跳到正文
研踪PaperTrail

论文库

检索计算机领域论文,按研究方向、作者或关键词筛选。

部分采集异常
正在读取账号与研究记录…

全站更新概况

+197新增论文
+151其中 AI 相关
+16已有论文修订
+91新增 AI 导读

arXiv 最近完成扫描:2026-11-21 12:19 · 全站最近入库:2026-11-21 12:19各信源进度 →

统计口径与数据新鲜度

统计 2026-11-21 04:53 — 2026-11-22 04:53(北京时间) · 点击数字查看对应简报;AI 相关包含在新增论文内。

最近完成检查:2026-11-21 12:19最近发现新增版本:2026-11-21 12:19库内论文最新更新时间:2026-11-20 02:59

每分钟刷新已保存的数据。空检查不代表采集异常;新收录可能包含历史论文。AI 导读按新增版本计数。

29,158 已收录论文24,181 AI 相关8,992 有修订112 HF 社区发现

浏览论文库

时间范围跟随排序:版本更新、本站收录或首次发表。最新收录可能包含历史论文。部分信源异常,查看状态 →

找到 21,490 篇论文 · 按版本更新时间排序上轮扫描起点 2026-11-21 12:17(北京时间)
新论文 · v1arXiv:2611.21449

Replicating and Extending Weight-Space Programming: Frontier LLMs Compile Programs into Weights and Read Input Layers

Roksana Goworek, Alexander Hickey, Federico Nanni, Martin O'Reilly

原文摘要

Researchers at MIT recently reported that GPT-6 Astra, a frontier large language model (LLM), can write the weights of small neural networks directly in a single reply, and can read the raw weights of small networks to say what they were trained to do. In a three-day sprint, we replicated and extended these results. Our four findings: (1) Writing weights reproduces. Astra's networks for the game Breakout scored 514, 550 and 307 (versus MIT's 517), and scored just as well on two modified versions of Breakout. Claude Opus 5.5 can also write working networks, but often runs out of output tokens. (2) What is new is translating programs into weights. Earlier models could write a good policy as code a generation before they could write network weights, and the networks the models write behave like compiled programs: two models independently decompiled Astra's Breakout network into Python matching its actions in all tested states. On classifying handwritten digits, Astra's networks scored 50% on average, below a simple classifier (82%). (3) Reading weights is mostly reading the input layer. Astra and Opus 5.5 identify which game a network plays almost perfectly, but on three new games we designed and trained ourselves, Astra recovered where things are on the board, while mostly missing how the game works. (4) Auditing from weights alone is possible but shallow. Without executing code, Astra picked out a small language model poisoned during training in 10 of 10 runs, one poisoned by extra training afterwards in 7 of 10, and a malicious adapter in 23 of 32 cases, but never identified the trigger or malicious behaviour. Open-weight models lag behind. From this early study, we conclude that LLMs can read and write small networks, but that this currently amounts to compiling hand-designed programs and reading the layout of the input layer. It is a promising but early tool for interpretability.

cs.CL · cs.AI

版本更新 · v2arXiv:2610.16937

Making Tensor Decoupling More Robust and Less Parametric

Maxime Rochkoulets, Joppe De Jonghe, Mariya Ishteva

原文摘要

Tensor decoupling algorithms recover compact representations of multivariate functions by combining linear factor matrices with univariate internal nonlinearities. We identify two limitations of recent methods. First, their performance is not invariant to mismatched output or gradient magnitudes, and second, their spline internals require manual tuning of their degrees of freedom. We propose two lightweight fixes. The first is a preprocessing step that rescales the input Jacobian tensor, removing both the bias towards larger magnitude outputs and the ill-conditioning of certain factor updates caused by such a mismatch. The second replaces B-spline internals by penalized splines coupled with generalized cross-validation, making the method less dependent on the user selecting the right number of control points. Finally, we release an efficient open-source JAX implementation of our algorithm.

cs.LG · eess.SP

版本更新 · v2arXiv:2610.15728

When Watermarks Turn Adversarial: Exposing Vulnerabilities in Document Understanding

Qingzhe Han, Fan Dang, Shengqi Chen, Xinyi Sun, Haoyu Xiong, Lingkun Li, Di Zhang

原文摘要

Visible watermarks are common in digital documents and persist during vision-language model (VLM) processing. We show that they can also serve as adversarial surfaces for document understanding. We propose Watermark-based Adversarial Surface Perturbation (WASP), a white-box attack that optimizes pixels inside fixed watermark strokes without requiring downstream questions, answers, or task-specific losses. We further introduce Balanced Aggregation Loss (BAL), which emphasizes visual token groups that remain highly similar to clean representations. Adversarial watermarks degrade performance substantially more than ordinary watermarks. On DocVQA, WASP achieves attack success rates of 96.81% and 71.93% against Qwen2.5-VL and InternVL2.5; BAL improves over a standard token hinge objective by 9.21 and 2.24 percentage points. On 1,651 OmniDocBench pages, WASP increases MinerU2.5's text edit distance by 0.7748 relative to ordinary watermarks. WASP also transfers within model families and across readers sharing a vision-tower architecture. Attack effectiveness depends on watermark geometry, model architecture, and preprocessing; in structured parsing, disrupted layout and routing can propagate errors downstream. These results expose an underexplored security risk in document understanding systems.

cs.CV

版本更新 · v2arXiv:2611.10736

Tunnels: Compressing the Solution Subgraph of Regular Path Queries

Adrián Gómez-Brandón, Gonzalo Navarro

原文摘要

Regular path queries (RPQs) are increasingly expected to return the paths that match them, not only their endpoints. Since the number of matching paths can be exponential or infinite, a standard way to deliver them (PMR) is the subgraph $\Hs$ of the product between the graph and the query automaton that contains exactly the nodes lying on some solution path. We show that $\Hs$ can be compressed much further when the graph is available to the decoder: most of $\Hs$ can be regenerated by re-running small parts of the traversal, and only the places where the traversal would waste significant effort on dead branches need to be stored. We formalize this as \emph{tunnels}: regions of $\Hs$ that are re-expanded from a stored entry point with a bounded amount of extra work, either an absolute budget or one proportional to the output produced. We handle the three classes of RPQs: constant-to-variable, constant-to-constant, and variable-to-variable. In the latter, we study the combination of tunnels with cutting the query at a selective articulation point of the automaton and encoding the answer as a union of Cartesian products of leftward and rightward traversals. Our experimental results show improvements of 2 to 3 orders of magnitude over the baseline space of PMR, for an acceptable increase in traversal time.

cs.DS

新论文 · v1arXiv:2611.21441

Learning to Reason with Predictive Gains for Generative Recommendation

Zhuoyang Li, Teng Shi, Jiakai Tang, Weijie Yu, Guofu Xie, Xiao Zhang, Lei Kou, Yang Yu 等 13 位作者

原文摘要

Reasoning-enhanced generative recommendation (GR) has attracted increasing attention for leveraging the reasoning capabilities of large language models (LLMs) to improve next-item prediction. However, existing methods primarily optimize reasoning through Group Relative Policy Optimization (GRPO) with rule-based rewards determined by final recommendation performance, which faces two key limitations: (i) unmeasured predictive contributions within reasoning, where the predictive contributions of individual reasoning steps are not directly reflected in optimization signals, and (ii) weakly discriminative learning signals within groups, where different reasoning trajectories within the group often receive indistinguishable rewards, limiting effective policy optimization. In this paper, we propose GainReason, a simple yet effective training framework that optimizes reasoning through Predictive Gain (PG). Specifically, PG measures changes in target Semantic ID (SID) log-likelihood induced by individual reasoning sentences, providing contribution-aware and discriminative rewards. We further theoretically establish the alignment between expected PG and the recommendation ranking objective. Building on this signal, we develop PG-based Reasoning Optimization (PGRO) to optimize reasoning generation with sentence-level rewards, supported by a custom causal attention mechanism for efficient PG computation. Additionally, an auxiliary PG-based Reasoning Utilization (PGRU) task adaptively strengthens reasoning-conditioned prediction to complement PGRO. Extensive experiments on three public benchmarks demonstrate that GainReason consistently outperforms competitive baselines and achieves substantial performance gains from reasoning.

cs.AI

新论文 · v1arXiv:2611.21439

A Reinforcement Learning Method to Navigate SMEFT Operator Space

Jacky Kumar

原文摘要

In particle physics, discrepancies between experimental measurements and Standard Model (SM) predictions are commonly interpreted within new theoretical models. In this context, the Standard Model Effective Field Theory (SMEFT) provides a general framework encompassing a broad class of models. However, exploring the enormous space of possible SMEFT operators rapidly becomes a challenging combinatorial problem: many operators can contribute simultaneously to different observables, their effects are correlated, and distinct operator combinations can provide comparably good descriptions of the data. In this work, we formulate the exploration of the SMEFT operator space as a sequential decision-making problem using reinforcement learning (RL). Our RL agent employs a lightweight actor-critic architecture with approximately 40K trainable parameters in total. Rather than searching for a single best-fit theory, our method is designed to discover a diverse set of high-quality theory hypotheses. We test the method in a complex phenomenological environment comprising 132 observables across several sectors of particle physics, including 24 measurements that deviate from their SM predictions by more than $1.5σ$. We show that the learned RL policy efficiently constructs multi-operator theories that substantially improve the global fit. Comparisons with independent random-subset and random-sequential searches demonstrate the ability of the RL policy to navigate this large combinatorial space. Our results establish RL as a systematic framework for exploring diverse patterns of new physics (NP) interactions within SMEFT.

cs.LG · cs.AI · hep-ex · hep-ph

版本更新 · v3arXiv:2610.04337

FLAT: Smoothing the Rugged Landscape for Learnable, Sample-Efficient Traffic Calibration

Haopeng Deng, Thomas He, Dayuan Wang

原文摘要

Calibrating microscopic traffic models for urban digital twins is an expensive black-box optimization problem: evaluating each set of car-following and lane-changing parameters requires a full simulation, limiting the evaluations available per recalibration window. Matching raw trajectories yields a rugged objective that sparse surrogates struggle to learn, reducing sequential acquisition to near-random probing. We present FLAT, which couples what to optimize with where to sample next. An eight-dimensional behavioral fingerprint smooths the parameter-error landscape, making the objective learnable from a few dozen samples; annealed lower-confidence-bound (LCB) acquisition then spends each remaining run where it most reduces error. The LCB loop supports a Gaussian process (GP), random forest (RF), or multi-layer-perceptron (MLP) ensemble. Across six heterogeneous real-world scenes, FLAT-GP achieves the lowest scene-averaged behavioral error, winning 6/6 scenes against SPSA, GA, and CMA-ES and 5/6 against TPE under the matched 100-run budget. Some baselines require up to 4.4 times as many simulations to match. Ablations show objective choice shifts final behavioral error by 80.6% on average, removing sequential LCB raises the six-scene mean by 21.7%, and surrogate choice shifts it by at most 4.2%, confirming gains trace to objective geometry and sequential allocation rather than surrogate capacity.

physics.soc-ph · cs.CE · cs.LG

版本更新 · v2arXiv:2610.07997

Can phenotypic activity be predicted without experimental readouts?

Télio Cropsal, Rocío Mercado

原文摘要

Molecular encoders contrastively pretrained on paired molecule-morphology data, such as CLOOME and CellCLIP, have been proposed as cheap surrogates for phenotypic prediction, avoiding the need to run a Cell Painting assay. We evaluate this idea under a protocol designed to control for two confounds that can inflate apparent performance: leakage across an encoder's own pretraining boundary, and the correlation between phenotypic activity and simple cytotoxicity. Testing six representations, including a non-pretrained MLP control matching CLOOME's input and layer count, on two Cell Painting screens of different scale and origin, we find that once these confounds are controlled for, CLOOME and CellCLIP show no clear advantage over plain physicochemical descriptors, also under stricter activity definitions and with non-linear probes, and that toxicity is generally easier to predict than genuine phenotypic activity across representations. The protocol is not tied to these two models, and we suggest that leakage-aware, confound-controlled evaluation should be standard practice before phenotype-pretrained encoders are trusted as surrogates for predictive phenotypic drug discovery.

cs.LG

版本更新 · v2arXiv:2611.16591

Singularity-Bench: Can Agents Build Agents?

Yang Zhou, Krish Agarwal, Zhuoming Chen, Zerui Wang, Zhenyu Zhang, Wenshuang Li, Xinpeng Wei, Zhichen Zeng 等 19 位作者

原文摘要

Frontier AI agents increasingly succeed on realistic, long-horizon tasks and are beginning to improve the systems through which they operate. Yet existing benchmarks often restrict agents to narrow action spaces, use simplified environments, or rely on verification methods that are unreliable for evaluating arbitrary agent-generated systems. We introduce Singularity-Bench, an open-ended benchmark for autonomous AI development with six end-to-end tasks spanning pre-training, mid-training, post-training, inference, the agent harness, and cluster orchestration. Each task requires agents to implement core systems from a minimal starting point and evaluates them against mature human-engineered baselines on identical hardware. Additionally, through extensive red-teaming covering 9 of the 10 categories in the OWASP Top 10 for Agentic Applications, we expose reward-hacking vulnerabilities in the conventional Harbor framework and develop Breakwater, an extension of Harbor that blocks all attacks tested in our study. Across 11 agents built from 7 frontier models and 2 harnesses, Opus 5.5 with Claude Code matches or outperforms the human baseline on all six tasks with a Singularity-Index of 1.25, suggesting it may have the skills needed to reach the singularity. The remaining agents average only 0.60. Our vision is for Singularity-Bench to track how much of AI development agents can already automate and when they may transition from assisting to taking end-to-end responsibility for building stronger successors.

cs.AI

新论文 · v1arXiv:2611.21434

Variational Monte Carlo for Quantum Excited States via Nested Low-Rank Approximation

Minchan Jeong, J. Jon Ryu, Se-Young Yun, Gregory W. Wornell

原文摘要

The electronic Schrödinger equation encodes a molecule's electronic structure and quantum properties, but solving it scales exponentially with the number of electrons. Solving this equation with neural-network variational Monte Carlo (VMC) has reached chemical accuracy on the ground state of small molecules; extending this success to excited states is constrained by two existing paradigms. Penalty-based methods need $K$ Laplacians per step but require tuning the overlap-penalty weights against the energy gaps they aim to compute. Natural excited states VMC (NES-VMC) is penalty-free, but it combines the $K$ ansatzes into one determinant, whose local energy requires $K^2$ Laplacians per step. We resolve this trade-off via low-rank approximation (LoRA) reparameterization of the rank-$K$ variational objective. The reparameterized objective admits a VMC-compatible stochastic gradient estimator via MCMC sampling from the per-state Born densities, and a sequential nesting scheme recovers the lowest $K$ eigenstates in eigenvalue order. The resulting algorithm, NestedLoRA-VMC, inherits a global optimality guarantee from the LoRA principle. On the first-row atoms Li through Ne at $K=10$, its accuracy is comparable to that of NES-VMC and of a state-of-the-art penalty-based method.

cs.LG · physics.chem-ph · physics.comp-ph

新论文 · v1arXiv:2611.21432

Just sort it! A note on the LP behind rank-one anti-loop edits

Matteo Fischetti

原文摘要

Weight-space programming removes the thinking loops of a small reasoning model by a rank-one edit W' = W + u v^T of an MLP output matrix, where u is a loop direction and v is the solution of a sparse Linear Program (LP) over the neuron activations. We investigate whether the LP is needed. We replace it by a sort: we take the 8 neurons whose activation has the smallest coefficient of variation, and give them equal shares of the push that the LP solution realizes. On Qwen3.5-0.8B the sorted edit scores 67.0% on 1000 problems (MATH-500 and GSM8K) against 67.4% for the LP edit and 9.5% for the unedited model; the paired difference is -0.4 points, with 95% confidence interval [-2.9, +2.1], inside a non-inferiority margin of 5 points declared before the runs. On three further sampling seeds the difference is -0.6 points, interval [-2.0, +0.8], and -0.8 points, interval [-2.3, +0.6], when the dose is set without the LP. The sort does not solve the LP, but the neurons it shares with the LP solution carry most of its mean push. The outcome is that, on this model, the edit acts as a nearly constant bias along u.

cs.LG · cs.CL · math.OC

版本更新 · v2arXiv:2610.21125

Noise-Space Reward Alignment for Few-Step Generative Models

Woocheol Shin, Taeyoung Yun, Jiwon Sung, Hyeongyu Kang, Jeongmin Park, Jaewoo Lee, Kyuil Sim, Dohyun Kim 等 10 位作者

原文摘要

Few-step generative models produce high-quality samples in one or a few evaluations. However, steering them toward desired properties remains difficult, especially when only black-box reward evaluations are available. While one could apply approaches developed for iterative diffusion and flow-based models, such as direct RL fine-tuning or test-time scaling, we find that neither transfers well to the few-step regime: direct fine-tuning suffers from severe reward over-optimization, and test-time scaling sacrifices the efficiency advantage of few-step generation. We propose Noise Policy Optimization (NPO), which explicitly targets the alignment problem for few-step generative models. The key idea is to learn an amortized flow policy over the initial noise space of the generator. The trained policy transports Gaussian noise to a reward-tilted noise distribution, whose pushforward distribution through the frozen generator is exactly the reward-tilted output distribution. By keeping the generator frozen, we can effectively utilize the knowledge encoded in the pretrained model, yielding a better reward-quality trade-off than direct fine-tuning. The amortized policy moves the search cost into training, so inference adds only a lightweight noise sampling step on top of the pretrained model. To improve sample efficiency during training, we use ELBO-based likelihood estimation that requires only sampled endpoints rather than simulated trajectories. Across image, language, and protein structure generation, NPO consistently achieves a better reward-quality trade-off than direct fine-tuning, and matches test-time search methods with 16 to 549 times lower inference cost.

cs.LG

版本更新 · v3arXiv:2512.12203

Illumination-Robust Localization around Unknown Space Objects using Multispectral Image Fusion

Eric J. Elias, Michael Esswein, Jonathan P. How, David W. Miller

原文摘要

As on-orbit operations grow, so does the need for precise navigation around unknown resident space objects (RSOs), such as spacecraft, debris, and asteroids. Simultaneous Localization and Mapping (SLAM) algorithms are often studied for mapping an RSO and estimating the inspector's relative pose, but conventional cameras struggle during eclipse, while lidar, though illumination-robust, has higher SWaP. Thermal-infrared cameras track targets through difficult lighting but lack the resolution and feature-richness of visible cameras. Images of a target satellite in low Earth orbit are simulated in both visible and thermal bands, and fused at the pixel level to combine each sensor's strengths for front-end tracking. Thermal imagery is further exploited by a novel silhouette-based visual place recognition (VPR) method that detects loop closures without training, runs entirely on CPU, and outperforms all baselines. Integrated SLAM performance is compared using visible, thermal-infrared, and fused imagery across lighting conditions, trajectories, and place recognition methods; using fused imagery for front-end tracking and our thermal-silhouette place recognition method performs best. The results show that visible and thermal modalities each provide unique information that is exploited differently at various stages of navigation. These low-SWaP sensors could ultimately enable reliable proximity operations on smaller, lower-cost spacecraft.

cs.RO · cs.CV

版本更新 · v3arXiv:2412.14291

Projected gradient methods for nonconvex and stochastic smooth optimization: new complexities and auto-conditioned stepsizes

Guanghui Lan, Tianjiao Li, Yangyang Xu

原文摘要

We present a novel class of projected gradient (PG) methods for minimizing a smooth but not necessarily convex function over a convex compact set. We first provide a novel analysis of the constant-stepsize PG method, achieving a unified iteration complexity for finding an approximate stationary point for convex and nonconvex problems. We then develop an "auto-conditioned" projected gradient (AC-PG) variant that achieves the same iteration complexity without requiring the input of the Lipschitz constant of the gradient or any line search procedure. The key idea is to estimate the Lipschitz constant using first-order information gathered from the previous iterations, and to show that the error caused by underestimating the Lipschitz constant can be properly controlled. We then generalize the PG methods to the stochastic setting, by proposing a stochastic projected gradient (SPG) method and a variance-reduced stochastic gradient (VR-SPG) method, achieving new unified complexity bounds in different oracle settings. We also present auto-conditioned stepsize policies for both stochastic PG methods and establish comparable convergence guarantees.

math.OC · cs.LG · stat.ML

版本更新 · v2arXiv:2609.39291

Physics-Informed GMDH for Adaptive Field Reconstruction

Mykhailo Minin

原文摘要

Physics-informed computational methods usually optimize parameters within a functional representation whose structure is fixed in advance. This work proposes a Physics-Informed Method of Group Data Handling (PI-GMDH), in which representations of coupled physical fields are progressively constructed during solution. Candidate functional directions are evaluated through the first variation of the complete physical and observational objective, introduced in packages, and followed by block-coordinate damped Gauss-Newton coefficient optimization. The framework is demonstrated with tensor-product Chebyshev functions on the incompressible Navier-Stokes equations using a two-dimensional time-dependent Taylor-Green benchmark. Under the tested configuration, adaptive PI-GMDH reached validation and held-out test losses of 5.299e-19 and 5.296e-19 with 204, 201, and 175 active functions for u, v, and p. Complete degree-by-degree and all-terms PI-GMDH variants, together with selected PINN and KAN reference configurations, are used to examine the effect of structural construction policy. The Taylor-Green experiments demonstrate that package-wise structural-growth policies can substantially affect the attained reconstruction accuracy, representation size, and computational cost under the tested configuration. These results support PI-GMDH as a self-organizing basis-construction framework for physics-informed regression and field reconstruction; they do not separately establish the advantage of its candidate-scoring mechanism or general superiority over other physics-informed approaches.

cs.LG · physics.flu-dyn

版本更新 · v2arXiv:2609.35910

The Decision Value of Perception Compute

Hoang Pham Cong, Ho Viet Duc Luong

原文摘要

Adaptive perception spends extra computation on inputs where perception is expected to improve. When perception feeds a downstream decision system, a better perception output need not produce a better decision. We define the decision value of perception compute as the change in downstream loss from escalating an input from a cheap to an expensive perception mode. Because this value can be negative, the allocation of perception compute should be judged against a budget-constrained decision oracle, with uniform full-fidelity inference as a baseline rather than an upper bound. We introduce DEEP (Decision Evaluation for Escalated Perception), a benchmark that scores pre-escalation allocators against this oracle under selection, latency and energy budgets, charging each allocator for its own computation. With deployed monocular geometry on KITTI and nuScenes, we find that 34--54% of the escalations that change downstream loss make it worse; harmful escalations also occur for the published PDM-Closed planner, evaluated open-loop on nuPlan with real detector outcomes. On nuScenes, perception-level gain frequently disagrees in sign with decision value. This mismatch has practical consequences: choosing among fixed deployable signals by missed-object perception gain rather than by decision value reduces realized test decision gain by 7.4% of the all-cheap loss on average. Learned allocators recover part of the oracle's value by finding beneficial escalations but select nearly as much harm as random, and once their own computation is charged at a 20% latency budget, only the lightweight routers, at about 3.5\% of a full detector pass, still beat random.

cs.LG · cs.CV

版本更新 · v2arXiv:2610.28537

Beyond 4 bits: Data-Centric Post-training Quantization and Recovery

Rhea Sanjay Sukthanker, Riccardo Grazzi, James Hensman, Pashmina Cameron

原文摘要

Post-training quantization (PTQ) compresses the weights of a trained model with a limited compute and data budget. Down to 4 bits, state-of-the-art PTQ methods are near-lossless; below this limit, the tested scalar PTQ configurations collapse to near-random accuracy and quality becomes acutely sensitive to choices such as the quantization representation and calibration data. At such low bitwidths, recovery fine-tuning becomes crucial, and its effectiveness depends heavily on these same choices. Yet there is little systematic evidence on how the quantization method, the quality and task alignment of calibration data, and the design of the recovery procedure interact to determine the final 2-bit model. We present a systematic study of 2-bit quantization and recovery fine-tuning across model families, scales, and data distributions, and distill our findings into practical, architecture-dependent best practices for obtaining an accurate 2-bit model. This recipe, which combines a vector codebook representation, a small task-aligned calibration mixture, and a short codebook-only distillation stage, restores 63-68% average accuracy across seven question-answering tasks at 2 bits (a 71-73% BF16 ceiling), versus the near-random 29% of the evaluated scalar PTQ pipelines, at roughly half the compute of a full-model QAT baseline. Given the same calibration and recovery data, EfficientQAT, a trained scalar baseline, avoids the collapse. However, at a matched rate it trails the vector recipe on generation on every model (2.8-10.9 AvgGen; within noise on Qwen3-MoE) and on question answering on three of the four studied models (2.7-7.8 AvgQA). We also find that the best recovery data depends on the calibration data, an interaction confirmed on a held-out model. Our architecture-aware map of the 2-bit design space shows that representation, calibration data, and recovery data must be chosen jointly.

cs.AI

新论文 · v1arXiv:2611.21422

Rubric-Based AES with LLM Ensembles: Scoring of Student Responses in Critical Online Reasoning Using Ensembles of LLMs

Daniel Baumartz, Alexander Mehler, Patrick Schrottenbacher, Leon Hammerla, Sebastian Gombert

原文摘要

Grading student responses is costly and time-consuming. Moreover, human annotators are prone to bias, fatigue, and rater drift, which further limits grading reliability. Various methods have been employed to develop automated essay scoring (AES) systems to reduce these costs and address bias-related issues. These methods range from early text-feature-based systems to current machine-learning and transformer-based models. Specifically, large language models (LLMs) offer the potential to streamline the development of AES systems. However, they are subject to well-known limitations, such as hallucinations and prompt sensitivity. This creates a need to explore the potential of a mixed-methods approach that leverages the synergies among individual LLMs through ensemble formation. We evaluate the potential of ensembles of LLMs to match human-rater agreement when grading student responses in economics and medical education across 103 criteria. We use 631 German-language responses and corresponding human ratings by 2 to 5 annotators from a previously conducted assessment. We experiment with 523 rubric-based prompt configurations across 15 open-weight LLMs and assess their consistency with human raters using Krippendorff's alpha for inter-annotator agreement (IAA) and the Alternative Annotator Test (alt-test). Our results show that replacing one human rater with the best-performing LLM--prompt configuration maintains or improves IAA for 91 of the 103 criteria and that LLMs pass the alt-test for 90 criteria. Our findings suggest that current state-of-the-art LLMs can support AES of German responses in economics and medical education, reducing human workload and enabling faster processing, while validation against human ratings remains necessary.

cs.CL

新论文 · v1arXiv:2611.21421

AuXSpec: Learning to Autoformalize Program Specifications via Executable Postconditions

Haocheng Wang, Weixuan Yuan, Florian Hübler, Yinya Huang, Zhijiang Guo, Sorrachai Yingchareonthawornchai, Rasmus Kyng

原文摘要

Formal specifications define intended program behavior and underpin verification, but writing them in theorem provers such as Lean is slow and expertise-intensive, making specification a major bottleneck to scaling formally correct code. Autoformalization from natural-language problems is promising, yet training specification models requires scalable feedback on whether generated specifications capture intended behavior. Prior approaches obtain this feedback through explicit proofs or separately implemented code with formal correspondence proofs, both expensive. We introduce AuXSpec, a Lean framework that pairs declarative postconditions with executable counterparts, obtained through type class resolution when possible and otherwise defined separately with an equivalence proof. Test cases can then evaluate declarative specifications by execution, letting us use agreement as a reinforcement-learning reward. Using AuXSpec, we synthesize 3,037 LeetCode problems and expand training to 19.5K samples through expert iteration on Codeforces. Experiments show that training Qwen3.5-9B and Qwen3.8-27B with the execution-based reward improves SpecGen pass@4 on AuXSpec-Test 0.7% $\to$ 82.7% and 14.3% $\to$ 84.3%, and the 27B model outperforms all frontier models. On the out-of-domain Verina benchmark with its original prompts, both models improve 7.0% $\to$ 90.5% and 21.9% $\to$ 90.5% and surpass VeriSpecGen-30B-A3B, the prior state of the art among open-source specification models, on pass@1 (theirs 47.6% vs. ours 73.7% and 84.5%) and pass@4 (theirs 63.9% at pass@10 vs. ours 90.5%). We release the weights of the trained 9B and 27B models and the 3,037 problems with test cases, including the 300-problem AuXSpec-Test benchmark with expert review and correctness proofs for future evaluation. Models and data: https://huggingface.co/collections/eth-algopt/auxspec. Website: https://auxspec.org

cs.SE · cs.CL

版本更新 · v2arXiv:2602.14869

Concept Influence: Leveraging Interpretability to Improve Performance and Efficiency in Training Data Attribution

Matthew Kowal, Goncalo Paulo, Louis Jaburi, Tom Tseng, Lev E McKinney, Stefan Heimersheim, Aaron David Tucker, Adam Gleave 等 9 位作者

原文摘要

As large language models are increasingly trained and fine-tuned, practitioners need methods to identify which training data drive specific behaviors, particularly unintended ones. Training Data Attribution (TDA) methods address this by estimating datapoint influence. Existing approaches like influence functions are both computationally expensive and attribute based on single test examples, which can bias results toward syntactic rather than semantic similarity. To address these issues of scalability and influence to abstract behavior, we leverage interpretable structures within the model during the attribution. First, we introduce Concept Influence which attribute model behavior to semantic directions (such as linear probes or sparse autoencoder features) rather than individual test examples. Second, we show that simple probe-based attribution methods are first-order approximations of Concept Influence that achieve comparable performance while being over an order-of-magnitude faster. We empirically validate Concept Influence and approximations across emergent misalignment benchmarks and real post-training datasets, and demonstrate they achieve comparable performance to classical influence functions while being substantially more scalable. More broadly, we show that incorporating interpretable structure within traditional TDA pipelines can enable more scalable, explainable, and better control of model behavior through data.

cs.AI · stat.ML

新论文 · v1arXiv:2611.21419

Quality and Similarity Indicators for Video-Based Heart Rate and Respiratory Rate Fusion: A Controlled Evaluation

Nhi Nguyen, Le Nguyen, Manuel Lage Cañellas, Miguel Bordallo López, Constantino Álvarez Casado

原文摘要

Camera-based heart rate (HR) and respiratory rate (RR) estimates vary in reliability from window to window, and a common remedy is to fuse several extraction methods with weights from reference-free signal quality indicators. However, few studies test whether these indicators identify the most accurate method in each window or outperform simpler fusion. We evaluate 11 quality indicators and 8 inter-method similarity measures for fusing 7 HR and 10 RR extractors on four public datasets (75 test subjects). Filters, indicator directions, weighting parameters and a learned ridge baseline are selected on a separate five-subject set before testing. Fusion is compared with equal averaging, a constant typical-rate predictor, the best single method selected in advance and in hindsight, and a per-window oracle with a permutation-based null, using subject-level bootstrap confidence intervals. For HR, quality-weighted fusion reduced mean absolute error by 10.9% [5.2, 18.4] relative to averaging and by 15.6% [8.1, 23.1] relative to the method selected in advance, and was statistically indistinguishable from the best method selected in hindsight. Weighting methods by their agreement did not improve on averaging. For RR, the indicators could not identify the more accurate estimates: sharpened quality weighting increased error (-2.6% [-4.8, -0.8]), fused estimates were uncorrelated with the reference (r <= 0.22), and averaging was less accurate than the typical-rate predictor (-3.1% [-6.1, -0.4]). With a second respiratory reference, apparent RR gains reflected estimates moving toward its typical rate. Quality-guided fusion is worthwhile for video HR; for RR, 30-s windows improved estimates but not fusion.

cs.CV

版本更新 · v2arXiv:2608.21847

CALE: Color-Adaptive Low-Light Enhancement via Learnable Intensity Coordinates

Yi Ai, Zheng Chen, Yuanhao Cai, Yulun Zhang, Xiaokang Yang

原文摘要

Small brightness differences in dark regions are difficult to recover, and their spacing in the coordinates used for restoration depends on the intensity mapping. A globally learned Box--Cox exponent controls the full response, coupling different brightness intervals and applying the same mapping to every image. We propose Color-Adaptive Low-Light Enhancement (CALE), which combines an image-conditioned intensity representation with structure-conditioned rectified flow (\mbox{SR-RF}). The representation, Box--Cox Intensity Adaptation (BIA), refines a normalized Box--Cox transform using a monotone piecewise-linear curve. Compact statistics of the low-light observation determine the output widths of brightness intervals, allowing the response to vary across images. Positive, normalized slopes preserve intensity order and endpoints and give the intensity mapping an analytic inverse. Combined with polar chromaticity, BIA defines the coordinates for the \mbox{SR-RF} network. The mapping and restoration network are trained jointly so that the reconstruction objective shapes the intensity response. Experiments on paired benchmarks and unseen datasets show competitive restoration performance across diverse low-light scenes, with qualitative comparisons illustrating the recovery of scene colors, contrast, and detail. Coordinate ablations separate the effects of interval flexibility and input conditioning, while Full and Tiny models offer quality--computation trade-offs.

cs.CV

新论文 · v1arXiv:2611.21412

Flexible Multi-Stack Super-Resolution of Clinical Brain MRI Using Hetero-Orientation Variational Encoding: Expert Reader Evaluation in a Nigerian Cohort

Kabir Hamzah Muhammad, Lorena Garcia-Foncillas Macias, Soumya Snigdha Kundu, Emily Guilhem, Jon Cleary, Amal Saleh Nour, Yodit Abraham Yaynishet, Chinasa Kalaiwo 等 13 位作者

原文摘要

High-resolution brain MRI is important for neuroradiological assessment, yet routine clinical protocols often rely on fast, anisotropic 2D acquisitions, particularly in resource-constrained settings. These acquisitions may differ in orientation, alignment and anatomical coverage, resulting in partially overlapping image stacks. Existing learning-based super-resolution methods typically rely on single, under-constrained inputs or fixed multi-stack configurations with idealised acquisition geometry. We introduce MS-HVED, a sequence-agnostic framework for flexible multi-stack brain MRI super-resolution. Available axial, coronal and sagittal stacks are independently encoded into multi-scale Gaussian posteriors and fused using a field-of-view-aware Product-of-Gaussians formulation that limits contributions from unsupported regions. Physics-informed domain randomisation simulates heterogeneous clinical acquisition conditions, including obliqueness, partial coverage and imaging artefacts. MS-HVED achieved competitive or superior reconstruction fidelity across synthetic benchmarks and improved downstream tissue segmentation relative to low-resolution inputs. To assess generalisation beyond the training cohorts, we performed external clinical validation on 30 routine brain MRI examinations from a Nigerian cohort, representing an under-represented population in medical imaging research. These examinations formed the largest component of a blinded reader study involving 44 real cases and seven expert neuroradiologists. In the Nigerian cohort, MS-HVED was preferred over the supervised baseline for overall image quality in 187 of 210 paired assessments. Across all clinical cohorts, MS-HVED achieved higher overall-quality scores with all available inputs (5.93 vs. 4.49, p<.001), with the preference persisting when axial or coronal inputs were omitted (p<.001). Code at github.com/cai4cai/ms-hved

cs.CV

版本更新 · v2arXiv:2605.26288

Beyond Differences: Estimating and Tuning Ratio-Based Treatment Effects

Michael Fuchs, Dominik Kreiss

原文摘要

When treatment effects are naturally expressed as ratios -- as in medicine, pricing, and marketing -- the ratio-based CATE $τ(x) = \mathbb{E}[Y|W=1,X=x] / \mathbb{E}[Y|W=0,X=x]$ is the appropriate estimand. Yet the common estimators divide by an estimated baseline, impose a parametric form, or regress the outcome, where a better fit need not mean a better effect estimate. We introduce the Q-Learner, which writes $τ(x)$ as a ratio of two odds, reducing ratio-CATE estimation for binary outcomes to two propensity classification tasks. In benchmarks on synthetic and randomized real-world data, the Q-Learner proves especially suited to industry practice. Standard hyperparameter tuning on its log-loss suffices to make it rank well, get the level of the effects right and their spread more reliably than the alternatives, whereas, for instance, the common S-Learner needs a costly, effect-specific tuning to rank well, often best, but then misses the spread of the effects. If the propensity is known, a single classifier fitted on only the converters is enough. This not only speeds up development, but is often a necessity in practice, where data are frequently collected only for converters.

stat.ML · cs.LG · stat.ME

版本更新 · v2arXiv:2608.03690

An ECG foundation model for zero-shot transfer from hospital to point-of-care devices

Edoardo Coppola, Stefano Fiorini, Pietro Liò, Mattia Savardi, Alberto Signoroni

原文摘要

Electrocardiogram (ECG) recording is moving from hospitals to point-of-care devices that record one or two leads, where public labelled data is too scarce to build foundation models (FMs) as for 12-lead ECGs. Whether existing 12-lead FMs transfer to these devices without retraining is unknown. We show that seven published FMs reach a zero-shot AUROC of 41-81% on two real point-of-care datasets and, when reduced to one or two leads, lose 12-35 AUROC points on average and gain 0.15-0.23 in mean absolute error across 29 benchmarks covering 1,673 clinical targets. We present RELAY, a 7-million-parameter FM that represents ECGs as graphs of segments from whichever leads are recorded, without filling missing ones or knowing their identities. Pretrained on 9.2 million 12-lead ECGs, RELAY is competitive with 12 leads, ranks first on both point-of-care datasets (AUROC 87.6% and 81.5%) and on 26 and 22 of 29 benchmarks with one and two random leads, and keeps the most informative operating point when 12-lead thresholds are applied to device recordings. Matched ablations show that omitting absent leads, rather than filling them, reduces what RELAY loses in transfer. ECG FMs should therefore be judged across all the acquisition settings, not on 12-lead benchmarks alone.

cs.LG

版本更新 · v2arXiv:2609.28782

When a better baseline makes delta learning worse: separating residual scale from learnability to guide target design

Kareem M. Gameel, Ihor Neporozhnii, Sjoerd Hoogland, Oleksandr Voznyy

原文摘要

In scientific machine learning, $Δ$-learning trains models on residual errors relative to approximate baselines, assuming that smaller residual scales inherently yield better downstream accuracy. Here, we demonstrate that residual scale alone is an insufficient heuristic for learnability. For molecular total energies, replacing a linear baseline with a neural network fitted to the same elemental counts reduced the residual scale, yet more than doubled GemNet's prediction error. Across local structural descriptors, more flexible baselines produced smaller residuals that varied sharply between structurally similar molecules. By contrast, semi-empirical baselines left larger residuals, yet yielded lower test error and improved out-of-domain transferability. Across the targets studied, scale-normalized graph Dirichlet roughness tracked scale-normalized prediction error, suggesting its use as a pre-training diagnostic for target selection. These results motivate baseline complementarity as a target-design principle in scientific machine learning: a useful baseline captures contributions the downstream architecture struggles to predict.

cs.LG · physics.chem-ph · physics.comp-ph

新论文 · v1arXiv:2611.21408

Do Embedding-Map Distances Guide LLM Ideation? Four Tests Against What Came Next

Javier Aguilar Martín

原文摘要

A popular ideation recipe treats an embedding map of existing work as a guide: pick points far from everything and decode them into ideas. I test four ways such distances could guide ideation against what appeared later. (1) Far random directions are not text: no later item reaches a raw cosine above about 0.10 with one (after centring or whitening, 22 of 24 cells stay out of reach), and an LLM decoder shown a target's neighbours, with score-only feedback, writes those neighbours in all six target sets. A neighbour-free search writes coherent text at median cosine 0.57 and 0.47 to interior gaps, beating the nearest existing item on 15 and 14 of 20, but only 0.07 and 0.06 to far points. (2) On k-means cells of a past-only map, the registered hole score mostly tracks cell size and position; three density-based gap scores forecast which cells grow fastest at or below chance (AUC 0.30 to 0.66) and add nothing to momentum in Steam and arXiv, with patents inconclusive; more isolated new Steam games are slightly better reviewed. (3) LLM continuations of arXiv papers stay closer to the source than papers that cite it, under three user prompts and against proposal rewrites (gpt-4o) and April-December 2023 papers (OLMo). Critique lengthens this step beyond content-free revision, replicated on 150 new papers. Revealing a real follow-up's title turns the ideas by 0.032 along a follow-up-specific direction beyond revision; reviewer, oracle and distance feedback add 0.13, 0.08 and 0.34 of that turn (upper bounds 0.46, 0.46, 0.70), and their turn is not separable from revision or their longer step. (4) For inspiration material graded by distance, no inverted U was detected, on an anticipation margin that tracks format more than content. Original tests were registered before running, follow-ups (thirteen requested in two AI-agent reviews) after the main results; all are dated only by local git commits.

cs.LG · cs.AI · cs.CL · cs.DL

新论文 · v1arXiv:2611.21407

Harness Leverage Across Model Size: A Ladder of Five Local Agent Harnesses from 0.75B to 4.7B

Hao Wang, Ting Huang

原文摘要

A harness turns a model into a system that completes tasks: it builds the prompt, exposes tools and decides when work is done. Harness comparisons hold the model fixed. Model comparisons hold the harness fixed. We vary both: five shipped harnesses, 18 artifact-producing tasks, four locally deployable open-weight models from 0.75B to 4.7B and a 35.5B anchor, with one deterministic checker scoring all 360 ladder cells. The completion-controlling harness, Mingbird, has the highest mean at every size. Of four pre-registered paired contrasts, two survive Holm correction, at 0.75B and 4.7B. Thirteen of sixteen paired intervals exclude zero under the deployment reading, with leads from 0.075 over the strongest rival to 0.819 over the weakest. The completion gate is not the carrier: disabling all six executable checks moves the mean by +0.080, +0.117, +0.003 and -0.004, every interval including zero, and only at 1.9B does disabling it create empty runs. Removing one repeat-blocking rule alone costs 0.411 at 0.75B and leaves eight of eighteen tasks with no artifact. The harness effect is a progress effect: it decides whether a run leaves anything on disk. Of 357 usable cells, 127 leave none of their declared artifacts and 85 of those exit normally, so a scorer that reads exit status credits work that was never written. Whether the effect grows or shrinks with size, four rungs cannot settle: a floor arm produces a rise, a ceiling a decline, and a cross-family rung a peak.

cs.AI · cs.SE

新论文 · v1arXiv:2611.21406

Do Apple M-Series Processors Enforce the Compound Memory Model? Sequential and Parallel Composition of Arm and TSO

Vijay Nagarajan

原文摘要

Apple's M-series processors can run each thread under the Arm memory model or under a TSO model, and threads of both kinds can share memory. This raises two questions of composition that Apple has not answered. Across threads, how does the machine behave as a whole? Within a thread, which Arm instructions play the roles of TSO's loads, stores and atomic read-modify-writes, and how are they ordered between themselves and across other Arm instructions? We propose that both answers are compositions of preserved program order. Across threads, each thread keeps the preserved program order of its own model: a parallel composition. Within a TSO thread, every Arm instruction keeps TSO's order and adds Arm's, and every atomic memory operation (AMO) is locked, as x86's read-modify-writes are: a sequential composition. We test both on an Apple M1 and an Apple M2 running Linux, with 936 litmus tests chosen from the two models' preserved program order, each run 6x10^8 times on each machine. Arm-mode threads follow Arm's official model, TSO-mode threads follow x86-TSO and the sequential composition, and programs that mix the two modes follow the compound model: neither machine shows an outcome these models forbid. Both machines also preserve three orders that these models do not require, and a compare-and-swap whose comparison fails behaves as if it writes back the value it read, which the Arm model does not require. All three orders held in runs up to 40 times longer.

cs.AR · cs.PL

新论文 · v1arXiv:2611.21405

Shaping AI's Beliefs Before It Generates: Human Sense of Agency in Creative AI

Zi Wang, Mahima Pushkarna

原文摘要

AI systems for creative media generation expand human capabilities yet risk diminishing human agency. A core problem is anchoring, where the user's thinking becomes anchored to a polished, generated artifact, shifting the user's role from active authoring to reactive editing. We propose pre-generation belief controls that can empower creators to inspect, revise, and steer the AI's interpretation before any media is generated, so that the initial anchor is set by the human. We formalize four design principles and instantiate them in the Proactive Co-Creator (PaCC), a system for early-stage conceptual ideation and prototyping that proactively surfaces its assumptions and uncertainty through clarifications, belief graphs, and media attributes, inviting users to shape the AI's beliefs before generating the final outcome. Using PaCC as a design probe in a small-scale, single-arm exploratory user study, we observed that participants reported directionally positive sense of agency and process control relative to their familiar prompt-based AI tools, noting that surfaced assumptions helped them think through details they had overlooked, whereas their feelings about ownership and authorship diverged

cs.AI · cs.HC