arXiv subjects

Chengjun Mao

Publications and source records attributed to Chengjun Mao.

8 recordsLinked to original sources

arXiv2026

When Top-K Misses the Decision: Tool-Call Drift in Multi-Teacher On-Policy Distillation

Top-$K$ teacher logits make on-policy distillation tractable, but retained teacher mass does not certify student-relative gradient fidelity. We study routed, two-teacher tool-use distillation. In a frozen Qwen3.5-9B audit, the tool teacher ranks the entry token first on 500 tool prompts, and top-32 reinforces it in every matched pair. The response teacher's top-32 instead retains 99.99\% mass yet contains the same token on only 0.4\% of 500 response prompts; student-aware support restores the coordinate in all matched pairs and nearly matches full-vocabulary descent. On the tool route, top-32 preserves the reinforcing direction but substantially attenuates its full-vocabulary magnitude despite mass displayed as 1.000000. Matched restoration connects the response-side omission to behavior: restoring the entry coordinate at every supervised response position lowers full-generation over-calling from $14.2 \pm 2.1\%$ to $3.7 \pm 0.5\%$ across three seeds, but also lowers call recall by 12.4 points. A non-tool placebo changes over-calling by only 0.95 points. A synthetic output-logit-gradient-matched reroute also fails to reproduce exact restoration. We compare three optimization layers: student-aware support changes distilled coordinates, loss shaping changes retained-signal strength, and validation-tuned entry bias shifts inference without retraining. They expose distinct trade-offs in correction scope, system access, and required-call retention. Llama-3.1-8B reproduces the directional support asymmetry under its native JSON protocol. The results causally implicate decision-critical support omission as one contributor in the primary Qwen setting and show why both support fidelity and deployment costs must be audited. Code and aggregate artifacts are available at https://github.com/shen-jiabin/decision-support-opd.

cs.CL↗

arXiv2026

Answer Presence Drives RAG Rewriting Gains

Retrieval-augmented QA pipelines often route retrieved passages through an LLM \emph{rewriter} before a smaller reader, lifting F1 by tens of points on multi-hop benchmarks; this gain is typically credited to improved evidence quality. We ask whether that lift is causally driven by the gold answer string appearing in the rewritten context rather than by curation per se, using a controlled intervention audit. For each rewritten context we re-run the reader after one of four controlled edits to the compile output: removing the gold answer span, replacing a length-matched random non-answer span (placebo), or injecting the gold into rewrites where it was absent (at the prefix or at a midpoint sentence boundary). Across twelve completed (cell, baseline) intervention runs spanning three reader families (Qwen2.5-7B, Qwen3.5-35B, GLM-4.7), two datasets (HotpotQA, 2WikiMultihopQA), and three compiler arrangements (MA-only, MB-only, MA$+$verify), removing the gold answer drops reader F1 by $28$ to $64$ points beyond the length-matched placebo on paired \texttt{answer-in-compile} strata, and prepending the gold into rewrites that lacked it raises F1 by $+0.7$ to $+9.7$ points in $10$ of $12$ (cell, baseline) combinations. A companion five-sentinel audit shows the conventional single-\texttt{[MASK]} probe is itself sentinel-fragile: on 2Wiki it reports a $+4.12$~F1 ``non-leakage residual'' that flips to $-3.33$ to $-7.81$~F1 under four alternative sentinels and fails an equivalence test for three of those four ($1/4$~pass). We do not propose a new rewriter or mitigation; we release the intervention runner and the sentinel panel so that other rewriter-gain claims can be tested against the same standard.

cs.AI↗

arXiv2026

Attention-guided Evidence Grounding for Spoken Question Answering

Spoken Question Answering (Spoken QA) presents a challenging cross-modal problem: effectively aligning acoustic queries with textual knowledge while avoiding the latency and error propagation inherent in cascaded ASR-based systems. In this paper, we introduce Attention-guided Evidence Grounding (AEG), a novel end-to-end framework that leverages the internal cross-modal attention of Speech Large Language Models (SpeechLLMs) to explicitly locate and ground key evidence in the model's latent space. To address the diffuse attention distribution in pre-trained models, we propose Learning to Focus on Evidence (LFE), a supervised fine-tuning paradigm that calibrates the model's attention mechanism to distinguish query-relevant segments from irrelevant context. Experiments on SQuAD, HotpotQA, and MuSiQue demonstrate that AEG reduces hallucinations and achieves strong efficiency gains, outperforming large-scale cascaded baselines (Whisper-Large-v3 + Reranker) while reducing inference latency by approximately 62%.

cs.CL↗

arXiv2026

Deep GraphRAG: A Balanced Approach to Hierarchical Retrieval and Adaptive Integration

Graph-based Retrieval-Augmented Generation (GraphRAG) frameworks face a trade-off between the comprehensiveness of global search and the efficiency of local search. Existing methods are often challenged by navigating large-scale hierarchical graphs, optimizing retrieval paths, and balancing exploration-exploitation dynamics, frequently lacking robust multi-stage re-ranking. To overcome these deficits, we propose Deep GraphRAG, a framework designed for a balanced approach to hierarchical retrieval and adaptive integration. It introduces a hierarchical global-to-local retrieval strategy that integrates macroscopic inter-community and microscopic intra-community contextual relations. This strategy employs a three-stage process: (1) inter-community filtering, which prunes the search space using local context; (2) community-level refinement, which prioritizes relevant subgraphs via entity-interaction analysis; and (3) entity-level fine-grained search within target communities. A beam search-optimized dynamic re-ranking module guides this process, continuously filtering candidates to balance efficiency and global comprehensiveness. Deep GraphRAG also features a Knowledge Integration Module leveraging a compact LLM, trained with Dynamic Weighting Reward GRPO (DW-GRPO). This novel reinforcement learning approach dynamically adjusts reward weights to balance three key objectives: relevance, faithfulness, and conciseness. This training enables compact models (1.5B) to approach the performance of large models (70B) in the integration task. Evaluations on Natural Questions and HotpotQA demonstrate that Deep GraphRAG significantly outperforms baseline graph retrieval methods in both accuracy and efficiency.

cs.IR↗

arXiv2025

Think in Blocks: Adaptive Reasoning from Direct Response to Deep Reasoning

Large Language Models (LLMs) with chains-of-thought have demonstrated strong performance on an increasing range of tasks, particularly those involving complex logical reasoning. However, excessively long chains can lead to overthinking, causing computational waste and slower responses. This raises a question: can LLMs dynamically adjust the length of their reasoning processes based on task complexity? To address this, we propose the Think in Blocks framework, which enables adaptive reasoning-from zero to deep reasoning-by partitioning the reasoning process into a tunable number of blocks. Our main contributions are: (1) Establishing an explicit block-structured paradigm in which the model first predicts an integer reasoning budget-the number of blocks-and then partitions its reasoning accordingly; (2) Training an adaptive model through a three-stage pipeline-Supervised Fine-Tuning, reward-guided Direct Preference Optimization, and Reinforcement Learning-that adjusts its reasoning depth to problem difficulty; (3) Exploiting the explicit block count to dynamically control reasoning depth at inference time, allowing flexible adjustment of chain-of-thought length during deployment.

cs.AI↗

arXiv2023

FAN: Fatigue-Aware Network for Click-Through Rate Prediction in E-commerce Recommendation

Since clicks usually contain heavy noise, increasing research efforts have been devoted to modeling implicit negative user behaviors (i.e., non-clicks). However, they either rely on explicit negative user behaviors (e.g., dislikes) or simply treat non-clicks as negative feedback, failing to learn negative user interests comprehensively. In such situations, users may experience fatigue because of seeing too many similar recommendations. In this paper, we propose Fatigue-Aware Network (FAN), a novel CTR model that directly perceives user fatigue from non-clicks. Specifically, we first apply Fourier Transformation to the time series generated from non-clicks, obtaining its frequency spectrum which contains comprehensive information about user fatigue. Then the frequency spectrum is modulated by category information of the target item to model the bias that both the upper bound of fatigue and users' patience is different for different categories. Moreover, a gating network is adopted to model the confidence of user fatigue and an auxiliary task is designed to guide the learning of user fatigue, so we can obtain a well-learned fatigue representation and combine it with user interests for the final CTR prediction. Experimental results on real-world datasets validate the superiority of FAN and online A/B tests also show FAN outperforms representative CTR models significantly.

cs.IR↗

arXiv2021

MOEF: Modeling Occasion Evolution in Frequency Domain for Promotion-Aware Click-Through Rate Prediction

Promotions are becoming more important and prevalent in e-commerce to attract customers and boost sales, leading to frequent changes of occasions, which drives users to behave differently. In such situations, most existing Click-Through Rate (CTR) models can't generalize well to online serving due to distribution uncertainty of the upcoming occasion. In this paper, we propose a novel CTR model named MOEF for recommendations under frequent changes of occasions. Firstly, we design a time series that consists of occasion signals generated from the online business scenario. Since occasion signals are more discriminative in the frequency domain, we apply Fourier Transformation to sliding time windows upon the time series, obtaining a sequence of frequency spectrum which is then processed by Occasion Evolution Layer (OEL). In this way, a high-order occasion representation can be learned to handle the online distribution uncertainty. Moreover, we adopt multiple experts to learn feature representations from multiple aspects, which are guided by the occasion representation via an attention mechanism. Accordingly, a mixture of feature representations is obtained adaptively for different occasions to predict the final CTR. Experimental results on real-world datasets validate the superiority of MOEF and online A/B tests also show MOEF outperforms representative CTR models significantly.

cs.LG↗

arXiv2021

MetaCVR: Conversion Rate Prediction via Meta Learning in Small-Scale Recommendation Scenarios

Different from large-scale platforms such as Taobao and Amazon, CVR modeling in small-scale recommendation scenarios is more challenging due to the severe Data Distribution Fluctuation (DDF) issue. DDF prevents existing CVR models from being effective since 1) several months of data are needed to train CVR models sufficiently in small scenarios, leading to considerable distribution discrepancy between training and online serving; and 2) e-commerce promotions have significant impacts on small scenarios, leading to distribution uncertainty of the upcoming time period. In this work, we propose a novel CVR method named MetaCVR from a perspective of meta learning to address the DDF issue. Firstly, a base CVR model which consists of a Feature Representation Network (FRN) and output layers is designed and trained sufficiently with samples across months. Then we treat time periods with different data distributions as different occasions and obtain positive and negative prototypes for each occasion using the corresponding samples and the pre-trained FRN. Subsequently, a Distance Metric Network (DMN) is devised to calculate the distance metrics between each sample and all prototypes to facilitate mitigating the distribution uncertainty. At last, we develop an Ensemble Prediction Network (EPN) which incorporates the output of FRN and DMN to make the final CVR prediction. In this stage, we freeze the FRN and train the DMN and EPN with samples from recent time period, therefore effectively easing the distribution discrepancy. To the best of our knowledge, this is the first study of CVR prediction targeting the DDF issue in small-scale recommendation scenarios. Experimental results on real-world datasets validate the superiority of our MetaCVR and online A/B test also shows our model achieves impressive gains of 11.92% on PCVR and 8.64% on GMV.

cs.LG↗