袁骁斐 Glofy

Travel-planning agents generate itineraries that may become infeasible after acceptance because of flight cancellations, hotel unavailability, or attraction closures. Revising these itineraries involves full replanning, classical plan repair, and LLM-based travel-agent revision, whose differing task formulations and evaluation protocols hinder comparison. We conduct a systematic empirical study using two TREK-derived benchmark sets: 500 single-disruption cases, including feasible and infeasible instances, and 200 feasible simultaneous compound-disruption cases. We compare LLM–Z3 full replanning, IPyHOPPER hierarchical repair, and an iTIMO local-revision adapter across effectiveness, plan stability, and computational cost. LLM–Z3 with Gemini achieved the highest observed compound-disruption success. IPyHOPPER nearly matched that configuration’s single-disruption overall success, while preserving substantially more of the accepted itinerary on successful repairs. Successful hierarchical and local repairs made fewer edits and retained more accepted commitments than full replanning. Computational profiles differed: IPyHOPPER used no LLM inference, the evaluated LLM–Z3 adapter used compact one-call inference, and the iTIMO adapter consumed substantially more tokens. The study provides practical guidelines for balancing feasibility recovery, commitment preservation, and computational cost within evaluated settings.

The Foreign Exchange (Forex) market is a decentralised, high-frequency trading environment where fraudulent behaviours can be concealed within high-volume and irregular temporal transaction streams. Detecting such activities is challenging due to complex behavioural dependencies, severe class imbalance, and weak supervision in real-world fraud labels. Existing fraud detection approaches often assume that fraud labels are reliable. However, these labels are typically generated through rule-based screening followed by expert verification, which inevitably introduces uncertainty and potential noise into the supervision signals. To tackle this challenge, we propose a label-quality-aware fraud detection framework. The framework integrates out-of-fold Random Forest prediction with Confident Learning to estimate sample-level label quality and identify potentially unreliable training labels. It further applies noise-handling strategies before passing the data to the ensemble fraud detection model that combines probabilistic screening with soft-voting integration. Extensive experiments on a real-world Forex transaction dataset evaluate model robustness under different levels of controlled label noise. Empirical evaluation demonstrates that increasing label corruption degrades fraud detection performance, while downweighting provides more stable performance recovery than removal and relabelling. Applying label-quality-aware downweighting directly to the original labels also improves detection performance, demonstrating the practical value of robustness-aware learning for scalable Forex fraud detection under noisy supervision.

Reinforcement learning has become an important paradigm for improving large language model reasoning, yet policy optimization methods such as GRPO still face two core limitations: reward signals are provided only at the sequence level, and gradient signals vanish when all trajectories within a group are uniformly correct or uniformly incorrect. To address these issues, we propose ODRPO (On-policy Dense Reward Policy Optimization), an on-policy reinforcement learning method that redistributes sequence-level credit using token-level signals derived from same-model privileged conditioning. ODRPO contains three components: (1) Smooth GRPO, which prevents degenerate groups from collapsing to zero advantage; (2) a privileged teacher view built from the same underlying model conditioned on the problem and reference solution; and (3) a JSD-based multiplicative reweighting mechanism that redistributes sequence-level advantages across token positions without introducing an external teacher or additional trainable model parameters. Experiments on multiple mathematical reasoning benchmarks show that, under the current setting, ODRPO achieves the strongest overall accuracy–length tradeoff among the compared methods, improving both Avg@1 and Avg@8 while showing the smallest reduction in response length. In addition, token-level analysis suggests that the resulting JSD weights are semantically selective, assigning larger weights to reasoning- and method-related tokens than to low-importance categories such as numerals, formatting tokens, and filler words.

Knowledge base question answering (KBQA) commonly uses semantic parsing to translate natural-language questions into executable programs over a knowledge base (KB). Existing systems typically select the single highest-scoring program, even when multiple programs return the same answer, and use execution only to verify program validity. Consequently, evidence for an answer is fragmented across programs, while mismatches between a question and the returned answer are not reflected in program scores. We propose ECHO, an execution-aware answer-level decoding framework. ECHO predicts a type skeleton to guide grounded program enumeration, executes all enumerated candidates, and groups programs by their returned answers. It scores each program using a noisy-channel model that combines an answer-aware channel term, measuring how well the program and its execution result explain the input question, with a skeleton-conditioned prior over the program. A tempered sum then aggregates the scores of programs supporting each answer. ECHO achieves state-of-the-art performance on three widely used benchmarks, GrailQA, WebQSP, and GraphQuestions, across all reported EM and F1 metrics, while answering each question an order of magnitude faster than systems with comparable accuracy. Controlled comparisons further confirm the effectiveness of our proposed answer-level decoding.

Large language models (LLMs) are increasingly deployed for resume screening, yet it remains unclear whether they respond fairly to gender-related cues embedded in resumes. To address this question, we introduce a controlled benchmark for auditing gender bias in LLM-based job-resume matching. The benchmark integrates publicly available job descriptions with de-identified real-world resumes and constructs counterfactual gender variants that differ solely in gender-related information, including names, pronouns, and explicit gender labels. In addition to conventional female and male variants, the benchmark incorporates a non-binary variant and a neopronoun-associated variant, denoted as Neutral and Neo, respectively. Using a unified evaluation protocol, we assess six widely used LLM families: ChatGPT, Claude, DeepSeek, Gemini, MiMo, and Qwen. We also conduct multi-factor experiments across job levels, industries, and progressively explicit gender-cue disclosure conditions. Our analysis reveals that the average scores do not exhibit a linear trend with increasing disclosure intensity, but instead vary across contexts. Specifically, Male and Female candidates achieve the highest scores when no gender cues are disclosed, whereas Neutral and Neo candidates obtain their highest scores under full disclosure. These findings suggest that fairness audits should attend not only to explicit demographic indicators but also to indirect signals, along with their interactions with occupational and evaluative contexts.

Large language models (LLMs) are increasingly used for resume screening, but their sensitivity to gender-related cues remains unclear. We present a controlled benchmark for auditing gender bias in LLM-based job-resume matching, combining public job descriptions, de-identified real resumes, and counterfactual variants that vary only identity signals such as names, pronouns, and explicit gender attributes. Beyond male and female conditions, our benchmark includes non-binary and neopronoun-based identity presentations. We evaluate six LLM families: ChatGPT, Gemini, Claude, DeepSeek, MiMo, and Qwen, using a unified evidence-based scoring rubric. Across overall, seniority-level, and industry-level analyses, we find that model family, job level, and industry explain larger score variation than identity condition, while gender cues still cause small but recurring shifts, especially for neopronoun-based profiles. Our results show that fairness risks in LLM-assisted hiring are context-dependent and require aggregate, stratified, and matched auditing.