Personalized RewardBench: Evaluating Reward Models with Human Aligned Personalization

Pluralistic alignment has emerged as a critical frontier in the development of Large Language Models (LLMs), with reward models (RMs) serving as a central mechanism for capturing diverse human values. While benchmarks for general response quality are prevalent, evaluating how well reward models account for individual user preferences remains an open challenge. To bridge this gap, we introduce Personalized RewardBench, a novel benchmark designed to rigorously assess reward models' capacity to model personalized preferences. We construct chosen and rejected response pairs based on strict adherence to (or violation of) user-specific rubrics, ensuring that preference distinctions are uniquely tailored to the individual. In particular, human evaluations confirm that the primary discriminative factor between pairs is strictly personal preference, with both responses maintaining high general quality (e.g., correctness, relevance and helpfulness). Extensive testing reveals that existing state-of-the-art reward models struggle significantly with personalization, peaking at an accuracy of just 75.94%. Crucially, because an effective reward model benchmark should predict a reward model's performance on downstream tasks, we conduct experiments demonstrating that our benchmark exhibits a significantly higher correlation with downstream performance in both Best-of-N (BoN) sampling and Proximal Policy Optimization (PPO) compared to existing baselines. These findings establish Personalized RewardBench as a robust and accurate proxy for evaluating reward models' performance in downstream applications.

ResearchEVO: An End-to-End Framework for Automated Scientific Discovery and Documentation

An important recurring pattern in scientific breakthroughs is a two-stage process: an initial phase of undirected experimentation that yields an unexpected finding, followed by a retrospective phase that explains why the finding works and situates it within existing theory. We present ResearchEVO, an end-to-end framework that computationally instantiates this discover-then-explain paradigm. The Evolution Phase employs LLM-guided bi-dimensional co-evolution -- simultaneously optimizing both algorithmic logic and overall architecture -- to search the space of code implementations purely by fitness, without requiring any understanding of the solutions it produces. The Writing Phase then takes the best-performing algorithm and autonomously generates a complete, publication-ready research paper through sentence-level retrieval-augmented generation with explicit anti-hallucination verification and automated experiment design. To our knowledge, ResearchEVO is the first system to cover this full pipeline end to end: no prior work jointly performs principled algorithm evolution and literature-grounded scientific documentation. We validate the framework on two cross-disciplinary scientific problems -- Quantum Error Correction using real Google quantum hardware data, and Physics-Informed Neural Networks -- where the Evolution Phase discovered human-interpretable algorithmic mechanisms that had not been previously proposed in the respective domain literatures. In both cases, the Writing Phase autonomously produced compilable LaTeX manuscripts that correctly grounded these blind discoveries in existing theory via RAG, with zero fabricated citations.

DebugLM: Learning Traceable Training Data Provenance for LLMs

Large language models (LLMs) are trained through multi-stage pipelines over heterogeneous data sources, yet developers lack a principled way to pinpoint the specific data responsible for an observed behavior. This lack of observability reduces debugging to reactive patching and makes failures prone to recur under distribution shift or subsequent model updates. To address this limitation, we propose DebugLM, a framework that equips LLMs with built-in data provenance, enabling them to explicitly trace the origins of their behaviors to specific training data sources. Specifically, the model learns to associate its responses with unique provenance tags that indicate the responsible dataset, empowering developers to precisely identify where undesirable behaviors are learned. Building on this capability, DebugLM further supports targeted test-time remediation, enabling developers to selectively trigger targeted refusal for specified data sources without retraining or modifying model parameters. Experiments demonstrate that DebugLM provides accurate behavior tracing in multi-stage training pipelines and effective test-time remediation while preserving the general utility of the model.

Logos: An evolvable reasoning engine for rational molecular design

The discovery and design of functional molecules remain central challenges across chemistry,biology, and materials science. While recent advances in machine learning have accelerated molecular property prediction and candidate generation, existing models tend to excel either in physical fidelity without transparent reasoning, or in flexible reasoning without guarantees of chemical validity. This imbalance limits the reliability of artificial intelligence systems in real scientific design workflows.Here we present Logos, a compact molecular reasoning model that integrates multi-step logical reasoning with strict chemical consistency. Logos is trained using a staged strategy that first exposes the model to explicit reasoning examples linking molecular descriptions to structural decisions, and then progressively aligns these reasoning patterns with molecular representations. In a final training phase, chemical rules and invariants are incorporated directly into the optimization objective, guiding the model toward chemically valid outputs. Across multiple benchmark datasets, Logos achieves strong performance in both structural accuracy and chemical validity, matching or surpassing substantially larger general-purpose language models while operating with a fraction of their parameters. Beyond benchmark evaluation, the model exhibits stable behaviour in molecular optimization tasks involving multiple, potentially conflicting constraints. By explicitly exposing intermediate reasoning steps, Logos enables human inspection and assessment of the design logic underlying each generated structure. These results indicate that jointly optimizing for reasoning structure and physical consistency offers a practical pathway toward reliable and interpretable AI systems for molecular science, supporting closer integration of artificial intelligence into scientific discovery processes.

EvoCodeBench: A Human-Performance Benchmark for Self-Evolving LLM-Driven Coding Systems

As large language models (LLMs) continue to advance in programming tasks, LLM-driven coding systems have evolved from one-shot code generation into complex systems capable of iterative improvement during inference. However, existing code benchmarks primarily emphasize static correctness and implicitly assume fixed model capability during inference. As a result, they do not capture inference-time self-evolution, such as whether accuracy and efficiency improve as an agent iteratively refines its solutions. They also provide limited accounting of resource costs and rarely calibrate model performance against that of human programmers. Moreover, many benchmarks are dominated by high-resource languages, leaving cross-language robustness and long-tail language stability underexplored. Therefore, we present EvoCodeBench, a benchmark for evaluating self-evolving LLM-driven coding systems across programming languages with direct comparison to human performance. EvoCodeBench tracks performance dynamics, measuring solution correctness alongside efficiency metrics such as solving time, memory consumption, and improvement algorithmic design over repeated problem-solving attempts. To ground evaluation in a human-centered reference frame, we directly compare model performance with that of human programmers on the same tasks, enabling relative performance assessment within the human ability distribution. Furthermore, EvoCodeBench supports multiple programming languages, enabling systematic cross-language and long-tail stability analyses under a unified protocol. Our results demonstrate that self-evolving systems exhibit measurable gains in efficiency over time, and that human-relative and multi-language analyses provide insights unavailable through accuracy alone. EvoCodeBench establishes a foundation for evaluating coding intelligence in evolving LLM-driven systems.

FaLW: A Forgetting-aware Loss Reweighting for Long-tailed Unlearning

Machine unlearning, which aims to efficiently remove the influence of specific data from trained models, is crucial for upholding data privacy regulations like the ``right to be forgotten". However, existing research predominantly evaluates unlearning methods on relatively balanced forget sets. This overlooks a common real-world scenario where data to be forgotten, such as a user's activity records, follows a long-tailed distribution. Our work is the first to investigate this critical research gap. We find that in such long-tailed settings, existing methods suffer from two key issues: \textit{Heterogeneous Unlearning Deviation} and \textit{Skewed Unlearning Deviation}. To address these challenges, we propose FaLW, a plug-and-play, instance-wise dynamic loss reweighting method. FaLW innovatively assesses the unlearning state of each sample by comparing its predictive probability to the distribution of unseen data from the same class. Based on this, it uses a forgetting-aware reweighting scheme, modulated by a balancing factor, to adaptively adjust the unlearning intensity for each sample. Extensive experiments demonstrate that FaLW achieves superior performance. Code is available at \textbf{Supplementary Material}.

CARPE: Context-Aware Image Representation Prioritization via Ensemble for Large Vision-Language Models

Recent advancements in Large Vision-Language Models (LVLMs) have pushed them closer to becoming general-purpose assistants. Despite their strong performance, LVLMs still struggle with vision-centric tasks such as image classification, underperforming compared to their base vision encoders, which are often CLIP-based models. To address this limitation, we propose Context-Aware Image Representation Prioritization via Ensemble (CARPE), a novel, model-agnostic framework which introduces vision-integration layers and a context-aware ensemble strategy to identify when to prioritize image representations or rely on the reasoning capabilities of the language model. This design enhances the model's ability to adaptively weight visual and textual modalities and enables the model to capture various aspects of image representations, leading to consistent improvements in generalization across classification and vision-language benchmarks. Extensive experiments demonstrate that CARPE not only improves performance on image classification benchmarks but also enhances results across various vision-language benchmarks. Finally, CARPE is designed to be effectively integrated with most open-source LVLMs that consist of a vision encoder and a language model, ensuring its adaptability across diverse architectures.

Closing the Reality Gap: Zero-Shot Sim-to-Real Deployment for Dexterous Force-Based Grasping and Manipulation

Human-like dexterous hands with multiple fingers offer human-level manipulation capabilities, but training control policies that can directly deploy on real hardware remains difficult due to contact-rich physics and imperfect actuation. We close this gap with a practical sim-to-real reinforcement learning (RL) framework that utilizes dense tactile feedback combined with joint torque sensing to explicitly regulate physical interactions. To enable effective sim-to-real transfer, we introduce (i) a computationally fast tactile simulation that computes distances between dense virtual tactile units and the object via parallel forward kinematics, providing high-rate, high-resolution touch signals needed by RL; (ii) a current-to-torque calibration that eliminates the need for torque sensors on dexterous hands by mapping motor current to joint torque; and (iii) actuator dynamics modeling to bridge the actuation gaps with randomization of non-ideal effects such as backlash, torque-speed saturation. Using an asymmetric actor-critic PPO pipeline trained entirely in simulation, our policies deploy directly to a five-finger hand. The resulting policies demonstrated two essential skills: (1) command-based, controllable grasp force tracking, and (2) reorientation of objects in the hand, both of which were robustly executed without fine-tuning on the robot. By combining tactile and torque in the observation space with effective sensing/actuation modeling, our system provides a practical solution to achieve reliable dexterous manipulation. To our knowledge, this is the first demonstration of controllable grasping on a multi-finger dexterous hand trained entirely in simulation and transferred zero-shot on real hardware.

Rethinking Crystal Symmetry Prediction: A Decoupled Perspective

Efficiently and accurately determining the symmetry is a crucial step in the structural analysis of crystalline materials. Existing methods usually mindlessly apply deep learning models while ignoring the underlying chemical rules. More importantly, experiments show that they face a serious sub-property confusion SPC problem. To address the above challenges, from a decoupled perspective, we introduce the XRDecoupler framework, a problem-solving arsenal specifically designed to tackle the SPC problem. Imitating the thinking process of chemists, we innovatively incorporate multidimensional crystal symmetry information as superclass guidance to ensure that the model's prediction process aligns with chemical intuition. We further design a hierarchical PXRD pattern learning model and a multi-objective optimization approach to achieve high-quality representation and balanced optimization. Comprehensive evaluations on three mainstream databases (e.g., CCDC, CoREMOF, and InorganicData) demonstrate that XRDecoupler excels in performance, interpretability, and generalization.

Domain Generalization in-the-Wild: Disentangling Classification from Domain-Aware Representations

Evaluating domain generalization (DG) for foundational models like CLIP is challenging, as web-scale pretraining data potentially covers many existing benchmarks. Consequently, current DG evaluation may neither be sufficiently challenging nor adequately test genuinely unseen data scenarios. To better assess the performance of CLIP on DG in-the-wild, a scenario where CLIP encounters challenging unseen data, we consider two approaches: (1) evaluating on 33 diverse datasets with quantified out-of-distribution (OOD) scores after fine-tuning CLIP on ImageNet, and (2) using unlearning to make CLIP `forget' some domains as an approximation. We observe that CLIP's performance deteriorates significantly on more OOD datasets. To address this, we present CLIP-DCA (Disentangling Classification from enhanced domain Aware representations). Our approach is motivated by the observation that while standard domain invariance losses aim to make representations domain-invariant, this can be harmful to foundation models by forcing the discarding of domain-aware representations beneficial for generalization. We instead hypothesize that enhancing domain awareness is a prerequisite for effective domain-invariant classification in foundation models. CLIP-DCA identifies and enhances domain awareness within CLIP's encoders using a separate domain head and synthetically generated diverse domain data. Simultaneously, it encourages domain-invariant classification through disentanglement from the domain features. CLIP-DCA shows significant improvements within this challenging evaluation compared to existing methods, particularly on datasets that are more OOD.