Keynote Speakers

Workshop Organizers

Accepted Papers

Posters:

1. Neural Routed Boosting: Robust Decision-Making under Heteroscedastic Noise

   • Puspak Chakraborty, Arun Rajkumar

   • Ensemble methods struggle with heteroscedastic label uncertainty, limiting their reliability in downstream decision-making tasks. Traditional algorithms like AdaBoost overfit to corrupted instances, while robust variants that rely on global conditional risk estimation often fail to capture spatially varying uncertainty. We propose Neural Routed Boosting, a novel framework for conditional reasoning and robust classification under localized uncertainty. Neural Routed Boosting utilizes a lightweight neural network to capture spatial variations in the data. This network partitions the input space into geometrically coherent regions. It then routes samples to local experts. This structural approach isolates high-uncertainty subspaces, preventing them from corrupting the decision boundaries of reliable regions. We prove that this composite architecture of a neural router and region-specific experts constitutes a valid weak learner, guaranteeing training error convergence. Evaluations on synthetic and real-world datasets demonstrate that our spatial approach to uncertainty quantification outperforms traditional and robust baselines in complex environments.

Oral Presentations:

1. From Topology To Trajectory: LLM-Driven World Models For Supply Chain Resilience

   • Jia Luo

   • Semiconductor supply chains face unprecedented resilience challenges amidst global geopolitical turbulence. Conventional Large Language Model (LLM) planners, when confronting such non-stationary “Policy Black Swan” events, frequently suffer from Decision Paralysis or a severe Grounding Gap due to the absence of physical environmental modeling. This paper introduces ReflectiChain, a cognitive agentic framework tailored for resilient macroeconomic supply chain planning. The core innovation lies in the integration of Latent Trajectory Rehearsal powered by a generative world model, which couples reflection-in-action (System 2 deliberation) with delayed reflection-on-action. Furthermore, we leverage a Retrospective Agentic RL mechanism to enable autonomous policy evolution during the deployment phase (test-time). Evaluations conducted on our high-fidelity benchmark, Semi-Sim, demonstrate that under extreme scenarios such as export bans and material shortages, Silicon Janus achieves a 250% improvement in average step rewards over the strongest LLM baselines. It successfully restores the Operability Ratio (OR) from a deficient 13.3% to over 88.5% while ensuring robust gradient convergence. Ablation studies further underscore that the synergy between physical grounding constraints and double-loop learning is fundamental to bridging the gap between semantic reasoning and physical reality for long-horizon strategic planning.

2. DBE-Net: A Dual-dimensional Band Encoder Network for Unsupervised Anomaly Detection in Self-Piercing Riveting

   • Chen Hu, Yujie Wan, Yang Luo, Xiaofeng Gao

   • Self-Piercing Riveting (SPR) is critical to ensuring the structural safety of New Energy Vehicle (NEV) body-in-white assembly. However, conventional anomaly detection approaches typically ignore inter-series process dependencies and fail to extract subtle defect features hidden in strong global trends. To address these issues, this paper proposes DBE-Net, an unsupervised anomaly detection framework tailored for SPR quality monitoring. It combines Induced Set Attention Block (ISAB) to capture process-level correlations across sequences, Variational Mode Decomposition (VMD) to separate multi-frequency signal components, and Deep SVDD for compact one-class normality modeling. Extensive experiments on the industrial SPRAD dataset demonstrate that DBE-Net outperforms state-of-the-art methods in AUC, precision, recall, and F1-score, accurately identifying micro-defects while suppressing high-frequency noise. This work provides a reliable, high-precision intelligent detection solution for real-world NEV manufacturing quality assurance.

3. WiSDoM: Frugal Winner Selection by Design of Matchups

   • Saranath P

   • Identifying the best of 𝑁 items from noisy pairwise comparisons is a core primitive of modern machine-learning pipelines — RLHF annotation, LLM-judge arenas, policy selection — where each comparison is expensive, and the available budget is a small multiple of 𝑁. Under the Bradley–Terry–Luce model, we recast winner identification as a winner-focused experimental design problem: pick a distribution over pairs that minimizes the worst-case posterior variance of the winner’s score gap against the top challengers. The design is convex but only useful with a reliable warm start, so we pair it with a best-of-𝑡 single-elimination bracket using Elo scoring to secure winner coverage. The resulting algorithm, WiSDoM, recovers the top item effectively under shoestring budgets and outperforms prior budgeted BTL winner-identification baselines across synthetic and real-world experiments

4. Less Can Be Safer: Tail-Risk Reduction via Capacity-Constrained Policies in Industrial Reinforcement Learning

   • Minu Baek, Gihun Gil, Yeojin Jang, Byounghoon Son, Beomdo Park, Minsung Jung, Junseong Park, Hyeonseok Jang, Sangkeum Lee 

   • Average reward can hide unacceptable rare failures in safety-critical industrial control. We present a workshop study of capacity-constrained reinforcement learning for papermaking control, using a data-validated digital twin built from 11 months of industrial operation and 687 production lots. The policy class uses shallow variational quantum circuit features as a deliberately small representation inside PPO; the workshop claim is not quantum computational advantage, but tail-risk reduction under process-derived feedback. Across 600 evaluation episodes on 120 lots and five seeds, the shallowest policy reduces failure rate to 0.33% compared with 6.0% for a classical PPO baseline, improves 5th-percentile safe-zone compliance from 13.6% to 55.5%, and reduces variance by 2.44x while mean performance is not significantly different. We introduce the Proactive Safety Index as a descriptive signal for anticipatory speed reduction. The evidence is limited to in-distribution digital-twin simulation. We submit this result as a case study in uncertainty-aware decision making where safety depends on rare-event behavior rather than mean performance.

5. Travel-Oriented Reasoning Large Language Model via Domain-Specific Knowledge Graphs

   • Vignesh Ram Nithin Kappagantula, Shayan Hassantabar, Samuel Simpson, Golnaz Moallem

   • Large language models (LLMs) demonstrate broad reasoning abilities but struggle with accuracy and reliability in specialized domains such as travel, where reasoning depends on precise definitions, rules, and expert-defined conceptual frameworks, and where confident but unfounded outputs arise from a reasoning failure in which the model has not internalized the underlying domain graph rather than from missing domain knowledge alone. We propose a modular pipeline for building a travel-domain reasoning LLM grounded in an expert-designed knowledge graph (KG). Our pipeline integrates a travel KG that encodes domain entities and their relationships, a bottom-up construction procedure that walks the KG to produce multi-hop question answer (QA) pairs, a supervised fine-tuning stage that embeds the domain knowledge into a reasoning-capable LLM using the generated QA pairs as auditable reasoning traces, and a travel-domain benchmark dataset that measures the fine-tuned model's accuracy and calibration. We evaluate our approach using Qwen3-4B with LoRA adaptation. Our reasoning model achieves an 82.4% exact match on the benchmark. This performance significantly outperforms the pretrained Qwen3-4B baseline at 22.4%. A calibration analysis decomposes the residual 17.57% of errors into two distinct failure modes: an over-confident multi-label decoder that predicts both correct answers plus one spurious option on most dual-answer mistakes, and a smaller reasoning failure on single-answer questions where the supporting facts are present in the KG but the model fails to reconstruct the correct multi-hop path. This split confirms that explicit KG-grounded reasoning substantially improves the accuracy and uncertainty interpretation of LLMs in specialized domains, and isolates per-option calibration and trace-length-aware decoding as the next axes of improvement.

6. Claim-Level Confidence Calibration for Reliable Decision Making with Large Language Models

   • Toghrul Abbasli

   • Large Language Models (LLMs) increasingly support decision-making in high-stakes domains, but they often hallucinate and express confidence that is misaligned with factual correctness. Response-level confidence is a coarse signal: a single generation can mix correct and incorrect statements, so a single number is not actionable for users that must accept, reject, or verify individual pieces of information. We study \emph{claim-level confidence calibration} as a decision-relevant uncertainty signal: each response is decomposed into atomic, verifiable claims, and each claim is assigned a calibrated confidence using inference-time signals from consistency across samples and self-verification. Our framework operates in closed-box settings (no logits, no fine-tuning) and applies post-hoc calibration directly at the claim level, enabling selective intervention such as evidence retrieval or human review for low-confidence claims. Across TriviaQA and TruthfulQA we evaluate seven baselines on six recent models (Llama-3.1, Mistral, Qwen2.5, DeepSeek-R1, GPT-4, GPT-4o), and show that claim-level decomposition combined with post-hoc calibration reduces expected calibration error on factual questions while exposing failure modes on adversarial false-premise questions where decision-makers most need reliable uncertainty estimates.

7. Optimization-based Online Conformal Prediction for Multi-step Forecasting

   • Ruipu Li, Daniel Menacho, Alexander Rodríguez

   • Conformal prediction (CP) is well-suited for uncertainty quantification in time series forecasting due to its distribution-free coverage guarantees. However, existing multi-step methods often struggle to balance coverage validity with efficiency: they either calibrate horizons independently—ignoring temporal correlations—or enforce strict simultaneous coverage, resulting in overly conservative intervals. In this work, we propose O2CP: Optimization-based Online Conformal Prediction, a unified framework for online conformal prediction that explicitly models multi-step error dependencies without sacrificing long-term marginal coverage guarantees. We first prove that standard online conformal updates maintain validity as long as calibration parameters remain within a defined "safe" region. Leveraging this theoretical insight, we introduce a two-layer architecture: an outer layer that defines admissible parameter sets to ensure validity, and an inner layer that performs constrained optimization to model joint error distributions and minimize horizon-wide objectives. To make this computationally feasible, we develop a lightweight sampling strategy that estimates joint distributions without requiring large calibration sets. Extensive experiments on real-world datasets—including autonomous driving, climate forecasting, and public health—demonstrate that O2CP consistently outperforms state-of-the-art baselines, achieving target coverage with significantly sharper prediction intervals and reduced regret over long horizons.

8. GLU: Global-Local Uncertainty Quantification in LLMs

   • Johanne Medina, Tianyi Zhou, Keivin Isufaj, Aristides Gionis, Sanjay Chawla

   • Large language models hallucinate confidently, making uncertainty quantification (UQ) essential for reliable deployment. Existing methods rely predominantly on token-level signals, leaving the geometric structure of intermediate hidden states underused. In this paper, we take the geometric complexity of hidden-state matrices as a measure of the global uncertainty of LLMs, while treating token-level uncertainty estimation as a local metric. We show that hidden-state geometric entropy (global uncertainty) and token-level entropy (local uncertainty) are statistically near-orthogonal, capturing distinct failure regimes for reliability prediction. In particular, global geometry recovers the confident-but-wrong failure mode that local signals systematically miss. Building on this, we propose Global-Local Uncertainty (GLU), an unsupervised, single-pass score that fuses the two signals via a multiplicative gate. Across three model families and six benchmarks, GLU matches or outperforms all unsupervised baselines while being a single forward pass, length-normalized, and architecture-agnostic.

9. NG-MMoE: Improving Multi-Task Mixture-of-Experts via Learnable Stochastic Gating

   • Connor Lee, Carissa Lee

   • Multi-gate Mixture-of-Experts (MMoE) has demonstrated strong performance in multi-task learning by learning task-specific gating over a shared pool of expert networks. However, the deterministic softmax gating in MMoE does not model uncertainty in the routing decision, causing it to collapse to suboptimal expert assignments, particularly when tasks are weakly correlated or training labels are sparse. We propose Noisy-Gate MMoE (NG-MMoE), which augments each per-task gate with input-conditioned learnable Gaussian noise during training. The noise acts as a principled uncertainty-aware exploration mechanism over the expert routing space, smoothing the loss landscape and preventing premature gate collapse. We provide theoretical motivation grounded in stochastic optimization, information theory, and the exploration-exploitation trade-off. Experiments on UCI Census-income and MovieLens 100k/1M show that NG-MMoE (i) achieves lower average loss and variance than MMoE and (ii) produces higher-entropy gate distributions, confirming better expert utilization under routing uncertainty.