Yanguang Liu

Shuang Liu, Yuxuan Bo, Qiuyang Zhao et al.

Reward models are central to large language model (LLM) alignment, but they remain vulnerable to reward hacking. To evaluate reward-model robustness, we introduce RewardHackBench containing 13 reward-hacking patterns covering real life high-stakes domains and general settings, and we find severe failures on specific subcategories across eight reward models. To mitigate these failures, we propose HARVE, a training-free reward-head editing method for scalar reward models. Instead of fine-tuning the reward model, HARVE identifies a multi-directional hacking subspace from residual stream directions associated with selected hacking subcategories, and removes the component of the reward-head vector aligned with that subspace. This directly reduces the reward head's sensitivity to hacking-related features using only a small set of contrastive gold-hacked examples, without gradient updates or fine-tuning. Comprehensive experiments across eight reward models indicates that \model improves hacking robustness, outperforms fine-tuning baselines, and preserves reward-models' general capability. Further analyses suggest that reward hacking is better captured as a multidimensional residual-space structure than by isolated surface cues.

Yidong Chai, Yanguang Liu, Xuan Tian et al.

We study investor learning and information acquisition in private markets using a large dataset of expert network calls. We develop a sequential Large Language Model (LLM)-Bayesian framework that treats expert interactions as sequential signals and recovers time-varying beliefs about firm success and associated uncertainty from unstructured conversations, providing a measurement system for how qualitative information is aggregated into investment expectations. We show that expert network calls contain decision-relevant information: a single call increases subsequent investment probability by 6.9 to 9.0 percentage points, while positive sentiment raises deal likelihood by 3.9 to 4.1 percentage points. Informativeness varies across topics and environments: discussions of technology adoption and customer acquisition increase deal probability by up to 14.7 percentage points, particularly in high-uncertainty settings. Information is asymmetric across horizons, with positive signals predicting short-term investment decisions and negative signals more informative about long-run firm performance. Consistent with a belief-based mechanism, investment decisions respond to inferred beliefs rather than raw signals. A one standard deviation increase in success belief raises deal probability by approximately 11 percentage points, while reductions in uncertainty further increase investment likelihood. Our framework improves capital allocation, increasing portfolio returns by 15.26% and F1 by 6.69%, with gains concentrated in the upper tail. Attention and ablation analyses show that conversational cues are particularly informative for technologically complex startups, young firms, diverse founding teams, and firms with low public visibility, where information frictions are severe.

Weiqi Liu, Yongliang Miao, Haiyan Zhao et al.

Neuron-level interpretation in large language models (LLMs) is fundamentally challenged by widespread polysemanticity, where individual neurons respond to multiple distinct semantic concepts. Existing single-pass interpretation methods struggle to faithfully capture such multi-concept behavior. In this work, we propose NeuronScope, a multi-agent framework that reformulates neuron interpretation as an iterative, activation-guided process. NeuronScope explicitly deconstructs neuron activations into atomic semantic components, clusters them into distinct semantic modes, and iteratively refines each explanation using neuron activation feedback. Experiments demonstrate that NeuronScope uncovers hidden polysemanticity and produces explanations with significantly higher activation correlation compared to single-pass baselines.

Dong Shu, Haoyang Yuan, Yuchen Wang et al.

Large vision-language models (LVLMs) have made significant progress in chart understanding. However, financial charts, characterized by complex temporal structures and domain-specific terminology, remain notably underexplored. We introduce FinChart-Bench, the first benchmark specifically focused on real-world financial charts. FinChart-Bench comprises 1,200 financial chart images collected from 2015 to 2024, each annotated with True/False (TF), Multiple Choice (MC), and Question Answering (QA) questions, totaling 7,016 questions. We conduct a comprehensive evaluation of 25 state-of-the-art LVLMs on FinChart-Bench. Our evaluation reveals critical insights: (1) the performance gap between open-source and closed-source models is narrowing, (2) performance degradation occurs in upgraded models within families, (3) many models struggle with instruction following, (4) both advanced models show significant limitations in spatial reasoning abilities, and (5) current LVLMs are not reliable enough to serve as automated evaluators. These findings highlight important limitations in current LVLM capabilities for financial chart understanding. The FinChart-Bench dataset is available at https://huggingface.co/datasets/Tizzzzy/FinChart-Bench.

Zirui He, Huopu Zhang, Yanguang Liu et al.

Financial prediction from long documents involves significant challenges, as actionable signals are often sparse and obscured by noise, and the optimal LLM for generating embeddings varies across tasks and time periods. In this paper, we propose FinAnchor(Financial Anchored Representations), a lightweight framework that integrates embeddings from multiple LLMs without fine-tuning the underlying models. FinAnchor addresses the incompatibility of feature spaces by selecting an anchor embedding space and learning linear mappings to align representations from other models into this anchor. These aligned features are then aggregated to form a unified representation for downstream prediction. Across multiple financial NLP tasks, FinAnchor consistently outperforms strong single-model baselines and standard ensemble methods, demonstrating the effectiveness of anchoring heterogeneous representations for robust financial prediction.

Huopu Zhang, Yanguang Liu, Miao Zhang et al.

Predicting earnings surprises from financial documents, such as earnings conference calls, regulatory filings, and financial news, has become increasingly important in financial economics. However, these financial documents present significant analytical challenges, typically containing over 5,000 words with substantial redundancy and industry-specific terminology that creates obstacles for language models. In this work, we propose the SAE-FiRE (Sparse Autoencoder for Financial Representation Enhancement) framework to address these limitations by extracting key information while eliminating redundancy. SAE-FiRE employs Sparse Autoencoders (SAEs) to decompose dense neural representations from large language models into interpretable sparse components, then applies statistical feature selection methods, including ANOVA F-tests and tree-based importance scoring, to identify the top-k most discriminative dimensions for classification. By systematically filtering out noise that might otherwise lead to overfitting, we enable more robust and generalizable predictions. Experimental results across three financial datasets demonstrate that SAE-FiRE significantly outperforms baseline approaches.