Optimization
-
SemEllipsoid: Semantic-aware Shape Abstraction with Ellipsoid
2026-10-30 17:59:14Kuankuan Cheng, Haocheng Yuan, Changjian LiAbstract
Representing 3D shapes using a set of primitives - known as shape abstraction - is an actively studied problem in computer graphics with wide applications in shape understanding, analysis, and simulation due to the compactness and computational efficiency of the primitive-based representation. Prior work has primarily focused on geometry-aware shape abstraction, where the objective is to approximate the input shape with a minimal set of primitives while preserving geometric fidelity. However, these methods often overlook semantic consistency, inevitably leading to primitives that span across multiple semantic parts. Such inconsistency can cause undesirable artifacts in downstream tasks such as part-level shape editing or manipulation. In this paper, we introduce a new task - semantic-aware shape abstraction, where the goal is to produce abstractions that are both geometrically faithful and semantically consistent, ensuring that each primitive adheres strictly to a single semantic part without crossing semantic boundaries. We propose SemEllipsoid, a novel method to address this problem using ellipsoidal primitives. The core idea is to inject semantic constraints into the primitive optimization process leveraging the availability of semantic cues from large-scale foundation models. We design a fully differentiable projection-blending-evaluation scheme to measure semantic divergence against input semantic masks, and formulate an alternating optimization strategy to iteratively solve for optimal semantic assignment and primitive configurations. Our method is non-learning-based, requiring no data collection or network training, and generalizes well to shapes from arbitrary categories. It also supports different levels of abstraction according to user demand. Extensive experiments and evaluations validate the effectiveness of our approach across several 3D shape collections.
Abstract (translated)
URL
https://arxiv.org/abs/2610.42145
PDF
-
RoboArchitect: Automating the Design of Robotic Agents with Persistent Memory
2026-10-30 17:58:48Shengguang Wu, Suning Huang, Cherie Ho, Ria Doshi, Hao Zhu, Serena Yeung-LevyAbstract
Recent work shows that a frozen robot policy performs far better inside an agent harness, where control code decides when to query the policy, consult vision tools, or command the arm directly. Yet today these harnesses are built by hand and redesigned for each new robot or benchmark. We show that harness design can be automated, offering a strong alternative to finetuning on a novel setup. In RoboArchitect, a coding model observes a robot's rollouts, diagnoses failures, and iteratively revises the harness around the frozen policy. The harness evolves on a set of episodes for its optimization and is evaluated on held-out scenes and tasks. Without updating any model weights, RoboArchitect raises average success from 9.7% to 43.3% on unseen scenes (4.5x) and from 12.7% to 40.2% on unseen tasks (3.2x), across multiple benchmarks and policies. The gains are largest where the benchmark is far from the policy's training data: wrapping a general-purpose policy, the evolved harness yields a specialist that often outperforms even the finetuned version. Intuitive control strategies are automatically designed, such as lifting a jammed arm clear or reaching in horizontally when an obstacle blocks the way from above. The harness also evolves its own persistent memory, carrying learned experience, such as effective grasp depths, into later decisions. These results suggest that substantial headroom in robotic manipulation lies in the agent harness, and its design can be effectively automated and optimized without weight updates.
Abstract (translated)
URL
https://arxiv.org/abs/2610.42141
PDF
-
IGDA-ICMH: an Importance Guided Feature and Network Dual Adaptation Framework towards Image Compression for Machine and Human Vision
2026-10-30 15:54:03Yanbo Gao, Xinyu Zhao, Shuai Li, Xun Cai, Mao Ye, Li LiAbstract
With the surge of machine vision applications, image compression for both machine and human vision (ICMH) has become an important research area. Existing studies mainly focus on task-specific adaptation or multi-task joint optimization, yet they often overlook the unequal contributions of different features and network layers to downstream machine vision tasks. To address this issue, we first quantitatively investigate the impact of different feature and network layers on machine vision performance, providing a principled understanding of their importance. Based on this analysis, an Importance Guided feature and network Dual Adaptation (IGDA) framework is developed for ICMH. It adopts implicit neural representation as the basic architecture and designs an efficient adaptation structure with both feature and network weight adaptation. An Importance-Guided Network Adaptation (IGNA) module is developed with a low-rank adaptation strategy to selectively adjust key network layers according to their importance in different tasks. An Importance-Guided Feature Adaptation (IGFA) module is developed with a low-rank and Kronecker product-based transform to explore the non-local property of the features in order to reduce the number of adaptation parameters. Extensive experiments on representative machine vision tasks, including image classification and object detection, demonstrate that our IGDA-ICMH framework achieves significantly better performance than the existing implicit neural representation-based methods for ICMH.
Abstract (translated)
URL
https://arxiv.org/abs/2610.41924
PDF
-
Denoising Multirobot Trajectories
2026-10-30 15:13:03Yuhao Zhang, Keisuke Okumura, Ajay Shankar, Amanda ProrokAbstract
Multi-robot trajectory planning is a fundamental problem in multi-robot coordination but remains computationally challenging due to its nonconvex, multimodal, and high-dimensional nature. This work builds upon D4orm, a dynamics-aware diffusion-denoising framework, and develops a family of planning architectures for diverse operational requirements. Unlike conventional numerical optimization methods, D4orm employs sampling-based optimization to generate solution trajectories through massively parallel sampling, leveraging modern computing architectures such as GPUs. Its diffusion-denoising structure iteratively optimizes deformations to candidate control trajectories, providing an efficient and versatile paradigm for generating kinodynamically feasible and conflict-free trajectories. Using D4orm as the building block for advanced planners, we present a decoupled planner for improved scalability, an online receding-horizon planner with feedback control, and a distributed planner for resource-constrained settings. Evaluations with differential-drive and holonomic robots in 2D and 3D environments demonstrate that D4orm-based approaches find high-quality solutions faster and more reliably than other sampling-based optimization methods, such as MPPI, as well as a learned diffusion-model-based method. We further demonstrate zero-shot deployment on ten real quadrotors with obstacles, large-scale deconfliction with 100 simulated robots, and fully onboard distributed `lifelong' operation with six ground robots. Overall, these results establish diffusion denoising as a scalable and reliable framework for multi-robot coordination. Code and video: this https URL
Abstract (translated)
URL
https://arxiv.org/abs/2609.35651
PDF
-
Hard Sample-Conditioned Decorrelated Attention for Robust Face Recognition
2026-10-30 14:21:43Tahar Chettaoui, Guray Ozgur, Eduarda Caldeira, Naser Damer, Fadi BoutrosAbstract
Multi-head attention in Vision Transformers (ViTs) allows independent attention patterns across heads, but without explicit regularization, heads may learn redundant patterns rather than complementary roles. Correlation-based penalties encourage head disentanglement, yet apply uniform regularization across all training samples. In face recognition (FR), hard samples whose embeddings lie close to competing identities are critical for optimization and generalization. Motivated by this, we propose ViT-HarDis, which conditions attention head disentanglement on sample difficulty: samples with high cosine similarity to negative class centers receive stronger head decorrelation, enforcing specialization where identity ambiguity is highest. Experiments show improvements over both the baseline and uniform disentanglement on most benchmarks. Our analysis shows that ViT-HarDis reduces the similarity of hard samples to competing class centers while decreasing inter-head attention similarity, and that applying disentanglement only to the final transformer layers outperforms regularizing the entire network. Trained on WebFace4M, ViT-HarDis achieves SOTA results among ViT-based methods on IJB-B, IJB-C, and TinyFace Rank-1, while achieving the best IJB-B and TinyFace results when trained on MS1MV3, improving FR without architectural modifications or additional inference cost. Code: this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.41789
PDF
-
Visual-Inertial RGB-D SLAM with Encoders for a Differential Wheeled Robot
2026-10-30 14:09:58Zhanghao Zhu, Yutaka Kaizu, Kenichi Furuhashi, Kenji ImouAbstract
Recent years have seen multiple impressive results in visual-inertial odometry (VIO) techniques, by which accurate state estimation can be achieved via the extended Kalman filter (EKF) or nonlinear optimization. However, these approaches are rarely open source, and they tend to fail in real experiments due to the temporary lack of feature points and fast motion. Therefore, in this study, we used encoders to overcome the temporary failure of purely vision-based simultaneous localization and mapping (SLAM). Here, we propose a generative measurement model for encoders and derive an expression for the maximum a posteriori (MAP) estimate and necessary Jacobians for optimization. We use our theory to present a novel tightly coupled visual-inertial encoder RGB-Depth (RGB-D) SLAM system. Tests on our system on an in-house dataset confirmed that our modeling effort led to accurate (with a root mean squared error (RMSE) of approximately 2-7 cm) and robust state estimation in real time. The source code and our dataset containing the encoder information have been published for verification.
Abstract (translated)
URL
https://arxiv.org/abs/2610.41768
PDF
-
Task Vectors Can Be Linearly Decomposed into Interpretable and Controllable Components
2026-10-30 13:48:48Yukiya Horiba, Hakaze Cho, Takahiro YakohAbstract
Task Vectors (TVs) provide a compact way to steer large language models, but they are typically treated as atomized directions merely extracted from hidden states, leaving their internal functional structure poorly understood. Motivated by prior work that decomposes in-context learning into Task Recognition (TR) and Task Learning (TL), we ask whether a single TV can likewise be decomposed into functionally meaningful components. We formulate a TR/TL-based decomposition from candidate-label unembeddings, separating each TV into TR, TL, and residual components, and try to optimize the TR and TL coefficients using Bayesian optimization for more effective intervention. We evaluate this formulation across multiple model families, classification tasks, and both instruction- and demonstration-based TVs. Across these settings, TR/TL-based optimization identifies intervention directions that outperform the original TVs. Although optimization over random directions can also improve performance, TR/TL-based search yields more reliable gains, particularly for instruction-based TVs. Moreover, isolated interventions on the decomposed components produce distinct logit-level effects that are consistent with their intended functional roles. Together, these results suggest that TVs are not monolithic steering directions, but can be decomposed into functionally meaningful components that support both interpretation and control.
Abstract (translated)
URL
https://arxiv.org/abs/2610.41743
PDF
-
TSBPO: Tree-Structured Branching for Policy Optimization in Interactive Language Agents
2026-10-30 13:04:05Rui Yu, Chong Mu, Jinchuan ZhangAbstract
Training interactive language agents from terminal outcomes poses three coupled problems: which task and state continuations to sample, how to assign delayed feedback to decisions, and how to aggregate token losses for multi-token actions. We present TSBPO, a tree-structured sampling-to-update paradigm that addresses these problems within one training loop. A likelihood-fitted Root/Branch allocator uses retained task and state evidence to target continuations likely to produce contrasting terminal outcomes under fixed rollout budgets. On the resulting replayed physical trees, trajectory advantages compare immediate sibling segments, while observation-group step advantages compare outcomes at matched task/observation states. The actor then sums clipped token surrogates over each complete response and normalizes across active turns, avoiding response-length averaging in the policy-gradient term. With a Qwen2.5-1.5B actor, fixed checkpoints obtain 97.79% seen and 96.64% unseen stochastic success on ALFWorld and an 81.07% per-episode success rate over six stochastic episodes for each of the 500 WebShop test goals. A WebShop step-term ablation yields 75.93% at the same update-100 evaluation; an ALFWorld aggregation comparison shows higher observed command validity with complete-turn weighting. These results support the evaluated training configuration on two interactive benchmarks.
Abstract (translated)
URL
https://arxiv.org/abs/2610.41697
PDF
-
Interaction-Aware Sampling-Based Motion Planning for Autonomous Surface Vessels in Urban Waterways
2026-10-30 12:44:27Elia Trevisan, Alejandro Gonzalez-Garcia, Jan Swevers, Javier Alonso-MoraAbstract
Autonomous surface vessels operating in urban waterways must coordinate with other vessels in confined spaces while complying with traffic regulations, often without knowing their intentions. We present a decentralized motion planner based on model predictive path integral control, a sampling-based model predictive control scheme. The planner jointly computes a trajectory for the ego vessel and a prediction of the trajectories of the other vessels, and therefore accounts for their interaction. Traffic regulations are encoded as discontinuous costs, which the sampling-based formulation admits without a smooth approximation. We evaluate the method in maps taken from real sections of Amsterdam's canals, in encounters between two vessels and in scenarios with four interacting vessels. Across 780 randomized runs the method reaches every goal without collisions or deadlocks, and outperforms an optimization-based planner designed for the same domain, as well as a version of the proposed planner that does not account for the interaction. We further validate the planner on two scaled vessel prototypes in an outdoor environment, where planning and localization run onboard at 10 Hz, and head-on and crossing encounters are resolved without collisions or rule violations.
Abstract (translated)
URL
https://arxiv.org/abs/2610.41676
PDF
-
Enhancing Knowledge Entity Relation Extraction in Abstracts and Full-text Content of Academic Articles Using Large Language Models
2026-10-30 10:50:28Yingyi Zhang, Chengzhi ZhangAbstract
This study investigates the effectiveness of large pre-trained language models (LLMs) for knowledge entity relation extraction (KERE) from academic papers. Existing studies have examined individual types of LLM-based strategies and evaluated model performance using metrics such as precision and recall. To provide a broader evaluation, this study investigates multiple strategies, including closed-source LLM-based CoT prompting, LoRA-adapted open-source LLMs, LLM-enhanced small pre-trained language models (SLMs), and SLM-enhanced closed-source LLMs. In addition, this study introduces two strategies, namely SLM-enhanced open-source LLMs and closed-source LLM-enhanced open-source LLMs. These strategies are evaluated from multiple perspectives, including metrics, their ability to mitigate errors observed in SLMs, and potential directions for model optimization. This study manually constructs two annotated datasets in the NLP domain, covering Abstract and Full-Text datasets. Experimental results show that the LoRA-adapted T5 model achieves the highest macro F1 scores of 0.7408 and 0.7013 on the Abstract and Full-Text datasets, respectively. SLMs also benefit from LLM-generated entity descriptions under the LLM-enhanced SLM strategy. Error analysis further reveals that the effectiveness of LLM-based strategies in mitigating SLM errors varies across error types. Based on the identified limitations and error patterns of these strategies, several optimization directions are proposed. In addition, combining strategies with different error patterns further improves KERE performance.
Abstract (translated)
URL
https://arxiv.org/abs/2610.41552
PDF
-
Contact-based optimization for collision-free motion planning for a million robots
2026-10-30 10:46:26Jan Kamm, Antonio Terpin, Raffaello D'AndreaAbstract
Multi-agent robotic systems have attracted growing interest in recent years, with applications ranging from coordinated drone shows to autonomous operations on land, in the air, and on water. At the planning level, robot motion is often modeled using double-integrator dynamics (with low-level controllers handling the physical dynamics during execution) with a minimum-energy objective subject to collision-avoidance constraints and limits on position, velocity, and acceleration. The complexity of the problem resides in the constraints, which couple the trajectories and make the optimization problem nonconvex. Existing numerical approaches typically discretize and optimize entire trajectories, which becomes expensive for large fleets. We exploit the structure of optimal trajectories by organizing optimization around contacts, at which robots reach the minimum allowed separation. We formulate a coupled algebraic system in contact variables and reconstruct trajectories analytically between contacts. Our algorithm, sequential contact insertion (SCI), starts from unconstrained trajectories, inserts contacts at separation violations, and jointly adjusts their times and states. Numerical experiments demonstrate planning for one million robots in minutes and thousands of robots in fractions of a second. On problems feasible for a fixed-grid sequential convex programming baseline, SCI achieves median speedups of several orders of magnitude at matching solution accuracy.
Abstract (translated)
URL
https://arxiv.org/abs/2610.41548
PDF
-
FOCUS: First-Order Control Using Optimal Transport Supervision for Imitation from Observations
2026-10-30 10:37:43Jo\~ao A. C\^andido Ramos, Lionel Blond\'e, Hugues Vinzant, Jacopo Castellini, Alexandros KalousisAbstract
Imitation learning from observations must recover behavior from state-only demonstrations without known temporal correspondence. Differentiable simulators provide pathwise policy gradients, but differentiating through long, contact-rich trajectories can make optimization unstable. We introduce FOCUS, which combines local optimal-transport alignment with short-horizon differentiable actor-critic learning. FOCUS matches agent rollout windows to multiple sampled expert segments using a soft best-of-$K$ objective that emphasizes locally compatible matches. The resulting transport-based costs provide differentiable imitation rewards, while a learned critic bootstraps value beyond the rollout horizon. FOCUS requires neither temporally aligned trajectories nor expert-aligned initialization or teacher forcing. Across contact-rich locomotion tasks in DFlex, including a musculoskeletal humanoid controlled by 152 muscle activations, FOCUS improves sample efficiency and training stability over adversarial imitation-from-observation methods and prior differentiable imitation baselines.
Abstract (translated)
URL
https://arxiv.org/abs/2610.41540
PDF
-
Physically Constrained Kinematic Calibration for High-Precision Robotic Needle Guidance
2026-10-30 08:35:37Prabhu Nithin Gollapudi, Andreas Maier, Franziska Mathis-Ullrich, Water Zhao, Jia Wang, Siming BayerAbstract
Percutaneous needle insertion is a common practice in minimally invasive procedures whose clinical success depends on accurate needle placement. However, achieving the necessary precision is a challenge for manual operators. Robotic assistance can improve targeting accuracy and reduce operator dependence, but residual millimeter-scale errors persist because nominal kinematic models do not fully represent the physical manipulator. This paper presents a robust bound-constrained kinematic calibration method designed to improve positional accuracy for a 6-Degree-of-Freedom (6-DoF) robotic needle-guidance manipulator. The approach uses physically plausible parameter ranges to regularize a compact 19-parameter geometric identification model. By treating these bounds as active optimization constraints rather than mere numerical safeguards, the framework effectively mitigates parameter coupling and overfitting, ensuring highly stable parameter identification. In simulation, the proposed method reduced end-effector Root Mean Square Error (RMSE) from 3.422 mm to 0.690 mm (a 79.8% improvement). In benchtop hardware experiments, the calibrated model achieved sub-millimeter validation accuracy, including 0.865 mm RMSE in experimental trajectory validation and 0.610 mm RMSE in a coordinate-lattice verification task. These results demonstrate that physically constrained kinematic calibration can enhance robotic needle-targeting accuracy, establishing a high-precision foundation for future clinical applications.
Abstract (translated)
URL
https://arxiv.org/abs/2610.41429
PDF
-
Geometry-Aware Humanoid Retargeting via Manifold ODEs
2026-10-30 07:06:57Pengyu Zhu, Heye Wang, Xi Wang, Xuyang Wu, He Kong, Guodong ShiAbstract
In this paper, we focus on interaction-preserving motion retargeting, which transfers human demonstrations to humanoids while maintaining spatial relationships with objects and terrain. Previous work, such as OmniRetarget, addresses this task through temporally coupled constrained optimization, requiring repeated sequential quadratic programming (SQP) iterations at each frame. We study this problem from an ordinary differential equation (ODE) perspective by modeling the continuous-time limit of its framewise Karush-Kuhn-Tucker (KKT) conditions as an ODE on the robot configuration manifold. This viewpoint offers two benefits: (i) it connects OmniRetarget's SQP iterations to a backward discretization of the ODE, enabling us to establish their local convergence under suitable regularity assumptions; and (ii) it leads to TaSP-KKT (tangent-space parametrization for KKT flow), a new retargeting method based on a forward discretization of the KKT flow. The proposed method achieves the same first-order convergence to the ODE solution as the corresponding backward scheme while requiring only one weighted tangent-cone projection and a closed-form manifold update per step, thereby avoiding repeated inner SQP iterations. Numerical experiments show that TaSP-KKT achieves retargeting quality comparable to, though slightly lower than, that of OmniRetarget while running approximately twice as fast.
Abstract (translated)
URL
https://arxiv.org/abs/2610.41349
PDF
-
TRACE: Target-State Risk-Aligned Conditional Execution for Embodied Tracking
2026-10-30 05:26:41Mingyang Lyu, Gongce Cao, Yiyang Jia, Lu Jia, Erliang Lin, Yuhao Zhang, Yinqian Sun, Pin Lyu, Feifei Zhao, Bo WangAbstract
Embodied visual tracking requires a robot to follow a language-specified person under occlusion, look-alike distractors, and ambiguous instructions. We show that improving tracking persistence can simultaneously increase safety violations. Using a compact 0.6B vision-language-action (VLA) tracker with a frozen backbone, we update only the planner head with proximal policy optimization (PPO). PPO improves success and following rates across all three EVT-Bench settings, but increases the collision rate on distracted tracking because episodes that previously ended in target loss can instead end in contact. We introduce TRACE, an execution-time target-state risk gate that preserves the learned policy and intervenes selectively. A lightweight risk head processes 84 cues, including candidate boxes, identity margins, a learned range-and-bearing probe, and the policy's commands. When the calibrated risk exceeds a threshold, TRACE rescales the translational component of the next waypoint while leaving yaw unchanged. The executed action is identical to PPO's on more than 98% of steps. In paired evaluations on identical episodes, TRACE reduces the distracted-tracking collision rate by 3.96 percentage points, while further improving success and following rates. Matched slowdown controls without target-state timing do not reproduce this improvement. Concurrent evaluation reproduces a collision reduction of 3.77 points, compared with 0.29 points for identical-arm null pairs, indicating that the effect exceeds evaluation variability. These results show that target-state-aware execution can recover the safety cost of policy post-training with minimal intervention.
Abstract (translated)
URL
https://arxiv.org/abs/2610.41275
PDF
-
Beyond Scarcity: When Encoder Adaptation Is Worthwhile
2026-10-30 03:49:33Zexing Zhang, Jichao Li, Tianyang Lei, Yiyin Zhang, Kewei YangAbstract
Long-tailed recognition often learns a shared representation before correcting the classifier for class imbalance. Once the classifier has been rebalanced, however, final accuracy does not reveal whether further encoder adaptation remains worthwhile. Standard imbalance benchmarks obscure this question because they jointly vary the number of distinct examples available to rare classes and the frequency with which those examples shape the representation. We introduce a controlled evaluation framework that separates unique support from representation allocation and estimates the incremental value of encoder updates through matched classifier-only and full-network continuations from a common corrected model. The framework combines stage-specific controls, multiple representation and adaptation objectives, whole-distribution metrics, and hierarchical inference over class assignments, support reconstructions, and optimization runs. Across the controlled study, the same low-support regime moves from one in which classifier-only training is adequate to one in which encoder adaptation yields a practically meaningful benefit solely as representation allocation becomes more imbalanced. The transition persists across classifier corrections, representation objectives, normalization and augmentation controls, and ImageNet-scale evaluation, while preserving frequent-class accuracy. Under the prespecified one-percentage-point operating margin, the two regimes are separated by simultaneous confidence bounds. These results show that benchmark construction can alter a post-correction training decision and that support size alone does not characterize the remaining value of representation adaptation.
Abstract (translated)
URL
https://arxiv.org/abs/2610.41213
PDF
-
A Capability-Aware Backup-Recoverability Filter for Heterogeneous Multi-Robot Navigation
2026-10-30 03:11:02Joonkyung Kim, Yanze Zhang, Wenhao Luo, Yiwei LyuAbstract
Safe navigation in heterogeneous multi-robot teams faces two coupled challenges: differences in robots' braking and turning capabilities limit how each robot can respond to an approaching neighbor, while simultaneous interactions require a single control action to satisfy multiple collision-avoidance constraints. Together, these challenges can make local safety constraints incompatible with a robot's admissible controls, leaving a decentralized safety controller without a feasible action. We propose a Capability-Aware Backup-Recoverability Filter (CA-BRF), based on the insight that navigation actions should preserve access to a certified recovery maneuver. First, we construct a backup-recoverable set in a shared Cartesian operational space by evaluating whether prescribed backup maneuvers maintain pairwise safety until the robots come to rest. Each maneuver respects the corresponding robot's kinematic and actuation constraints, making the recovery certificate capability-dependent. Second, CA-BRF filters candidate controls by selecting admissible interpolations between candidate and backup actions while coordinating the selection across simultaneous interactions. This preserves backup recoverability while retaining as much of the candidate action as possible and ensures that all pairwise recovery conditions are satisfied by a single executable control for each robot. Together, we establish recursive feasibility and safety for the discrete-time closed-loop system even when the nominal safety-constrained optimization becomes infeasible. The validity and effectiveness of the proposed CA-BRF are validated in simulations with up to 30 heterogeneous robots and experiments of robots with five different kinematic classes.
Abstract (translated)
URL
https://arxiv.org/abs/2610.41188
PDF
-
Lineage Is Memory: Parent Inheritance and Feedback Disclosure in Short-Budget LLM-Guided Optimization
Abstract
Iterative language-model optimizers refine prompts and programs by evaluating a candidate, showing a proposer model a sample of failing cases, and keeping revisions that improve performance. Recent work attributes the limits of this loop to narrow single-round reflection, introducing full-set error clustering, targeted minibatches, or larger proposer models so the optimizer sees more failure modes per step. Yet every existing evaluation tests these mechanisms exclusively inside lineage-preserving search, where each revision edits a surviving parent, leaving the contribution of parent inheritance confounded with the breadth of per-round feedback. Across 1,750 pre-registered runs spanning nine prompt and program synthesis benchmarks, we separate parent inheritance from feedback disclosure under matched evaluation budgets. We find that greedy parent inheritance drives the vast majority of optimization progress, outperforming seed-reset resampling even when seed-reset proposals receive a clustered summary of all failures in every round or use a larger proposer model. While full-set error clustering and diverse failure selection improve single-step rewrites from the seed and accelerate the opening rounds of search, sequential parent inheritance rapidly closes this gap as solved error modes are retained in the candidate and filtered from subsequent failure lists. Moreover, decomposing token expenditure in GEPA reveals that proposer reflection accounts for only a small fraction of total cost whereas validation and minibatch gating consume nearly the entire budget, showing that greedy parent inheritance over raw failures captures the gains of complex reflective loops at a fraction of their compute.
Abstract (translated)
URL
https://arxiv.org/abs/2610.41157
PDF
-
HyperRepair: Recovering Forgotten General Capabilities of Fine-Tuned Language Models
2026-10-30 02:09:34Junrong Tang, Jianghui Wang, Hao Kong, Yang Lin, Yibo YangAbstract
Fine-tuning improves a language model's performance in a target domain but can weaken its general capabilities. Existing methods often face a trade-off between general capabilities and domain specialization, making it difficult to preserve both. We introduce HyperRepair, a framework for recovering general capabilities while retaining domain specialization at only inference cost. To make recovery generalizable across models sharing a pretrained backbone, HyperRepair learns a hypernetwork from repairs trained on source models. The hypernetwork uses activation fingerprints that summarize changes from the pretrained backbone to generate a tailored Repairing Side Net (RSN), whose prompt-conditioned gate applies corrections selectively. This enables recovery on models from unseen adaptation domains without target-domain examples or additional optimization. We evaluate HyperRpair across two model families, 12 LoRA adaptation datasets, and 15 general benchmarks. On the default split, HyperRepair achieves 93.4% of the base models' average general score while retaining 99.5% of the fine-tuned models' average domain score.
Abstract (translated)
URL
https://arxiv.org/abs/2610.41147
PDF
-
MUON-RT Reusing the Teacher's Trajectory for Momentum-Free Muon for LLM Self-Distillation
2026-10-30 00:50:49Aryan Das, Vinay Kumar Verma, Aakansha MishraAbstract
Self-distillation methods for large language models with an exponential moving average (EMA) teacher already retain a history of student updates. Muon nevertheless stores a separate momentum buffer, increasing persistent training memory. We propose Muon with Reused Teacher (Muon-RT), which reuses the teacher-student displacement to eliminate momentum buffers for hidden weight matrices. We show that this displacement encodes an exponentially weighted history of applied updates and smooths temporally uncorrelated noise under a local model. Muon-RT removes its radial component, normalizes it spectrally, and mixes it with the current gradient after Newton-Schulz (NS) orthogonalization. Under explicit regularity and alignment assumptions, we establish linear contraction of an expected objective-gap bound toward a residual error level. With Self-Distillation Fine-Tuning (SDFT) on Qwen2.5-7B-Instruct, Muon-RT improves average accuracy over AdamW by 2.6 points with 42% less persistent memory, and over momentum Muon by 5.4 points with 27% less persistent memory. With Self-Distillation Policy Optimization (SDPO) on Olmo-3-7B-Instruct, it improves average accuracy over AdamW by 2.0 points.
Abstract (translated)
URL
https://arxiv.org/abs/2610.41104
PDF
