Lessons from Training Grounded LLMs with Verifiable Rewards

Generating grounded and trustworthy responses remains a key challenge for large language models (LLMs). While retrieval-augmented generation (RAG) with citation-based grounding holds promise, instruction-tuned models frequently fail even in straightforward scenarios: missing explicitly stated answers, citing incorrectly, or refusing when evidence is available. In this work, we explore how reinforcement learning (RL) and internal reasoning can enhance grounding in LLMs. We use the GRPO (Group Relative Policy Optimization) method to train models using verifiable outcome-based rewards targeting answer correctness, citation sufficiency, and refusal quality, without requiring gold reasoning traces or expensive annotations. Through comprehensive experiments across ASQA, QAMPARI, ELI5, and ExpertQA we show that reasoning-augmented models significantly outperform instruction-only variants, especially in handling unanswerable queries and generating well-cited responses. A two-stage training setup, first optimizing answer and citation behavior and then refusal, further improves grounding by stabilizing the learning signal. Additionally, we revisit instruction tuning via GPT-4 distillation and find that combining it with GRPO enhances performance on long-form, generative QA tasks. Overall, our findings highlight the value of reasoning, stage-wise optimization, and outcome-driven RL for building more verifiable and reliable LLMs.

DOVA-PATBM: An Intelligent, Adaptive, and Scalable Framework for Optimizing Large-Scale EV Charging Infrastructure

The accelerating uptake of battery-electric vehicles demands infrastructure planning tools that are both data-rich and geographically scalable. Whereas most prior studies optimise charging locations for single cities, state-wide and national networks must reconcile the conflicting requirements of dense metropolitan cores, car-dependent exurbs, and power-constrained rural corridors. We present DOVA-PATBM (Deployment Optimisation with Voronoi-oriented, Adaptive, POI-Aware Temporal Behaviour Model), a geo-computational framework that unifies these contexts in a single pipeline. The method rasterises heterogeneous data (roads, population, night lights, POIs, and feeder lines) onto a hierarchical H3 grid, infers intersection importance with a zone-normalised graph neural network centrality model, and overlays a Voronoi tessellation that guarantees at least one five-port DC fast charger within every 30 km radius. Hourly arrival profiles, learned from loop-detector and floating-car traces, feed a finite M/M/c queue to size ports under feeder-capacity and outage-risk constraints. A greedy maximal-coverage heuristic with income-weighted penalties then selects the minimum number of sites that satisfy coverage and equity targets. Applied to the State of Georgia, USA, DOVA-PATBM (i) increases 30 km tile coverage by 12 percentage points, (ii) halves the mean distance that low-income residents travel to the nearest charger, and (iii) meets sub-transmission headroom everywhere -- all while remaining computationally tractable for national-scale roll-outs. These results demonstrate that a tightly integrated, GNN-driven, multi-resolution approach can bridge the gap between academic optimisation and deployable infrastructure policy.

AttentionDrag: Exploiting Latent Correlation Knowledge in Pre-trained Diffusion Models for Image Editing

Traditional point-based image editing methods rely on iterative latent optimization or geometric transformations, which are either inefficient in their processing or fail to capture the semantic relationships within the image. These methods often overlook the powerful yet underutilized image editing capabilities inherent in pre-trained diffusion models. In this work, we propose a novel one-step point-based image editing method, named AttentionDrag, which leverages the inherent latent knowledge and feature correlations within pre-trained diffusion models for image editing tasks. This framework enables semantic consistency and high-quality manipulation without the need for extensive re-optimization or retraining. Specifically, we reutilize the latent correlations knowledge learned by the self-attention mechanism in the U-Net module during the DDIM inversion process to automatically identify and adjust relevant image regions, ensuring semantic validity and consistency. Additionally, AttentionDrag adaptively generates masks to guide the editing process, enabling precise and context-aware modifications with friendly interaction. Our results demonstrate a performance that surpasses most state-of-the-art methods with significantly faster speeds, showing a more efficient and semantically coherent solution for point-based image editing tasks.

Natural Language Guided Ligand-Binding Protein Design

Can AI protein models follow human language instructions and design proteins with desired functions (e.g. binding to a ligand)? Designing proteins that bind to a given ligand is crucial in a wide range of applications in biology and chemistry. Most prior AI models are trained on protein-ligand complex data, which is scarce due to the high cost and time requirements of laboratory experiments. In contrast, there is a substantial body of human-curated text descriptions about protein-ligand interactions and ligand formula. In this paper, we propose InstructPro, a family of protein generative models that follow natural language instructions to design ligand-binding proteins. Given a textual description of the desired function and a ligand formula in SMILES, InstructPro generates protein sequences that are functionally consistent with the specified instructions. We develop the model architecture, training strategy, and a large-scale dataset, InstructProBench, to support both training and evaluation. InstructProBench consists of 9,592,829 triples of (function description, ligand formula, protein sequence). We train two model variants: InstructPro-1B (with 1 billion parameters) and InstructPro-3B~(with 3 billion parameters). Both variants consistently outperform strong baselines, including ProGen2, ESM3, and Pinal. Notably, InstructPro-1B achieves the highest docking success rate (81.52% at moderate confidence) and the lowest average root mean square deviation (RMSD) compared to ground truth structures (4.026{\AA}). InstructPro-3B further descreases the average RMSD to 2.527{\AA}, demonstrating InstructPro's ability to generate ligand-binding proteins that align with the functional specifications.

Error Typing for Smarter Rewards: Improving Process Reward Models with Error-Aware Hierarchical Supervision

Large Language Models (LLMs) are prone to hallucination, especially during multi-hop and reasoning-intensive tasks such as mathematical problem solving. While Outcome Reward Models verify only final answers, Process Reward Models (PRMs) score each intermediate step to steer generation toward coherent solutions. We introduce PathFinder-PRM, a novel hierarchical, error-aware discriminative PRM that first classifies math and consistency errors at each step, then combines these fine-grained signals to estimate step correctness. To train PathFinder-PRM, we construct a 400K-sample dataset by enriching the human-annotated PRM800K corpus and RLHFlow Mistral traces with three-dimensional step-level labels. On PRMBench, PathFinder-PRM achieves a new state-of-the-art PRMScore of 67.7, outperforming the prior best (65.5) while using 3 times less data. When applied to reward guided greedy search, our model yields prm@8 48.3, a +1.5 point gain over the strongest baseline. These results demonstrate that decoupled error detection and reward estimation not only boost fine-grained error detection but also substantially improve end-to-end, reward-guided mathematical reasoning with greater data efficiency.

VeriFastScore: Speeding up long-form factuality evaluation

Metrics like FactScore and VeriScore that evaluate long-form factuality operate by decomposing an input response into atomic claims and then individually verifying each claim. While effective and interpretable, these methods incur numerous LLM calls and can take upwards of 100 seconds to evaluate a single response, limiting their practicality in large-scale evaluation and training scenarios. To address this, we propose VeriFastScore, which leverages synthetic data to fine-tune Llama3.1 8B for simultaneously extracting and verifying all verifiable claims within a given text based on evidence from Google Search. We show that this task cannot be solved via few-shot prompting with closed LLMs due to its complexity: the model receives ~4K tokens of evidence on average and needs to concurrently decompose claims, judge their verifiability, and verify them against noisy evidence. However, our fine-tuned VeriFastScore model demonstrates strong correlation with the original VeriScore pipeline at both the example level (r=0.80) and system level (r=0.94) while achieving an overall speedup of 6.6x (9.9x excluding evidence retrieval) over VeriScore. To facilitate future factuality research, we publicly release our VeriFastScore model and synthetic datasets.

BLEUBERI: BLEU is a surprisingly effective reward for instruction following

Reward models are central to aligning LLMs with human preferences, but they are costly to train, requiring large-scale human-labeled preference data and powerful pretrained LLM backbones. Meanwhile, the increasing availability of high-quality synthetic instruction-following datasets raises the question: can simpler, reference-based metrics serve as viable alternatives to reward models during RL-based alignment? In this paper, we show first that BLEU, a basic string-matching metric, surprisingly matches strong reward models in agreement with human preferences on general instruction-following datasets. Based on this insight, we develop BLEUBERI, a method that first identifies challenging instructions and then applies Group Relative Policy Optimization (GRPO) using BLEU directly as the reward function. We demonstrate that BLEUBERI-trained models are competitive with models trained via reward model-guided RL across four challenging instruction-following benchmarks and three different base language models. A human evaluation further supports that the quality of BLEUBERI model outputs is on par with those from reward model-aligned models. Moreover, BLEUBERI models generate outputs that are more factually grounded than competing methods. Overall, we show that given access to high-quality reference outputs (easily obtained via existing instruction-following datasets or synthetic data generation), string matching-based metrics are cheap yet effective proxies for reward models during alignment. We release our code and data at https://github.com/lilakk/BLEUBERI.

A 2D Semantic-Aware Position Encoding for Vision Transformers

Vision transformers have demonstrated significant advantages in computer vision tasks due to their ability to capture long-range dependencies and contextual relationships through self-attention. However, existing position encoding techniques, which are largely borrowed from natural language processing, fail to effectively capture semantic-aware positional relationships between image patches. Traditional approaches like absolute position encoding and relative position encoding primarily focus on 1D linear position relationship, often neglecting the semantic similarity between distant yet contextually related patches. These limitations hinder model generalization, translation equivariance, and the ability to effectively handle repetitive or structured patterns in images. In this paper, we propose 2-Dimensional Semantic-Aware Position Encoding ($\text{SaPE}^2$), a novel position encoding method with semantic awareness that dynamically adapts position representations by leveraging local content instead of fixed linear position relationship or spatial coordinates. Our method enhances the model's ability to generalize across varying image resolutions and scales, improves translation equivariance, and better aggregates features for visually similar but spatially distant patches. By integrating $\text{SaPE}^2$ into vision transformers, we bridge the gap between position encoding and perceptual similarity, thereby improving performance on computer vision tasks.

NORA: A Small Open-Sourced Generalist Vision Language Action Model for Embodied Tasks

Existing Visual-Language-Action (VLA) models have shown promising performance in zero-shot scenarios, demonstrating impressive task execution and reasoning capabilities. However, a significant challenge arises from the limitations of visual encoding, which can result in failures during tasks such as object grasping. Moreover, these models typically suffer from high computational overhead due to their large sizes, often exceeding 7B parameters. While these models excel in reasoning and task planning, the substantial computational overhead they incur makes them impractical for real-time robotic environments, where speed and efficiency are paramount. To address the limitations of existing VLA models, we propose NORA, a 3B-parameter model designed to reduce computational overhead while maintaining strong task performance. NORA adopts the Qwen-2.5-VL-3B multimodal model as its backbone, leveraging its superior visual-semantic understanding to enhance visual reasoning and action grounding. Additionally, our \model{} is trained on 970k real-world robot demonstrations and equipped with the FAST+ tokenizer for efficient action sequence generation. Experimental results demonstrate that NORA outperforms existing large-scale VLA models, achieving better task performance with significantly reduced computational overhead, making it a more practical solution for real-time robotic autonomy.

PromptDistill: Query-based Selective Token Retention in Intermediate Layers for Efficient Large Language Model Inference

As large language models (LLMs) tackle increasingly complex tasks and longer documents, their computational and memory costs during inference become a major bottleneck. To address this, we propose PromptDistill, a novel, training-free method that improves inference efficiency while preserving generation quality. PromptDistill identifies and retains the most informative tokens by leveraging attention interactions in early layers, preserving their hidden states while reducing the computational burden in later layers. This allows the model to focus on essential contextual information without fully processing all tokens. Unlike previous methods such as H2O and SnapKV, which perform compression only after processing the entire input, or GemFilter, which selects a fixed portion of the initial prompt without considering contextual dependencies, PromptDistill dynamically allocates computational resources to the most relevant tokens while maintaining a global awareness of the input. Experiments using our method and baseline approaches with base models such as LLaMA 3.1 8B Instruct, Phi 3.5 Mini Instruct, and Qwen2 7B Instruct on benchmarks including LongBench, InfBench, and Needle in a Haystack demonstrate that PromptDistill significantly improves efficiency while having minimal impact on output quality compared to the original models. With a single-stage selection strategy, PromptDistill effectively balances performance and efficiency, outperforming prior methods like GemFilter, H2O, and SnapKV due to its superior ability to retain essential information. Specifically, compared to GemFilter, PromptDistill achieves an overall $1\%$ to $5\%$ performance improvement while also offering better time efficiency. Additionally, we explore multi-stage selection, which further improves efficiency while maintaining strong generation performance.