Ruosen Li

Ruosen Li, Ruochen Li, Barry Wang et al. · cmu

To evaluate Large Language Models (LLMs) for question answering (QA), traditional methods typically focus on assessing single-turn responses to given questions. However, this approach doesn't capture the dynamic nature of human-AI interactions, where humans actively seek information through conversation. Recent works in human-computer interaction (HCI) have employed human evaluators to conduct interactions and evaluations, but they are often prohibitively expensive and time-consuming to scale. We introduce an automatic evaluation framework IQA-EVAL to achieve Interactive Question Answering Evaluations, more specifically, we introduce a LLM-based Evaluation Agent (LEA) that can: (1) simulate human behaviors to generate interactions with IQA models; (2) automatically evaluate the generated interactions. Moreover, we propose assigning personas to LEAs to better simulate groups of real human evaluators. We show that: (1) our evaluation framework with GPT-4 (or Claude) as the backbone model achieves a high correlation with human evaluations on the IQA task; (2) assigning personas to LEA to better represent the crowd further significantly improves correlations. Finally, we use our automatic metric to evaluate five recent representative LLMs with over 1000 questions from complex and ambiguous question answering tasks, which comes with a substantial cost of $5k if evaluated by humans.

Ruosen Li, Teerth Patel, Xinya Du

Nowadays, the quality of responses generated by different modern large language models (LLMs) is hard to evaluate and compare automatically. Recent studies suggest and predominantly use LLMs for reference-free evaluation of open-ended question answering. More specifically, they use the recognized "strongest" LLM as the evaluator, which conducts pairwise comparisons of candidate models' answers and provides a ranking score. However, this intuitive method has multiple problems, such as bringing in self-enhancement (favoring its own answers) and positional bias. We draw insights and lessons from the educational domain (Cho & MacArthur, 2011; Walsh, 2014) to improve LLM-based evaluations. Specifically, we propose (1) the peer rank (PR) algorithm that takes into account each peer LLM's pairwise preferences of all answer pairs, and outputs a final ranking of models; and (2) peer discussion (PD), where we prompt two LLMs to discuss and try to reach a mutual agreement on the preferences of two answers. We conduct experiments on two benchmark datasets. We find that our approaches achieve higher accuracy and align better with human judgments. Interestingly, PR can induce a relatively accurate self-ranking of models under the anonymous setting, where each model's name is unrevealed. Our work provides space to explore evaluating models that are hard to compare for humans.

Liqiang Jing, Ruosen Li, Yunmo Chen et al.

We introduce FaithScore (Faithfulness to Atomic Image Facts Score), a reference-free and fine-grained evaluation metric that measures the faithfulness of the generated free-form answers from large vision-language models (LVLMs). The FaithScore evaluation first identifies sub-sentences containing descriptive statements that need to be verified, then extracts a comprehensive list of atomic facts from these sub-sentences, and finally conducts consistency verification between fine-grained atomic facts and the input image. Meta-evaluation demonstrates that our metric highly correlates with human judgments of faithfulness. We collect two benchmark datasets (i.e. LLaVA-1k and MSCOCO-Cap) for evaluating LVLMs instruction-following hallucinations. We measure hallucinations in state-of-the-art LVLMs with FaithScore on the datasets. Results reveal that current systems are prone to generate hallucinated content unfaithful to the image, which leaves room for future improvements. We hope our metric FaithScore can help evaluate future LVLMs in terms of faithfulness and provide insightful advice for enhancing LVLMs' faithfulness.

Ruosen Li, Xinya Du

Neural models, including large language models (LLMs), achieve superior performance on multi-hop question-answering. To elicit reasoning capabilities from LLMs, recent works propose using the chain-of-thought (CoT) mechanism to generate both the reasoning chain and the answer, which enhances the model's capabilities in conducting multi-hop reasoning. However, several challenges still remain: such as struggling with inaccurate reasoning, hallucinations, and lack of interpretability. On the other hand, information extraction (IE) identifies entities, relations, and events grounded to the text. The extracted structured information can be easily interpreted by humans and machines (Grishman, 2019). In this work, we investigate constructing and leveraging extracted semantic structures (graphs) for multi-hop question answering, especially the reasoning process. Empirical results and human evaluations show that our framework: generates more faithful reasoning chains and substantially improves the QA performance on two benchmark datasets. Moreover, the extracted structures themselves naturally provide grounded explanations that are preferred by humans, as compared to the generated reasoning chains and saliency-based explanations.

Chengsong Huang, Wenhao Yu, Xiaoyang Wang et al.

Self-evolving Large Language Models (LLMs) offer a scalable path toward super-intelligence by autonomously generating, refining, and learning from their own experiences. However, existing methods for training such models still rely heavily on vast human-curated tasks and labels, typically via fine-tuning or reinforcement learning, which poses a fundamental bottleneck to advancing AI systems toward capabilities beyond human intelligence. To overcome this limitation, we introduce R-Zero, a fully autonomous framework that generates its own training data from scratch. Starting from a single base LLM, R-Zero initializes two independent models with distinct roles, a Challenger and a Solver. These models are optimized separately and co-evolve through interaction: the Challenger is rewarded for proposing tasks near the edge of the Solver capability, and the Solver is rewarded for solving increasingly challenging tasks posed by the Challenger. This process yields a targeted, self-improving curriculum without any pre-existing tasks and labels. Empirically, R-Zero substantially improves reasoning capability across different backbone LLMs, e.g., boosting the Qwen3-4B-Base by +6.49 on math-reasoning benchmarks and +7.54 on general-domain reasoning benchmarks.

Ruosen Li, Ziming Luo, Quan Zhang et al.

Large reasoning models (LRMs) achieve impressive reasoning capabilities by generating lengthy chain-of-thoughts, but this "overthinking" incurs high latency and cost without commensurate accuracy gains. In this work, we introduce AALC, a lightweight, accuracy-aware length reward integrated into reinforcement learning that dynamically balances correctness and brevity during training. By incorporating validation accuracy into the reward and employing a smooth, dynamically scheduled length penalty, AALC delays length penalty until target performance is met. Through extensive experiments across standard and out-of-distribution math benchmarks, we show that our approach reduces response length by over 50% while maintaining or even improving the original accuracy. Furthermore, qualitative analysis reveals that our method curbs redundant reasoning patterns such as excessive subgoal setting and verification, leading to structurally refined outputs rather than naive truncation. We also identify that efficiency gains are accompanied by reduced interpretability: models trained with AALC omit some narrative framing and explanatory context. These findings highlight the potential of reward-based strategies to guide LRMs toward more efficient, generalizable reasoning paths.

Yangxiao Lu, Ruosen Li, Liqiang Jing et al.

Visual grounding focuses on detecting objects from images based on language expressions. Recent Large Vision-Language Models (LVLMs) have significantly advanced visual grounding performance by training large models with large-scale datasets. However, the problem remains challenging, especially when similar objects appear in the input image. For example, an LVLM may not be able to differentiate Diet Coke and regular Coke in an image. In this case, if additional reference images of Diet Coke and regular Coke are available, it can help the visual grounding of similar objects. In this work, we introduce a new task named Multimodal Reference Visual Grounding (MRVG). In this task, a model has access to a set of reference images of objects in a database. Based on these reference images and a language expression, the model is required to detect a target object from a query image. We first introduce a new dataset to study the MRVG problem. Then we introduce a novel method, named MRVG-Net, to solve this visual grounding problem. We show that by efficiently using reference images with few-shot object detection and using Large Language Models (LLMs) for object matching, our method achieves superior visual grounding performance compared to the state-of-the-art LVLMs such as Qwen2.5-VL-72B. Our approach bridges the gap between few-shot detection and visual grounding, unlocking new capabilities for visual understanding, which has wide applications in robotics. Project page with our video, code, and dataset: https://irvlutd.github.io/MultiGrounding

Zhenwen Liang, Ruosen Li, Yujun Zhou et al.

Assessing the quality of Large Language Model (LLM) outputs presents a critical challenge. Previous methods either rely on text-level information (e.g., reward models, majority voting), which can overfit to superficial cues, or on calibrated confidence from token probabilities, which would fail on less-calibrated models. Yet both of these signals are, in fact, partial projections of a richer source of information: the model's internal hidden states. Early layers, closer to token embeddings, preserve semantic and lexical features that underpin text-based judgments, while later layers increasingly align with output logits, embedding confidence-related information. This paper explores hidden states directly as a unified foundation for verification. We show that the correctness of a solution is encoded as a geometrically separable signature within the trajectory of hidden activations. To validate this, we present Clue (Clustering and Experience-based Verification), a deliberately minimalist, non-parametric verifier. With no trainable parameters, CLUE only summarizes each reasoning trace by an hidden state delta and classifies correctness via nearest-centroid distance to ``success'' and ``failure'' clusters formed from past experience. The simplicity of this method highlights the strength of the underlying signal. Empirically, CLUE consistently outperforms LLM-as-a-judge baselines and matches or exceeds modern confidence-based methods in reranking candidates, improving both top-1 and majority-vote accuracy across AIME 24/25 and GPQA. As a highlight, on AIME 24 with a 1.5B model, CLUE boosts accuracy from 56.7% (majority@64) to 70.0% (top-maj@16).