictive performance without modifying model architectures. Our Code is available at https://github.com/iDEA-iSAIL-Lab-UIUC/TaTS.
Forough Majidi, Mohammad Mehdi Morovati, Foutse Khomh, Heng Li
原文摘要
Large Language Models (LLMs) are integrated into software systems and AI services, making efficient LLM serving a concern for software engineering. Serving LLMs is challenging because inference requires computation, memory, GPU resources, and execution while maintaining latency and throughput. Although prior research has proposed LLM inference, optimization, and serving techniques and frameworks, little is known about how they are adopted in practice. In this study, we investigate the use of LLM serving frameworks and serving methods in open-source software systems. We identify and analyze five LLM-specific frameworks: vLLM, SGLang, TensorRT-LLM, LMDeploy, and FlashInfer. We examine how these frameworks and techniques are adopted individually and in combination, how adoption varies across categories of LLMs, and how repositories differ in intent, focus, use case, and architectural design. Our results show that vLLM is the most visible framework in popularity and adoption, while parallel computation, memory management, and network pruning are the most frequently used serving-method categories. Multi-framework usage is limited, suggesting that developers rely on a single serving framework; however, combined frameworks connect complementary capabilities across the serving stack. Framework adoption varies across model families, modalities, model sizes, domain specializations, and deployment settings. Repository-level analysis shows that LLM serving frameworks support applications and architectures, including Reinforcement Learning (RL)-based reasoning, multimodal generation and understanding, microservices, and cloud infrastructure. Overall, this study provides a large-scale empirical characterization of LLM serving framework adoption in practice and offers insights for researchers, framework maintainers, and practitioners working on LLM systems.
Received OFDM grids preserve both pilot and payload observations, yet wireless foundation models are rarely pre-trained on them. We introduce LWM-TF, a 9.58M-parameter factorized masked autoencoder pre-trained on complex 5G NR uplink grids taken after the FFT and before channel estimation. Each block attends first across OFDM symbols and then across resource blocks and receive antennas, and the mask follows the same two axes. Because the payload phase is unpredictable, the model reconstructs the normalized log-power of each masked token and predicts its scale separately; ablations show that the two terms are complementary and that this target outperforms complex-signal reconstruction. The encoder is pre-trained on 850,496 physical uplink shared channel (PUSCH) slots from a 10^6-slot dataset simulated with the NR transmit chain under various channel conditions. With the encoder frozen, linear probes classify modulation, predict CRC success, and detect UE mobility. On a separately generated test set, 16 labeled slots per class give 0.85 modulation accuracy, 0.82 CRC area under the ROC curve (AUROC), and 0.73 mobility AUROC, ahead of a supervised CNN on every paired draw. With 10,000 labels the probes reach 0.94, 0.96, and 0.92, well above engineered receiver statistics, and detect mobility where the CNN stays at chance. Without any impairment labels, distances in the frozen feature space flag unseen waveform impairments with 0.78 macro AUROC, against 0.68 for receiver statistics. A single pre-trained encoder thus provides receiver features that transfer across tasks, most strongly when labels are scarce.
Michael Skotnica, Marek Pecha, Jana Rušajová, Bohdan Rieznikov, Vít Wandrol, Markéta Rösnerová, Jaromír Knejzlík
原文摘要
The north-eastern part of the Czech Republic, including the Czech part of the Upper Silesian Coal Basin, is among the most seismically active areas of the country. The seismic stations in the region record mining-induced events, which continue after the cessation of mining, natural earthquakes, which are rare but may originate in the same area, and quarry blasts. A rapid and reliable distinction between the three types is required for the routine evaluation of the continuously recorded data. We study their automatic discrimination with neural networks from short three-component windows starting one second before the P-wave onset. The main data set is the archive of the Seismic Polygon Frenštát (1992--2002), a triggered network whose records were recovered from the original format, cleaned with deterministic and learned quality checks and completed with automatically picked P-wave arrivals; it contains about 40,000 records of all three event types. Among five compared architectures, convolutional networks operating on short-time spectrograms outperform recurrent networks operating on magnitude spectra or on the raw waveform. The best model, EfficientNet-B0 applied to 24-s windows, reaches a macro-averaged F1 score of 0.96, and blasts and natural earthquakes are practically never confused with each other; class weighting and undersampling, the usual remedies for the strong class imbalance, lower the score. Finally, we examine the transfer to the station OKC, which monitors the region today with different instruments. A model trained on the archive alone fails on the modern records, whereas a single model trained on both data sets reaches a macro-F1 score of 0.96 on the archive and 0.92 on the modern station, where the accuracy is limited by the small number of catalogued natural earthquakes rather than by the model.
Best-of-N selection for code generation is only as good as the judge that picks among the candidates, and the judges accurate enough for this role are costly frontier LLMs. A small judge can be trained on abundant earlier data labeled by unit tests, but it is deployed on newer problems: the earlier problems are easier and teach it little, their tasks differ, and their candidates come from weaker policies, while data close to deployment are scarce. We present JURIST, a 1.7B reasoning judge that uses a small meta set close to deployment to learn how much each earlier training sample should count. In our experiments, we find that training on the meta set alone, or adding it to the training data, does not help, while using it to weight the earlier data through a bi-level objective adds 3.9 Hard points over adding it to the training data. On LiveCodeBench, JURIST reaches 84.7% Overall and matches Claude Sonnet 4.6 on Hard problems (73.8%) at 463x lower cost; on BigCodeBench, it surpasses all test-time scaling baselines (35.9%). Abundant earlier data, guided by a small deployment-aligned meta set, can thus train a compact judge that selects accurately at low cost.
When a cooperative team receives feedback only at the end of an episode, a single outcome must guide many agents' earlier actions. Learning a dense team reward helps, but its decomposition does not automatically give each agent a useful policy update. We introduce IMAP, an online method that fits a shared Q/V reward model by comparing completed episodes according to their observed terminal outcomes. Local Q/V residuals provide agent-specific rewards; the same normalized weights combine them for a centralized critic and scale local advantages for decentralized PPO actors. We characterize when these local updates follow the learned team objective. With full returns and frozen reward heads, a sharp bound relates their gradient error to variation in future Q/V compatibility; an exact identity identifies additional continuation terms for general GAE. In experiments with terminal-only feedback, IMAP's local-actor configuration attains the highest reported mean among the compared methods on all nine SMACv2 scenarios and four MAMuJoCo tasks, and on two of three SMACLite scenarios. Controlled relay interventions confirm the predicted changes in actor credit while holding the learned team reward fixed. The gradient result concerns the fitted objective; benchmark returns measure performance on the original task outcomes.
Matthew Smart, Soumya Ganguly, Nilava Metya, Alexandre V. Morozov, Anirvan M. Sengupta
原文摘要
We study minimal attention-only transformers under all-token corruption and show they admit a two-stage empirical Bayes interpretation. A single attention step computes a kernel-weighted posterior mean with respect to the empirical distribution defined by the context. Depth refines this distribution through particle dynamics (Stage 1), while a long-range skip-connection carries the noisy input as a query for posterior inference (Stage 2), revealing distinct statistical roles for depth and attention residuals. The framework isolates a minimal setting in which the context itself induces a depth-dependent energy landscape governing in-context inference. We show that effective denoising can emerge without an explicit noise schedule: a fixed kernel bandwidth and finite integration horizon suffice, yielding a principled depth-noise relationship. We further establish a posterior-mean recovery guarantee for Gaussian priors, showing the empirical estimator converges to the Bayes-optimal predictor under sequential asymptotic conditions. Connecting these dynamics to reverse-diffusion limits, our results provide a statistical interpretation of attention as in-context inference via sample-based posterior estimation, without explicit density modeling.
Prominent AI safety discussions emphasize existential threats and catastrophic events, such as AI-enabled nuclear conflict or engineered biological outbreaks. This position paper argues that there is another type of AI risk that arises and accumulates from the widespread use of AI. Drawing on lessons from social media, we argue that safety assessments must also address accumulative harms that develop through repeated use, recurring decisions, and feedback across interconnected deployments. Repeated reassurance, selective exposure, recurring decision biases, and growing personal profiles may gradually narrow perspectives, weaken independent judgment, or enable exploitative influence without a single clearly harmful event. We advance the Trajectory Neglect Thesis: a focus on vivid catastrophic events and identifiable incidents risks overlooking less visible harms that accumulate through the widespread daily use of AI. Social media research motivates this position by showing how reciprocal media effects, social reinforcement, and individual susceptibility shape outcomes beyond individual encounters. We examine how related mechanisms may operate in AI systems, using conversational AI as a concrete setting where generation, semantic inference, and persistent memory may tighten feedback loops. Sycophancy, manipulation, and privacy loss provide a case study of how these risks may interact and accumulate. This comparison identifies mechanisms to investigate rather than assuming that repeated AI use is harmful. We therefore call for evaluation of cumulative patterns of use and deployment, individual and societal outcomes, careful causal inference, and proportionate safeguards that preserve useful assistance and informed user choice.
Zhen Qin, Casey Jameson, Alireza Goldar, Michael B. Wakin, Zhexuan Gong, Zhihui Zhu
原文摘要
In this paper, we answer this fundamental question affirmatively by using a class of informationally complete positive operator-valued measures (IC-POVMs), including symmetric IC-POVMs (SIC-POVMs) and spherical $t$-designs, focusing on sample complexity while not accounting for the implementation complexity of the measurement settings. For SIC-POVMs and spherical 2-designs, we show that the number of state copies to guarantee bounded recovery error of an MPO state with a constrained least-squares estimator depends on the probability distribution of the MPO under the POVM but scales only linearly with $n$ when the distribution is approximately uniform. For general spherical $t$-designs with a sufficiently large $t$, we prove that only a number of state copies proportional to the number of independent parameters in the MPO is sufficient for a guaranteed recovery of any state represented by an MPO. In addition, we establish a sample complexity lower bound based on the Cramer-Rao bound, showing that, for strictly positive definite MPO states and unbiased estimators, a number of state copies proportional to the number of independent parameters in the MPO is also necessary, although the upper and lower bounds do not yet match in their dependence on the target accuracy. Furthermore, we propose a spectral method and a projected gradient descent (PGD) algorithm to solve the constrained least-squares problem, both equipped with rigorous recovery error guarantees.
Current multimodal large language models (MLLMs) suffer from unstable reasoning due to the lack of mechanisms for controlling reasoning trajectories, often leading to hallucinated conclusions. Our analysis reveals that early-stage uncertainty induces trajectory switching, significantly increasing hallucination risk. We propose MedMMV, an evidence-grounded trajectory control framework that treats reasoning as a search-and-selection process over candidate trajectories. By constructing, refining, and selecting among multiple reasoning paths, MedMMV improves reliability and produces structured supervision signals over trajectory quality. We further convert these signals into learnable objectives for post-training, including reject sampling and pairwise preference learning, enabling smaller models to learn trajectory selection as an efficient single-pass capability. Instead of merely imitating reasoning traces, the model is trained to assign higher likelihood to preferred trajectories. Experiments show that richer supervision leads to stronger transfer, with accuracy improving from 10.5% to 29.0%. Our results demonstrate that trajectory selection is the key transferable capability for reliable reasoning.
Knowledge graph learning offers a powerful framework for representing and inferring structured knowledge, with broad practical applications. Two obstacles limit current methods: labeled triples are scarce for each entity and relation, which makes expressive models hard to train, and scoring functions are designed ad hoc, which limits their generalizability and leaves them without theoretical grounding. We address both obstacles with a theoretically grounded framework, with end-to-end guarantees, that extends and subsumes existing methods. The framework proceeds in two stages: multi-view unsupervised pretraining that integrates heterogeneous sources of entity side information into a shared representation, followed by supervised learning across multiple relation types. We establish a unified nonasymptotic risk bound that combines the subspace recovery error of pretraining and the sample complexity of the labeled data in a single guarantee, quantifying formally how large-scale unlabeled data from multiple views benefits downstream knowledge prediction. Synthetic experiments validate each component of the theory, and experiments on large-scale real-world knowledge graph benchmarks confirm the effectiveness of our approach.
The number of spaceborne synthetic aperture radar (SAR) collectors is growing exponentially, each returning raw collections that must be formed into images. Exact backprojection needs 1.6 trillion interpolated reads and 90 trillion floating-point operations for a 108-megapixel Umbra image. How long it takes to form an image, and at what price, depends on the algorithm, the arithmetic, and the hardware. Open-source implementations have never been compared on matched hardware against a common reference, and none run on Google tensor processing units (TPUs). This paper presents FastSAR, an open-source SAR image formation library for central processing units (CPUs), NVIDIA graphics processing units (GPUs), and TPUs. With it, we compare three SAR image formation algorithms, four floating-point formats, and six cloud devices in every supported combination, with cost per image at on-demand cloud prices as the fair common measure. FastSAR is 1.06 to 22,000 times faster than the tested open-source implementations at equal or better accuracy, and the full suite of hardware, algorithm, and floating-point combinations lets a user choose the Pareto-optimal configuration for their balance of accuracy, speed, and cost. The L4 GPU forms images the cheapest overall. Larger GPUs are faster but cost more per image because their price rises faster than their speed. The TPUs come close to the GPUs in cost, but only for the factorized algorithm, and the CPU is far less efficient in both speed and cost. Lower precision formats buy less of a speedup than expected, and each step down in precision costs accuracy.
Current safety alignment of foundation models largely follows a one-size-fits-all paradigm, applying the same refusal policy across users and contexts. As a result, models may refuse requests that are unsafe for general users but legitimate for authorized professionals, limiting helpfulness in specialized professional settings. Existing approaches either require costly realignment or rely on inference-time steering that suffers from imprecise control and added latency. To this end, we propose Palette, a modular, controllable, and efficient framework that selectively relaxes refusal behavior on authorized target domains while preserving standard safety elsewhere. Our method identifies a refusal direction via multi-objective search and internalizes it into the model through lightweight adaptation. Palette further supports modular composition: it learns domain-specific safety controls independently and composes them through parameter merging, enabling on-demand multi-domain authorization without retraining. Experiments across four safety benchmarks, multiple model variants, and both LLMs and VLMs show that Palette delivers precise safety control without sacrificing general utility, offering a practical path toward foundation models that adapt to diverse professional needs.
Albert Gong, Annabelle Michael Carrell, Raaz Dwivedi, Lester Mackey
原文摘要
We introduce a new tool, Express, for converting a non-causal attention approximation into a causal approximation with matching approximation guarantees. When combined with the state-of-the-art Thinformer approximation, Express improves upon the best known causal attention guarantees, delivering $\log^{3/2}(n)/s$ approximation error with only $O(s)$ memory and $O(s^2 \log^2(n))$ compression overhead for a sequence of length $n$. We pair these developments with an efficient I/O-aware Triton implementation, demonstrate substantial speedups over FlashAttention 2, and use Express to overcome four resource bottlenecks in the language modeling pipeline: long-context prefill, KV cache compression, long-form memory-constrained decoding, and long-form compute-constrained decoding.
Low-resource automatic speech recognition (ASR) often relies on cross-lingual transfer from higher-resource languages due to limited availability of transcribed speech. However, selecting an effective donor language is a nontrivial issue, particularly in settings with spontaneous speech, linguistic variation, and uneven resource availability. We present DonorRank, a learning-to-rank framework that uses linguistic and dataset-level features along with previously observed transfer outcomes to predict promising donor languages for zero-shot ASR. We evaluate DonorRank on two multilingual speech corpora of Indic and African languages. DonorRank produces informative donor rankings across both collections, outperforming common heuristics in the more closely related Indic setting while remaining competitive in the more linguistically diverse African setting. Our analysis reveals that effective donors are not always the closest linguistic relatives and that the most informative predictors vary across language collections. These findings highlight the value of learned donor selection for prioritizing transfer experiments in low-resource ASR. We will also release all of our evaluation results, feature matrices for all languages, scripts for training the DonorRank framework, and pairwise transfer performance matrices for our datasets.
Zhongmou He, Yee Man Choi, Kexun Zhang, Jiabao Ji, Junting Zhou, Dejia Xu, Ivan Bercovich, Aidan Zhang 等 10 位作者
原文摘要
Verifiers provide important reward signals for reinforcement learning of large language models (LLMs). However, it is challenging to develop or create reliable verifiers, especially for code generation tasks. A well-disguised wrong solution program may only be detected by carefully human-written edge cases that are difficult to synthesize automatically. To address this issue, we propose HARDTESTGEN, an approach to synthesize high-quality test cases for algorithmic coding problems. We curate a comprehensive algorithmic programming dataset HARDTESTS with 26.6k problems and high-quality synthetic tests. Compared with existing tests, HARDTESTGEN tests demonstrate significantly higher accuracy in verifying LLM-generated code (+11.22 percentage points in precision, the percentage of actually correct code within the predicted correct ones). We also show that downstream post-training -- including rejection sampling and reinforcement learning (RL) -- using HARDTESTS verifier results in improved performance of LLM code generation. We open-source our dataset and synthesis pipeline at https://leililab.github.io/HardTests/.
Foundation models are post-trained with reinforcement learning (RL) to maximize specific rewards, such as helpfulness, correctness, or instruction following. This post-training process is computationally intensive, sometimes unstable, and has to be run from scratch every time the reward model changes or when we want to combine multiple rewards. We hence ask: given a new reward function, is it possible to predict the policy RL would produce without actually training on that reward? We answer this in the affirmative by introducing PoEM, a framework for predicting RL outcomes using models previously trained on other rewards. First, we build on the fact that if the new reward function can be written as a linear combination of existing ones, then the optimal policy can be expressed as a combination of the existing policies in log-probability space. Surprisingly, even in cases where the rewards are not linearly connected, we observe that policies learned through RL often span an approximately low-rank subspace across rewards. To our benefit, the weights for this combination can be estimated using only the reward or basis policy outputs on the samples. We turn these observations into an algorithm that takes post-trained models and a target reward function, and approximates the target RL policy without any additional RL training. We experimentally evaluate our approach across programmatic and learned rewards for open-ended generation, with additional experiments on math reasoning, tool calling, and image generation.
Yiheng Xiong, Luisa Gallée, Daniel Santak Wolf, Heiko Hillenhagen, Michael Götz
原文摘要
Numerous unsupervised domain adaptation (UDA) algorithms exist, but for clinical practice, selecting the best-suited one along with proper hyperparameters often remains unclear, as the unlabeled deployment (target) domain prevents direct evaluation. We propose a label-free criterion that jointly selects the algorithm and hyperparameters for UDA. Given a pool of candidate models from multiple algorithms trained with different hyperparameters, our approach scores each candidate against an agreement reference, and selects the one with the highest score. The agreement reference is constructed in two levels without using target labels. First, we leverage multiple label-free selection signals, using each to nominate a model within every algorithm. Second, the nominated models are aggregated across algorithms to form a reference prediction for each unlabeled target sample. The candidate whose predictions agree most with this reference is then selected for deployment. Experimental results on four brain MRI and four chest X-ray datasets across seven clinically relevant transfer scenarios show that our method achieves better selection performance than individual validators and remains effective across different algorithm pools. Our approach takes a step towards practical, label-free algorithm selection for clinical deployment of UDA. Our code is open sourced at https://github.com/xiongyiheng/Complete-Unsupervised-Domain-Adaptation-Pipeline/tree/main/two-level-algorithm-selection.
When used in automated decision-making systems, machine learning (ML) models are vulnerable to data-manipulation attacks. Some defense mechanisms (e.g., adversarial regularization) directly affect the ML models while others (e.g., anomaly detection) act within the broader system. However, many defenses operate under a predefined threat model that assumes characteristics of the adversary, including their knowledge, capabilities, and objectives, are known and fixed. In practice, these characteristics may be unknown or evolve over time, creating a mismatch between the assumptions of the defender and the true attacker. In this paper, we study a new perspective on adversarial defense: focusing on the attacker, rather than the attack. We present and demonstrate SHERLOC, a framework for adversarial learning forensics -- namely, a detective that learns about the attacker from observed attack(s). We prove that, without additional knowledge, the attacker is non-identifiable --- multiple potential attackers would perform the same observed attack. This reveals an important distinction between reconstructing an attacker's behavior and recovering the attacker that generated it. We demonstrate empirically that SHERLOC is able to reconstruct a hypothesis attacker that is consistent with the observed attacks against various defender models (i.e., generalized least-squares linear regressors, logistic regressors, and polynomial classifiers) across various security-critical domains. Finally, we show that these inferred attackers need not match the ground truth attacker exactly to have downstream utility. Defenses using the SHERLOC-inferred attackers achieve approximately 90-100% of robustness improvement that would be achieved if the defender had access to the ground-truth attacker.
Genoria AI Team, Yuejie Hou, Lixiang Yang, Yisi Liu, Weiting Tang, Yueyuxiao Yang, Meng Yang
原文摘要
Enzyme engineering usually optimises which candidates to assay within a fixed model and workflow; when the bottleneck is a mismatched method or a failing workflow, that route is hard to escape. We present EnzymeRSI, a dual-loop system for recursive self-improvement (RSI). The outer loop evolves harness assets without changing model weights; the inner loop develops prediction and selection algorithms under controlled conditions; a cloud laboratory system is specified for closing the physical loop with quality-controlled feedback. The outer loop raised the mean development-set Spearman correlation on ProteinGym by 0.044 across five assays, reaching 0.335 on CD19 and 0.683 on OPSD (rhodopsin); of the two first places, only CD19 was produced by the outer loop itself. In ThermoMPNN self-improvement the inner loop reached a PCC of 0.849, and on TadA-Bench a Spearman correlation of 0.307 in 1.13 h. We examined the system's experimental decision-making in three enzyme systems. For the luciferase RLuc, 10 of the 32 nominated substitutions not pre-specified reached at least twice wild-type activity in a single round. For the cytidine deaminase A3A, a variant selected by a rule that caps loss of unmodified C raised 5mC conversion in the weaker CT context from 11.4% to 46.5% and reversed the substrate preference in both contexts. For the DNA polymerase KOD-E, 189 single-substitution variants defined a landscape on which different objective functions selected divergent next steps from the same parent; one assayed combination reached 31.1 times the parent on a modified substrate. By treating the organisation of research and the development of algorithms as separately scored objects of evolution, and linking a frozen decision record to a physical measurement, this work offers an evolutionary paradigm for moving beyond the blind spots of a fixed method and for autonomous enzyme engineering.
We study the problem of differentially private continual counting in the unbounded setting where the input size $n$ is not known in advance. Current state-of-the-art algorithms based on optimal instantiations of the matrix mechanism cannot be directly applied here because their privacy guarantees only hold when key parameters are tuned to $n$. Using the common `doubling trick' avoids knowledge of $n$ but leads to suboptimal and non-smooth error. We solve this problem by introducing novel matrix factorizations based on logarithmic perturbations of the function $\frac{1}{\sqrt{1-z}}$ studied in prior works, which may be of independent interest. The resulting algorithm has smooth error, and for any $α> 0$ and $t\leq n$, it is able to privately estimate the sum of the first $t$ data points with $O(\log^{2}(t)(\log\log t)^{1+2α})$ variance. It requires $O(t)$ space and amortized $O(\log t)$ time per round, compared to $O(\log(n)\log(t))$ variance, $O(n)$ space and $O(n \log n)$ pre-processing time for the nearly-optimal bounded-input algorithm of Henzinger et al. Empirically, we find that our algorithm's performance is also comparable to theirs in absolute terms: with optimized parameters, our variance is less than $2\times$ theirs for $t$ as large as $2^{32}$.
Shoaib Sadiq Salehmohamed, Jinal Prashant Thakkar, Hansika Aredla, Shaik Mohammed Omar, Naveen Bera, Pulijala Sai Nikhila, Kondaguduru Abhiram, Shaik Gayaz Ali 等 9 位作者
原文摘要
Existing hallucination detection methods for large language models (LLMs) rely on external verification at inference time, requiring gold answers, retrieval systems, or auxiliary judge models. We ask whether this external supervision can instead be distilled into the model's own representations during training, enabling hallucination detection from internal activations alone at inference time. We introduce a weak supervision framework that combines three complementary grounding signals: substring matching, sentence embedding similarity, and an LLM as a judge verdict to label generated responses as grounded or hallucinated without human annotation. Using this framework, we construct a 15000-sample dataset from SQuAD v2 (10500 train/development samples and a separate 5000-sample test set), where each example pairs a LLaMA-2-7B generated answer with its full per-layer hidden states and structured hallucination labels. We then train five probing classifiers: ProbeMLP (M0), LayerWiseMLP (M1), CrossLayerTransformer (M2), HierarchicalTransformer (M3), and CrossLayerAttentionTransformerV2 (M4), directly on these hidden states, treating external grounding signals as training-time supervision only. Our central hypothesis is that hallucination detection signals can be distilled into transformer representations, enabling internal detection without any external verification at inference time. Results support this hypothesis. Transformer-based probes achieve the strongest discrimination, with M2 performing best on 5-fold average AUC/F1, and M3 performing best on both single-fold validation and held-out test evaluation. We also benchmark inference efficiency: probe latency ranges from 0.15 to 5.62 ms (batched) and 1.55 to 6.66 ms (single sample), while end-to-end generation plus probe throughput remains approximately 0.231 queries per second, indicating negligible practical overhead.
Han Li, Lingxiang Hu, Jiacheng Huang, Ziqian Jiang, Jingkai Luo, Wei Gao, Yunfan Tan, Zun Wang 等 9 位作者
原文摘要
Large language model (LLM) coding agents have advanced test generation across diverse programming tasks. However, the common practice of evaluating tests against a single reference solution overlooks alternative valid implementations and can overstate test quality. We introduce TestPrism, comprising 300 test tasks from 17 sources and 3000 candidate implementations, evenly split between valid and invalid solutions. Its primary metric, Joint Success Function, requires the generated tests to fail on the initial program state, accept every valid candidate, and reject every invalid candidate. Across fourteen baseline coding agent configurations, Joint Success Function reaches only 28.00%, whereas single reference success reaches 59.67%. Our analysis reveals missed behaviors, unsupported assertions, and faulty test construction. To address these weaknesses, we introduce TestHelix, which combines heterogeneous synthesis of test and repair pairs with peer cross validation and recursive self improvement (RSI). Across two models, TestHelix improves Joint Success Function by 8.67 to 9.00 percentage points over the native harness comparators in the TestHelix evaluation