Bo Zou

Bo Zou, Chao Xu

High-quality expert chain-of-thought (CoT) data is one of the core bottlenecks in large language model (LLM) post-training. Existing data production methods each have structural limitations: crowdsourced annotation lacks deep reasoning paths; expert solo writing is constrained by the "expert blind spot" -- experts structurally skip reasoning steps they consider obvious; RLHF only produces preference signals rather than reasoning chains. This paper proposes the BC Protocol -- a structured dual-expert elicitation method for LLM post-training data production. The method carefully pairs a domain expert (crystallized intelligence) with a knowledge engineer (fluid intelligence), systematically externalizing the expert's implicit judgments as natural language reasoning chains. We introduce the Participant Aptitude Model, which defines six participant characteristic dimensions that affect elicitation quality. "Calibrated Ignorance" is an original concept proposed in this paper. We further propose "Selection-over-Prescription" as a methodological principle: for implicit knowledge elicitation tasks, investing quality-control resources in personnel selection yields a higher return than investing the same resources in process design. In a controlled experiment in the narrative fiction domain, we directly compared CoT produced by BC Protocol dual dialogue (Group A, (n=20)) against CoT written independently by the same domain expert (Group B, (n=20)). Three cross-vendor judge models -- GPT-4o, Claude Opus 4.5, and Gemini 2.5 Pro -- conducted blind evaluation across five dimensions (600 ratings total). Results show that the BC Protocol achieves an overwhelming advantage in "naturalness of reasoning process" (Group A mean 4.80 vs. Group B mean 1.30, (p=2.4\times10^{-8}), Cliff's (δ=1.0)).

Bo Zou, Chao Xu

Large language models (LLMs) face a dual challenge in creative capability evaluation: existing benchmarks (e.g., Story Cloze Test, HellaSwag) measure models' discriminative ability over narrative continuation using multiple-choice recognition paradigms, rather than directly measuring creative generation capability; rubric-based scoring and LLM-as-Judge methods rely on subjective dimension assessment or natural language model outputs, and cannot provide objective, automated scoring mechanisms. This paper proposes QUIET (Quality Understanding via Interlocked Evaluation Testing), a diagnostic benchmark for LLM creative capability based on multi-blank cascaded story cloze. QUIET sets N blanks (10-20) in a story with complete structure, with each blank accompanied by an explicit content constraint, and cascade dependency relationships between blanks -- the content filled into earlier blanks constrains the feasible solution space for later blanks. The evaluated model (or human participants) fills all blanks in open-ended generation mode; the results are scored by an information-theoretic automated scoring protocol without human grading. The scoring protocol directly operationalizes the "calibrated surprise" theoretical framework (Zou & Xu, 2026a). For each blank k, a composite score is computed: score = satisfy * (1 + lambda * surprise), where lambda = 1.0. Here, "satisfy" measures how well the blank filling satisfies the content constraint (objective logical reasoning judgment, not subjective aesthetic scoring), and "surprise" measures the degree of surprise given that the constraint is satisfied. Creative answers that do not satisfy the constraint score zero; answers that satisfy the constraint but are mediocre score low; answers that satisfy the constraint and are surprising score high.

Bo Zou, Chao Xu

This paper provides an empirical implementation of the creative quality metric proposed in Calibrated Surprise (Zou & Xu, 2026a). The question this paper addresses is: does this mathematical claim hold at the engineering level? To make the answer as general as possible, we deliberately choose the strictest engineering conditions: low data cost and a small base model. Training data comes from approximately 100 expert chain-of-thought (CoT) annotations produced by the BC Protocol (Zou & Xu, 2026b). We also identify a data bias: most publicly available alignment datasets are skewed toward craft-related knowledge, while audience modeling and reality-logic coverage are systematically weak. We use the term Creative Quality Alignment (CQA) to describe this class of engineering methods. We also offer a supporting theoretical observation: in an LLM with a single conditional distribution architecture, calibrating the appreciation side automatically transfers to the generation side via architectural duality. This is the structural reason why ~100 CoT examples are sufficient -- not a purely empirical observation like LIMA (Zhou et al., 2023).

Jiachen Ma, Jiawen Zhang, Xiangtian Li et al.

While Large Language Models (LLMs) demonstrate remarkable capabilities, they remain susceptible to sophisticated, multi-step jailbreak attacks that circumvent conventional surface-level safety alignment by exploiting the internal generation process. To address these vulnerabilities, we propose Reflector, a principled two-stage framework that internalizes self-reflection within the generation trajectory. Reflector first leverages teacher-guided generation to produce high-quality reflection data for supervised fine-tuning (SFT), establishing structured reflection patterns. It subsequently uses Reinforcement Learning (RL) with outcome-driven and reward-validity supervision to instill robust, autonomous self-reflection capabilities. Empirical results show that Reflector achieves Defense Success Rates (DSR) exceeding 90% against complex indirect attacks while generalizing robustly across diverse threat scenarios. Notably, the framework enhances both task-specific and general utility, yielding a 5.85% gain on GSM8K alongside improved performance on knowledge-intensive benchmarks. By internalizing trajectory-level safety, Reflector overcomes the fundamental limitations of surface alignment without significant computational overhead, offering an efficient and scalable solution for the development of safe and capable LLMs.

Bo Zou, Shaofeng Wang, Hao Liu et al.

Teeth localization, segmentation, and labeling in 2D images have great potential in modern dentistry to enhance dental diagnostics, treatment planning, and population-based studies on oral health. However, general instance segmentation frameworks are incompetent due to 1) the subtle differences between some teeth' shapes (e.g., maxillary first premolar and second premolar), 2) the teeth's position and shape variation across subjects, and 3) the presence of abnormalities in the dentition (e.g., caries and edentulism). To address these problems, we propose a ViT-based framework named TeethSEG, which consists of stacked Multi-Scale Aggregation (MSA) blocks and an Anthropic Prior Knowledge (APK) layer. Specifically, to compose the two modules, we design 1) a unique permutation-based upscaler to ensure high efficiency while establishing clear segmentation boundaries with 2) multi-head self/cross-gating layers to emphasize particular semantics meanwhile maintaining the divergence between token embeddings. Besides, we collect 3) the first open-sourced intraoral image dataset IO150K, which comprises over 150k intraoral photos, and all photos are annotated by orthodontists using a human-machine hybrid algorithm. Experiments on IO150K demonstrate that our TeethSEG outperforms the state-of-the-art segmentation models on dental image segmentation.

Bo Zou, Chao Xu

The essence of good creative writing is calibrated surprise: when constraints from all relevant dimensions act together, the feasible solution space collapses into a narrow region, and the surviving choices look least predictable from an unconstrained view. "Calibrated" has a precise meaning: the author's intent, the reader's reasonable expectation, and the logic of reality converge. When these three independent judgements agree on every dimension, the set of admissible writing choices is forced into a very small region. A mathematical corollary follows: full-dimensional accuracy and mediocrity are mutually exclusive -- two sides of one constraint structure, not separate goals. We use Shannon's mutual information $I(X;Y) = H(X) - H(X|Y)$ as our analysis tool. "Calibrated" corresponds to conditional entropy going to zero; "surprise" to entropy going up; mutual information is the precise measure of the joint quantity. The argument rests on two pillars. Static: when constraints from ethos, mythos, lexis, and dianoia are imposed together, the admissible set collapses sharply, and surviving solutions show up as low-probability choices from an unconstrained view. Dynamic: the chain rule shows each writing choice is constrained by what came before and constrains what comes after; macro-level decisions naturally contribute a larger share of information, removing the need for hand-tuned weighting. Through case studies and lightweight LLM-logprob computations, we show the framework is both analytically useful and operational, laying the theoretical groundwork for Creative Quality Alignment (CQA) and a professional evaluation benchmark.

Bo Zou, Chao Yang, Yu Qiao et al.

Existing methods to fine-tune LLMs, like Adapter, Prefix-tuning, and LoRA, which introduce extra modules or additional input sequences to inject new skills or knowledge, may compromise the innate abilities of LLMs. In this paper, we propose LLaMA-Excitor, a lightweight method that stimulates the LLMs' potential to better follow instructions by gradually paying more attention to worthwhile information. Specifically, the LLaMA-Excitor does not directly change the intermediate hidden state during the self-attention calculation of the transformer structure. We designed the Excitor block as a bypass module for the similarity score computation in LLMs' self-attention to reconstruct keys and change the importance of values by learnable prompts. LLaMA-Excitor ensures a self-adaptive allocation of additional attention to input instructions, thus effectively preserving LLMs' pre-trained knowledge when fine-tuning LLMs on low-quality instruction-following datasets. Furthermore, we unify the modeling of multi-modal tuning and language-only tuning, extending LLaMA-Excitor to a powerful visual instruction follower without the need for complex multi-modal alignment. Our proposed approach is evaluated in language-only and multi-modal tuning experimental scenarios. Notably, LLaMA-Excitor is the only method that maintains basic capabilities while achieving a significant improvement (+6%) on the MMLU benchmark. In the visual instruction tuning, we achieve a new state-of-the-art image captioning performance of 157.5 CIDEr on MSCOCO, and a comparable performance (88.39%) on ScienceQA to cutting-edge models with more parameters and extensive vision-language pertaining.

Bo Zou, Chao Yang, Yu Qiao et al.

Significant advancements in video question answering (VideoQA) have been made thanks to thriving large image-language pretraining frameworks. Although these image-language models can efficiently represent both video and language branches, they typically employ a goal-free vision perception process and do not interact vision with language well during the answer generation, thus omitting crucial visual cues. In this paper, we are inspired by the human recognition and learning pattern and propose VideoDistill, a framework with language-aware (i.e., goal-driven) behavior in both vision perception and answer generation process. VideoDistill generates answers only from question-related visual embeddings and follows a thinking-observing-answering approach that closely resembles human behavior, distinguishing it from previous research. Specifically, we develop a language-aware gating mechanism to replace the standard cross-attention, avoiding language's direct fusion into visual representations. We incorporate this mechanism into two key components of the entire framework. The first component is a differentiable sparse sampling module, which selects frames containing the necessary dynamics and semantics relevant to the questions. The second component is a vision refinement module that merges existing spatial-temporal attention layers to ensure the extraction of multi-grained visual semantics associated with the questions. We conduct experimental evaluations on various challenging video question-answering benchmarks, and VideoDistill achieves state-of-the-art performance in both general and long-form VideoQA datasets. In Addition, we verify that VideoDistill can effectively alleviate the utilization of language shortcut solutions in the EgoTaskQA dataset.

Limon Barua, Bo Zou, Yan et al.

Despite the rapid growth of online shopping and research interest in the relationship between online and in-store shopping, national-level modeling and investigation of the demand for online shopping with a prediction focus remain limited in the literature. This paper differs from prior work and leverages two recent releases of the U.S. National Household Travel Survey (NHTS) data for 2009 and 2017 to develop machine learning (ML) models, specifically gradient boosting machine (GBM), for predicting household-level online shopping purchases. The NHTS data allow for not only conducting nationwide investigation but also at the level of households, which is more appropriate than at the individual level given the connected consumption and shopping needs of members in a household. We follow a systematic procedure for model development including employing Recursive Feature Elimination algorithm to select input variables (features) in order to reduce the risk of model overfitting and increase model explainability. Extensive post-modeling investigation is conducted in a comparative manner between 2009 and 2017, including quantifying the importance of each input variable in predicting online shopping demand, and characterizing value-dependent relationships between demand and the input variables. In doing so, two latest advances in machine learning techniques, namely Shapley value-based feature importance and Accumulated Local Effects plots, are adopted to overcome inherent drawbacks of the popular techniques in current ML modeling. The modeling and investigation are performed both at the national level and for three of the largest cities (New York, Los Angeles, and Houston). The models developed and insights gained can be used for online shopping-related freight demand generation and may also be considered for evaluating the potential impact of relevant policies on online shopping demand.

Tanvir Ahamed, Bo Zou, Nahid Parvez Farazi et al.

This paper investigates the problem of assigning shipping requests to ad hoc couriers in the context of crowdsourced urban delivery. The shipping requests are spatially distributed each with a limited time window between the earliest time for pickup and latest time for delivery. The ad hoc couriers, termed crowdsourcees, also have limited time availability and carrying capacity. We propose a new deep reinforcement learning (DRL)-based approach to tackling this assignment problem. A deep Q network (DQN) algorithm is trained which entails two salient features of experience replay and target network that enhance the efficiency, convergence, and stability of DRL training. More importantly, this paper makes three methodological contributions: 1) presenting a comprehensive and novel characterization of crowdshipping system states that encompasses spatial-temporal and capacity information of crowdsourcees and requests; 2) embedding heuristics that leverage the information offered by the state representation and are based on intuitive reasoning to guide specific actions to take, to preserve tractability and enhance efficiency of training; and 3) integrating rule-interposing to prevent repeated visiting of the same routes and node sequences during routing improvement, thereby further enhancing the training efficiency by accelerating learning. The effectiveness of the proposed approach is demonstrated through extensive numerical analysis. The results show the benefits brought by the heuristics-guided action choice and rule-interposing in DRL training, and the superiority of the proposed approach over existing heuristics in both solution quality, time, and scalability. Besides the potential to improve the efficiency of crowdshipping operation planning, the proposed approach also provides a new avenue and generic framework for other problems in the vehicle routing context.

Nahid Parvez Farazi, Tanvir Ahamed, Limon Barua et al.

Deep reinforcement learning (DRL) is an emerging methodology that is transforming the way many complicated transportation decision-making problems are tackled. Researchers have been increasingly turning to this powerful learning-based methodology to solve challenging problems across transportation fields. While many promising applications have been reported in the literature, there remains a lack of comprehensive synthesis of the many DRL algorithms and their uses and adaptations. The objective of this paper is to fill this gap by conducting a comprehensive, synthesized review of DRL applications in transportation. We start by offering an overview of the DRL mathematical background, popular and promising DRL algorithms, and some highly effective DRL extensions. Building on this overview, a systematic investigation of about 150 DRL studies that have appeared in the transportation literature, divided into seven different categories, is performed. Building on this review, we continue to examine the applicability, strengths, shortcomings, and common and application-specific issues of DRL techniques with regard to their applications in transportation. In the end, we recommend directions for future research and present available resources for actually implementing DRL.