arXiv:2605.15601v1 Announce Type: new Abstract: Quantum programs often produce probability distributions rather than deterministic outputs, making verification inherently statistical and increasingly costly on real hardware. In practice, developers still frequently rely on testing with fixed shot budgets on simulators, which are simple but time-consuming and poorly suited to noisy backends. What is missing is a verification approach that is both statistically explicit and budget-aware. This paper formulates Bayesian sequential verification as a reference-based Bayesian hypothesis testing workflow in which priors are derived from explicit reference sources, such as finite-shot reference runs or ideal/statevector-based computation, and verification decisions are updated batch by batch as measurement evidence accumulates. This approach is evaluated in Qiskit on two complementary workloads: Bell-state and QAOA-MaxCut. Across both case studies, the results show that Bayesian sequential verification can substantially reduce measurement costs compared to fixed-budget baselines when the success probability of the program exceeds the target threshold. The findings position Bayesian sequential verification as a practical verification workflow for quantum programs. The approach provides a foundation for future quantum continuous-integration pipelines that require reliable, budget-aware pass/fail decisions and motivates validation on real quantum hardware.
Science Journals
arXiv:2605.15290v1 Announce Type: new Abstract: Hyperparameter transfer across model architectures dramatically reduces the amount of compute necessary for tuning large language models (LLMs). The maximal update parameterization ({\mu}P) ensures transfer through principled mathematical analysis but can be challenging to derive for new model architectures. Building on the spectral feature-learning view of Yang et al. (2023a), we make two advances. First, we promote spectral norm conditions on the weights from a heuristic to the definition of feature learning, and as a consequence arrive at the Complete-P depth and weight-decay scalings without recourse to lazy-learning. Second, we consider a modified spectral norm that preserves the valid scaling law of network weights when weight matrices are not full rank. This enables (to our knowledge, the first) derivation of {\mu}P scalings for grouped-query attention (GQA). We demonstrate the efficacy of our theoretical derivations by showing learning rate transfer across the GQA repetition hyperparameter as well as experiments regarding transfer over weight decay.
arXiv:2605.15294v1 Announce Type: new Abstract: A DFA separates two disjoint languages $L_1$ and $L_2$ if it accepts every word in $L_1$ and rejects every word in $L_2$. Algorithms for active learning of small separating DFAs have many applications, e.g., for learning network invariants, learning contextual assumptions in compositional verification, learning state machines from large amounts of log data, and learning bug pattern descriptions. We propose a simple active learning algorithm, inspired by $L^{\#}$, that learns a minimal separating DFA for disjoint languages $L_1$ and $L_2$ if one exists. Experiments show that our algorithm significantly outperforms existing active learning algorithms on both randomly generated and industrial benchmarks.
arXiv:2605.15707v1 Announce Type: cross Abstract: Whole-heart multi-compartment CT segmentation is clinically important, but standard CNNs do not explicitly enforce anatomical plausibility. Based on statistics derived from the training data, we evaluate whether lightweight explicit shape priors, implemented as shape-aware losses and spatial label distribution heatmap-guided U-Net variants, improve 3D cardiac segmentation on MM-WHS CT and WHS++. Across all experiments, a standard 3D U-Net surprisingly remained a very strong baseline, with handcrafted priors yielding at best marginal and inconsistent changes and often degrading performance. These results suggest that the baseline already captures substantial implicit anatomical regularities and that future gains will likely require more expressive learned priors rather than simple handcrafted anatomical shape constraints.
arXiv:2605.15673v1 Announce Type: cross Abstract: We present a highly detailed instance segmentation model for delineating individual tree crowns in natural broadleaf forests using aerial imagery acquired by unmanned aerial vehicles (UAVs). Tree crown delineation in broadleaf forests is more challenging than in other forest types due to diversity of crown shapes and the lack of clearly defined treetops. To address this issue, we developed a deep-learning-based crown segmentation model trained on high-quality annotated crown outlines. We manually delineated 18,507 crown polygons from orthomosaic images collected across seven forests in Japan by skilled annotators, and developed a model based on Mask2Former with multiple backbone architectures. The best model achieved high segmentation performance in structurally complex broadleaf forests using only RGB imagery. This performance was maintained when applied to geographically distinct forests within Japan, as well as to biologically distinct tropical rainforests in Borneo. These results demonstrate that using a large number of high-quality annotated datasets is critical for achieving detailed and generalizable crown segmentation across diverse forest ecosystems. The developed model has been integrated into DF Scanner Pro, a software that supports practical forest monitoring using UAVs, and this implementation is expected to enable a wide range of users to analyze tree-level information in broadleaf forest from UAVs.
arXiv:2605.15584v1 Announce Type: new Abstract: Vision-language models like CLIP have demonstrated remarkable zero-shot transfer capabilities. However, their susceptibility to imperceptible adversarial perturbations remains a critical security concern. While test-time defenses offer a pragmatic solution for deployed models, existing approaches typically rely on gradient-based optimization during inference, incurring significant computational overhead. In this paper, we revisit the role of data augmentation in CLIP robustness and observe that augmentations are not equally effective: specific augmentations consistently provide robust geometric cues that align with correct class semantics in the hyperspherical feature space. Based on this, we propose Adaptive Geodesic Correction (AGC), a training-free defense mechanism that requires no parameter updates. AGC identifies a reliable augmentation as a geometric anchor and corrects the input feature towards it, utilizing an adaptive step size to balance robustness against clean accuracy preservation. AGC achieves superior performance across eight fine-grained datasets and three CLIP backbones, improving average robust accuracy by 44.4\% over state-of-the-art baseline while delivering a 10$\times$ reduction in inference latency. Our findings reveal a fundamental geometric property of CLIP features, offering a highly efficient and effective paradigm for robust multimodal deployment.
arXiv:2605.15295v1 Announce Type: new Abstract: Machine learning (ML) algorithms are increasingly deployed in high-stakes decision-making domains such as loan approvals, hiring, and recidivism predictions. While existing fairness metrics (e.g., statistical parity, equal opportunity) effectively quantify outcome-oriented disparities, they offer limited insight into the procedure or explanation behind biased decisions. To address this gap, we propose Group-level Explanation Stability Disparity (GESD), a \textit{procedural-oriented} fairness metric that measures disparities in the stability, robustness, and sensitivity of model explanations across different subgroups in a protected category. %GESD is explainer-agnostic, model-agnostic, and extends the scope of fairness analyses to the level of explainability. We further integrate GESD into a multi-objective optimization framework that jointly optimizes for utility, outcome-based fairness, and explanation-based fairness called FEU (Fairness--Explainability--Utility). Empirical results on multiple benchmark datasets show that GESD effectively captures group-wise discrepancies in explanation quality, and that FEU improves both utility and fairness over state-of-the-art methods. By bridging outcome-based and explanation-based fairness, GESD offers a comprehensive tool for diagnosing and mitigating bias in predictive modeling. Our code and datasets are available on GitHub {\hyperlink{https://github.com/horlahsunbo/GESD}{https://github.com/horlahsunbo/GESD}}
arXiv:2605.15298v1 Announce Type: new Abstract: Vision-language-action models have advanced rapidly, but robot trajectories alone provide limited coverage for learning broad physical understanding. PhysBrain 1.0 studies a complementary route: converting large-scale human egocentric video into structured physical commonsense supervision before robot adaptation. Our data engine extracts scene elements, spatial dynamics, action execution, and depth-aware relations, then turns them into question-answer supervision for training PhysBrain VLMs. The resulting physical priors are further transferred to VLA policies through a capability-preserving and language-sensitive adaptation design. Across multimodal QA benchmarks and embodied control benchmarks, including ERQA, PhysBench, SimplerEnv-WidowX, LIBERO, and RoboCasa, PhysBrain 1.0 achieves SOTA results and shows especially strong out-of-domain performance on SimplerEnv. These results suggest that scaling physical commonsense from human interaction video can provide an effective bridge from multimodal understanding to robot action.
arXiv:2605.15202v1 Announce Type: new Abstract: Presentations are a primary medium for scholarly communication, yet most AI slide generators optimize the artifact (a visually plausible deck) while under-optimizing the delivery process (pacing, narrative, and presentation preparation). We present DeepSlide, a human-in-the-loop multi-agent system that supports preparing the full presentation process, from requirement elicitation and time-budgeted narrative planning, to evidence-grounded slide--script generation, attention augmentation, and rehearsal support. DeepSlide integrates (i) a controllable logical-chain planner with per-node time budgets, (ii) a lightweight content-tree retriever for grounding, (iii) Markov-style sequential rendering with style inheritance, and (iv) sandboxed execution with minimal repair to ensure renderability. We further introduce a dual-scoreboard benchmark that cleanly separates static artifact quality from dynamic delivery excellence. Across 20 domains and diverse audience profiles, DeepSlide matches strong baselines on artifact quality while consistently achieving larger gains on delivery metrics, improving narrative flow, pacing precision, and slide--script synergy with clearer attention guidance.
arXiv:2605.15538v1 Announce Type: cross Abstract: We study a stochastic optimization problem in which the sampling distribution depends on the decision variable, and the available samples are generated through an iterate-dependent Markov chain. Such settings arise naturally in problems with decision-dependent uncertainty; however, they introduce bias and temporal dependence, which render standard techniques developed for i.i.d.\ noise inapplicable. In this work, we analyze the stochastic mirror descent algorithm under iterate-dependent Markov noise. We first establish almost sure convergence for both convex and non-convex problems under the mild assumption of Lipschitz continuity of the objective function, without requiring differentiability. We then derive finite-time concentration bounds for smooth objectives. In the convex setting, the resulting sample complexity matches the classical rate of stochastic mirror descent under i.i.d.\ noise. In the non-convex setting, we obtain a sample complexity bound in terms of the norm of the Riemannian gradient over the probability simplex. Overall, our results establish a unified convergence framework for stochastic mirror descent with state-dependent Markov noise, and highlight its behavior in both convex and non-convex regimes.
arXiv:2602.00841v4 Announce Type: replace Abstract: Visual Place Recognition (VPR) demands representations robust to drastic environmental and viewpoint shifts. Existing aggregation paradigms either depend on extensive supervised training or rely on first-order pooling, often struggling to preserve structural correlations under extreme shifts or incurring high adaptation costs. In this work, we propose Riemannian Invariant Aggregation (RIA), a unified geometric framework that explicitly models second-order scene structure on the Symmetric Positive Definite (SPD) manifold. By treating perturbations as tractable congruence transformations, RIA leverages geometry-aware Riemannian mappings to project covariance descriptors into a linearized Euclidean space, effectively preserving invariant structural components while suppressing noise. Extensive evaluations demonstrate that RIA achieves zero-shot performance comparable to supervised methods, and establishes state-of-the-art accuracy with simple fine-tuning, particularly in unstructured environments. The source code will be released.
arXiv:2605.15299v1 Announce Type: new Abstract: In search and recommendation systems, predictive models often suffer from temporal instability when certain input features introduce volatility in output scores. This instability can degrade model reliability and user experience especially in multi-stage systems where consistent predictions are critical for downstream decision making. We introduce Fortress, a general framework for enhancing model stability and accuracy by identifying and pruning features that contribute to inconsistent prediction scores over time. Fortress leverages historical snapshots temporally partitioned datasets capturing score fluctuations for the same entity across periods and follows a four-step process: (1) collect historical snapshots, (2) identify samples with unstable predictions, (3) isolate and remove instability-inducing features, and (4) retrain models using only stable features. While semantic features from LLMs and BERT-based models improve generalization, they often lack full query or entity coverage. Engagement-based features offer strong predictive power but tend to introduce temporal instability. Fortress mitigates this trade-off by suppressing the volatility of engagement signals while retaining their predictive value leading to more stable and accurate models. We validate Fortress on a query-to-app relevance model in a large-scale app marketplace. Offline experiments demonstrate notable improvements in prediction stability (measured by Coefficient of Variation) and classification performance (measured by PR-AUC).
arXiv:2605.15300v1 Announce Type: new Abstract: Most Vision Language Models (VLMs) directly map outputs from ViT encoders to the LLM via a lightweight projector. While effective, recent analysis suggests this architecture suffers from an alignment challenge: visual features remain distant from the text space in the initial layers of the LLM, forcing the model to waste critical depth~\cite{zhang-etal-2024-investigating,artzy-schwartz-2024-attend} on superficial modality alignment rather than deep understanding and complex reasoning. In this work, we propose Deep Pre-Alignment (DPA), a novel architecture that replaces the standard ViT encoder with a small VLM as perceiver, ensuring visual features are deeply aligned with the text space of the target large language model. Comprehensive experiments demonstrate the effectiveness of DPA. On the 4B parameter scale, DPA outperforms baselines by 1.9 points across 8 multimodal benchmarks, with gains widening to 3.0 points at the 32B scale. Moreover, by offloading alignment to the perceiver, DPA achieves a 32.9\% reduction in language capability forgetting over 3 text benchmarks. We further demonstrate that these gains are consistent across different LLM families including Qwen3 and LLaMA 3.2, highlighting the generality of our approach. Beyond performance, DPA also offers a seamless upgrade path for current VLM development, requiring only a modular replacement for the visual encoder with marginal computation overhead.
arXiv:2605.15301v1 Announce Type: new Abstract: Large language models (LLMs) still struggle with the rigorous reasoning demands of hard competitive programming. While recent multi-agent frameworks attempt to bridge this reliability gap, they remain fundamentally stateless: they rely on static retrieval and discard the valuable problem-solving and debugging experience gained from previous tasks. To address this, we present Solvita, an agentic evolution framework that enables continuous learning without requiring weight updates to the underlying LLM. Solvita reorganizes problem-solving into a closed-loop system of strategy selection, program synthesis, certified supervision, and targeted hacking, executed by four specialized agents: Planner, Solver, Oracle, and Hacker. Crucially, each agent is paired with a trainable, graph-structured knowledge network. As the system operates, outcome signals, such as pass/fail verdicts, test certification quality, and adversarial vulnerabilities discovered by the Hacker, are recast as reinforcement learning updates to these network weights. This allows the agents to dynamically route future queries based on past successes and failures, effectively accumulating transferable reasoning experience over time. Evaluated across CodeContests, APPS, AetherCode, and live Codeforces rounds, Solvita establishes a new state-of-the-art among code-generation agents, outperforming existing multi-agent pipelines and nearly doubling the accuracy of single-pass baselines.
arXiv:2605.15304v1 Announce Type: new Abstract: The relations connecting propositions in discourse such as cause (A because B) or concession (A although B) are a subject of intense interest in Computational Linguistics and Pragmatics, but challenging to study and compare across languages. Recent progress in standardizing discourse relation inventories across datasets offers the potential to facilitate such studies, but is hindered by the complexity of relevant data and the lack of easily accessible interfaces to analyze it. In this paper we present DiscoExplorer, a new open source web interface, capable of running on local computers, which we use to make datasets from the DISRPT Shared Task on discourse relation classification publicly available, covering 16 different languages. We present the query language, search and visualization facilities for relations and signaling devices such as connectives, as well as some example studies.
arXiv:2605.15307v1 Announce Type: new Abstract: Motion-centric video editing remains difficult for large generative video models, which often respond well to appearance changes but struggle to produce specific, localized actions or state transitions in an existing clip. We introduce Sound Sparks Motion, a training-free framework that enables motion editing in an audio-visual video generation model by tuning its internal multimodal conditioning signals at test time. Rather than modifying model weights, our method tunes only two lightweight variables: an audio latent derived from the source video and a residual perturbation in the text-conditioning. We find that this combination can encourage motion edits that the underlying model often struggles to realize under prompt-only control. Since there is no direct way to evaluate temporal alignment between text and motion, we guide the tuning process using a vision-language model that provides feedback indicating whether the intended motion appears in the generated video. This simple supervision yields an effective semantic objective for motion editing, while regularization and perceptual-temporal constraints help preserve content and visual quality. Beyond per-video tuning, we show that the learned latent controls are transferable across videos, suggesting that they capture reusable motion-edit directions rather than overfitting to a single example. Our results highlight multimodal conditioning tuning, particularly through the audio pathway, as a promising direction for motion-aware video editing, and suggest that test-time tuning can serve as a lightweight probing mechanism that helps reveal latent motion controls embedded in the model's multimodal conditioning. Code and data are available via our project page: https://amirhossein-razlighi.github.io/Sound_Sparks_Motion/
arXiv:2605.15439v1 Announce Type: cross Abstract: We reformulate the R\'enyi entanglement of purification as a constrained minimum output R\'enyi entropy problem. Equivalently, for $p>1$, this formulation can be expressed in terms of a constrained maximal output Schatten $p$-norm. More precisely, for a completely positive map $\Omega:L(B')\to L(A)$, we consider the quantity $\upsilon_p(\Omega)$ defined by optimizing $\|(\Omega\otimes \mathrm{id}_E)(\sigma^{B'E})\|_p$ over all bipartite states $\sigma^{B'E}$ whose $B'$-marginal is maximally mixed. We focus on the case $p=2$. First, we compute $\upsilon_2$ for the transpose-depolarizing channel and prove that it is multiplicative under tensor powers. We then establish a general multiplicativity criterion: whenever a completely positive map $N:L(B')\to L(A)$ satisfies $N^{\dagger} \mathbin{\circ} N=a\,\mathrm{id}_A+b\,\mathrm{Tr}[\cdot]\,I_d$ for some constants $a,b\ge 0$, where $N^{\dagger}$ denotes the Hilbert-Schmidt adjoint of $N$, the quantity $\upsilon_2(N)$ is multiplicative under tensor powers. Examples of channels satisfying this criterion include the transpose-depolarizing channel, the depolarizing channel, and their respective complementary channels. Furthermore, we show that, for every completely positive map $\Omega$, multiplicativity of $\upsilon_p(\Omega)$ implies multiplicativity for its complementary map. This yields the corresponding additivity statements for the associated R\'enyi-2 entanglement of purification.
arXiv:2605.15581v1 Announce Type: new Abstract: LLM-based root cause analysis (RCA) agents have recently emerged as a promising paradigm for incident diagnosis in microservice AIOps. However, their reliability remains fragile: an error in early evidence collection, hypothesis formulation, or causal analysis can propagate through the reasoning trace and eventually corrupt the final diagnosis. In this paper, we present \textbf{STAR}, a \emph{Stage-attributed Triage and Repair} framework for repairing erroneous RCA traces. STAR explicitly decomposes an RCA workflow into four structured stages, namely \emph{Evidence Package} (EP), \emph{Hypothesis Set} (HS), \emph{Analysis Structure} (AS), and \emph{Decision Report} (DR), and treats agent failure as a stage-localizable reasoning bug rather than a monolithic end-to-end error. Built on top of LangGraph, STAR performs stage-wise auditing, budget-aware \emph{Fast/Slow Routing}, \emph{decisive stage localization via counterfactual candidate evaluation}, and stage-specific patch-and-replay repair. We evaluate STAR on a public large-scale benchmark and a real-world production dataset, using two RCA agent workflows and three foundation models. Experimental results show that STAR consistently improves both root cause localization and fault type classification over strong baselines. Moreover, STAR identifies the decisive faulty stage with high accuracy, repairs most initially incorrect traces within one or two replay rounds, and benefits substantially from both Fast/Slow Routing and counterfactual stage evaluation. These results suggest that explicitly modeling \emph{where} an RCA agent fails is an effective path toward reliable, debuggable, and self-repairing agentic RCA systems.
arXiv:2605.14665v2 Announce Type: replace Abstract: Legal reasoning is not semantic similarity search. A court judgment encodes constrained symbolic reasoning: precedent propagation, procedural state transitions, and statute-bound inference. These are properties that vector-based retrieval-augmented generation (RAG) cannot faithfully represent. Hallucinated precedents, outdated statute citations, and unsupported reasoning chains remain persistent failure modes in LLM-based legal AI, with real consequences for access to justice in high-caseload jurisdictions such as India. This paper presents Falkor-IRAC, a graph-constrained generation framework for Indian legal AI that grounds generation in structured reasoning over an IRAC (Issue, Rule, Analysis, Conclusion) knowledge graph. Judgments from the Supreme Court and High Courts of India are ingested as IRAC node structures enriched with procedural state transitions, precedent relationships, and statutory references, stored in FalkorDB for low-latency agentic traversal. At inference time, LLM-generated answers are accepted only if a valid supporting path can be traced through the graph, a check performed by a falsifiability oracle called the Verifier Agent. The system also detects doctrinal conflicts as a first-class output rather than silently resolving them. Falkor-IRAC is evaluated using graph-native metrics: citation grounding accuracy, path validity rate, hallucinated precedent rate, and conflict detection rate. These metrics are argued to be more appropriate for legal reasoning evaluation than BLEU and ROUGE. On a proof-of-concept corpus of 51 Supreme Court judgments, the Verifier Agent correctly validated citations on completed queries and correctly rejected fabricated citations. Evaluation against vector-only RAG baselines is left for future work. The companion InIRAC dataset, 500+ structured Indian court judgments with IRAC annotations, is released alongside this paper.
arXiv:2605.15328v1 Announce Type: new Abstract: Fully Connected Neural Networks (FCNNs) are often regarded as simple and intuitive architectures, yet they serve as the foundation for more complex models. Nonetheless, the lack of consensus on their interpretability continues to pose challenges, underscoring the enduring relevance of simpler, attribution-based approaches for understanding even the most advanced neural architectures. In this regard, we explore a novel idea for estimating feature attribution, by applying perturbation to the features' attached weights instead of their values. This method offers a fresh perspective aimed at mitigating common limitations in Occlusion techniques, such as Added Bias and Out-of-Distribution data. The application of this rule leads to the formation of a pair of novel attribution methods we call XWP and XWP_c. Founded on simple rules, our methods achieve competitive performance in identifying image signals for simple DNNs, competing with the most established attribution methods on standard baseline metrics. Our work thus contributes to the field of Explainability by introducing a robust framework that paves the way for addressing these long-standing vulnerabilities, and leads to more reliable and interpretable model explanations.
arXiv:2605.13143v2 Announce Type: replace Abstract: Knowledge distillation is widely used to improve generalization in practice, yet its theoretical understanding remains elusive. In the standard distillation setting, a teacher model provides soft predictions to guide the training of a student model. We model teacher and student training as coupled stochastic processes and introduce a distillation divergence, defined as the Kullback-Leibler divergence between these two stochastic kernels. Within this framework, we derive two generalization bounds for the student model relative to the teacher's generalization gap: an upper bound under a sub-Gaussian assumption via algorithmic stability, and a lower bound under a central condition with sharper dependence on the distillation divergence. We further develop a loss-sharpness-aware bound with an explicit tightness regime, showing that the teacher's local flatness can strictly tighten the bound. Additionally, in a linear Gaussian case study, the distillation divergence admits an interpretable decomposition into bias, variance, and rank-bottleneck costs, yielding practical guidance for distillation design.
arXiv:2605.15331v1 Announce Type: new Abstract: We study a repeated information design setting in which the receiver, who is also the decision-maker, updates beliefs in a systematically biased way. More specifically, a distorted posterior in our model can be written as a convex combination of the prior and the Bayesian posterior, governed by a fixed but unknown parameter. Over repeated interactions, the sender chooses persuasive signaling schemes, observes only the receiver's realized actions, and seeks to minimize regret relative to a full-information oracle that knows the receiver's biased updating rule. We propose a safe exploration algorithm for learning the receiver's bias while maintaining high persuasion value. The algorithm exploits the asymmetric cost of probing: conservative probes incur only local loss, whereas overly aggressive probes may lose the persuasive opportunity entirely. For general finite state and action spaces and arbitrary bounded utilities, our method achieves $O(\log\log T)$ regret. A matching $\Omega(\log\log T)$ lower bound shows that this rate is optimal. We further discuss the influence on receiver welfare, as well as extensions to jointly unknown prior and bias, and contextual settings with time-varying priors and utilities.
arXiv:2605.15333v1 Announce Type: new Abstract: Large language models have recently reached near-parity with classical planners on well-known planning domains, yet this competence relies on world-knowledge exploitation rather than genuine symbolic reasoning. Goal recognition is a complementary abductive task structurally better suited to LLM strengths: it consists of evaluating consistency with world knowledge rather than generating novel action sequences. This paper provides the first systematic zero-shot evaluation of frontier LLMs as goal recognisers on key classical PDDL benchmarks. Our results show that LLM competence on goal recognition is uneven: some models scale with evidence and approach landmark-based accuracy at full observations, while others remain anchored to world-knowledge priors regardless of how much evidence accumulates. Qualitative analysis of model reasoning traces reveals that this divergence reflects a fundamental difference in evidence integration rather than domain familiarity. These findings position goal recognition as a principled benchmark for the foundational planning knowledge of LLMs.
arXiv:2605.12667v2 Announce Type: replace Abstract: The alignment of Large Language Models (LLMs) utilizes Reinforcement Learning from AI Feedback (RLAIF) for non-verifiable domains such as long-form question answering and open-ended instruction following. These domains often rely on LLM based auto-raters to provide granular, multi-tier discrete rewards (e.g., 1-10 rubrics) that are inherently stochastic due to prompt sensitivity and sampling randomness. We empirically verify the stochasticity of auto-raters that can propagate and corrupt standard advantage estimators like GRPO and MaxRL, as a noisy reward samples can skew normalization statistics and degrade the global learning signal. Empirically, sampling more rewards and taking majority voting may reduce the noise and improve performance, but this approach is computationally expensive. To address this bottleneck, we introduce $\textbf{O}$rdinal $\textbf{D}$ecomposition for $\textbf{R}$obust $\textbf{P}$olicy $\textbf{O}$ptimization ($\textbf{ODRPO}$), a framework that structurally isolates evaluation noise by decomposing discrete rewards into a sequence of ordinal binary indicators. By independently computing and accumulating advantages across these progressively challenging success thresholds, ODRPO prevents outlier evaluations from corrupting the global update while establishing an implicit, variance-aware learning curriculum. Empirically, ODRPO achieves robust performance on Qwen2.5-7B and Qwen3-4B models, outperforming baselines with relative improvements of upto 14.8% on FACTS-grounding-v2 and 7.5% on Alpaca-Evals. Critically, these gains are achieved with negligible training-time overhead, as ODRPO requires no additional compute per step compared to standard estimators. Supported by theoretical analysis confirming its optimization stability, ODRPO provides a scalable and robust framework for aligning models within the noisy, discrete evaluation landscape of modern RLAIF.
arXiv:2605.15334v1 Announce Type: new Abstract: The automatic synthesis of a program from any form of specification is regarded as a holy grail of computer science. Fueled by LLMs, NL2Code has achieved tremendous success, yet the fundamentally more challenging task of synthesizing programs from input-output behavior, which we refer to as IO2Code, remains largely unsolved. Whereas NL2Code can exploit the semantic alignment between natural language and code acquired during pretraining, IO2Code requires recovering underlying principles from concrete computational behavior, navigating a vast and underspecified hypothesis space. To address this, we propose DIO-Agent, a discovery agent for IO2Code. Our method frames IO2Code as an evolutionary search over discrete program space, in which an LLM serves as the mutation operator and concrete error signals from execution guide each mutation. To prevent the search from wandering into structurally complex yet incorrect dead ends, we introduce the Transformation Priority Premise as a mutation prior that biases the LLM toward the simplest hypothesis consistent with current evidence, progressively escalating from constants to conditionals to iteration only when simpler constructs are insufficient. To facilitate systematic study, we further construct an IO2CodeBench spanning multiple difficulty levels. Extensive experiments show that DIO-Agent consistently outperforms both traditional program-by-example method and SOTA evolution-agent baselines across all difficulty levels and various LLMs, while substantially surpassing test-time scaling strategies with equivalent sampling budgets.