Ibne Farabi Shihab

Ibne Farabi Shihab, Fariya Afrin, Sanjeda Akter et al.

Process reward models (PRMs) are widely used in language-model training with dense step-level supervision. They assume PRM scores are stable proxies for step correctness under label-preserving transformations. These transformations change reasoning structure but preserve final answers. We argue this assumption is not well validated. Such transformations can change how PRM scores relate to correctness signals, leading to different failure modes across models.To address this gap, we introduce \textbf{EST-PRM}, a stress-testing framework for dense process rewards. It applies three transformations: (1) step inflation, (2) dependency-aware step reordering, and (3) confidence markers. A vulnerability decomposition is defined that separates reward inflation from loss of correctness sensitivity. Five PRM-style models are evaluated on 4,687 reasoning chains from MATH-500, GSM8K, and PRMBench.The results indicate clear differences in vulnerability patterns across models. Math-Shepherd shows the strongest sensitivity to position perturbations, with a Pearson correlation drop of $0.152 \pm 0.038$ and a $32.8 \pm 4.9\%$ score inflation rate. Qwen2.5-Math-PRM is most affected by step inflation, reaching a $47.6 \pm 4.3\%$ inflation rate. Confidence-based perturbations also distort reward calibration, revealing inconsistencies in correctness estimation. Three mitigation strategies are evaluated, highlighting trade-offs between robustness coverage and false-positive rates.

Ibne Farabi Shihab, Fariya Afrin, Anuj Sharma

Continual instruction tuning updates a language model through a sequence of new domains, yet each update can progressively erode previously learned capabilities and alignment behavior. Replay is the standard mitigation, but fixed replay ratios are inherently limited because the optimal mixture varies with the current domain, the training stage, and the evolving vulnerability of prior behaviors. We propose PROX-YMIX, a framework that learns a dynamic replay controller on a small proxy model and transfers the frozen controller to a larger target. The controller never observes future tasks and constructs its state from normalized validation losses and their temporal dynamics, producing a masked mixture over the current task and accessible replay buffers. Our core empirical hypothesis is forgetting mirroring: task vulnerability rankings remain largely consistent across model scales even when absolute loss magnitudes differ. We validate this assumption empirically before transferring controllers across scales. On LLaMA-3-8B across five continual instruction tuning sequences, PROXYMIX improves average accuracy by 3.4 points, reduces final forgetting by 3.5 points, and raises safety score by 5.8 points over the strongest non-oracle baseline, at roughly 50x lower policy learning cost than Oracle Target RL. The framework is leakage free and architecture independent at the interface level, and we also identify settings where the proxy assumption breaks down, highlighting limitations for robust deployment.

Ibne Farabi Shihab, Abu Sa-Adat Mohamed Moon-Im Al Ahsan, Anuj Sharma

Federated continual learning (FCL) lets distributed clients adapt language-model heads to evolving NLP tasks without sharing raw text. Under user-level differential privacy (DP), replay-based continual learning faces a structural obstacle: clients can release only small noisy lists of candidate replay summaries, and those lists are unordered across clients. We introduce Canonicalized Stable-List Replay (CSLR), where clients privately produce candidate replay distributions over a shared sentence-embedding space and the server aligns them using signatures induced by public anchor sentences. The anchors provide identifiability for aggregation rather than additional replay data. We prove that, under an observable anchor-signature margin, $O(\log(N/η)/p)$ anchors distinguish $N$ candidate list elements with probability at least $1-η$, and we give a scoped anchorless non-identifiability result for unordered-label oracle models. Across five seeds on continual classification, NER, and dialogue benchmarks, CSLR improves the final average task metric by 3.9--5.6 points over the strongest non-CSLR DP baseline at $\eps=4$ under the reported replay-release budget, while also outperforming Hungarian and optimal-transport matchers. The formal privacy guarantee covers replay release; end-to-end private training additionally requires composition with a private optimizer for task-head updates.

Ibne Farabi Shihab, Fariya Afrin, Sanjeda Akter et al.

As retrieval-augmented generation (RAG) systems scale, it becomes increasingly challenging to ensure faithful grounding in external evidence. Large language models may still prioritize parametric knowledge over retrieved information when conflicts arise. We propose a novel training-free decoding framework, \emph{Grounded Decoding}, designed to improve factual consistency in RAG without modifying model parameters. Unlike standard approaches that rely on a single conditional distribution, our method constructs two matched-prompt distributions at every generation step: (1) a full RAG distribution conditioned on the query, retrieved documents, and generated prefix, and (2) a retrieval-only distribution conditioned solely on retrieved evidence and the same prefix. The final next-token distribution is derived as the unique solution to a KL-barycenter objective over the probability simplex, yielding a normalized geometric fusion of the two distributions.This formulation naturally recovers standard RAG when the grounding weight is zero and smoothly shifts probability mass toward retrieved evidence as grounding strength increases. We further introduce a conflict-aware adaptive weighting scheme that dynamically adjusts grounding based on distributional disagreement and retriever confidence. Experiments on ALCE, Natural Questions, and FActScore demonstrate consistent improvements in factual accuracy and citation quality over standard RAG and competitive decoding-time baselines, while maintaining fluency. Our results indicate that probability-level fusion provides a strong and efficient alternative to logit-level intervention methods for faithful RAG decoding.

Ibne Farabi Shihab, Sanjeda Akter, Anuj Sharma

When many reinforcement-learning policies achieve near-optimal return, a post-hoc auditor may have to distinguish among many behaviorally distinct but return-equivalent policies. We formalize this phenomenon through an occupancy-measure analogue of Rashomon capacity: the metric entropy of the near-optimal occupancy region, computed relative to an audited deployment class. Because occupancy measures identify behavior only up to occupancy equivalence, we formulate auditing at the occupancy-class level and distinguish exact local-query oracles from noisy sample-query oracles. Our main exact-query result is conditional: if the audited class contains a $2/H$-separated near-optimal packing whose local signatures are $b$-sparse, then exact local-query auditing requires $Ω(M/b)$ queries; when the packing realizes deployment-class capacity and $b=O(1)$, this becomes $Ω(2^{\Hopt^\cF(\eps)})$. We give a finite discounted hidden-branch MDP attaining this bound and show the exact Bayes success law. For noisy hidden-trigger testing, we prove a mixture lower bound of order $M/β$, where $β$ is the per-sample KL signal, yielding $Ω(2^{\Hopt^\cF(\eps)}/(ρ^2Δ^2))$ for capacity-order packings with $β=O(ρ^2Δ^2)$. We also provide a static target-recognition information lower bound, a transcript-compatible oracle-cover verification upper bound, and a canonical occupancy regularizer whose regularized audited capacity collapses when a trusted reference occupancy is available. Controlled benchmarks distinguish positive sparse-signature instances from high-capacity negative controls where exact auditing is easy, and map the noisy-trigger law to post-processed continuous-control and visual-RL auditing regimes.

Ibne Farabi Shihab, Sanjeda Akter, Anuj Sharma

State abstraction in reinforcement learning is usually formulated as a partition of states based on reward and transition similarity. This excludes a common structural pattern in navigation, graph, and hierarchical decision problems: interface states such as doors, hubs, and bottlenecks naturally participate in more than one region. We introduce \emph{tangle-core abstraction}, an overlapping state-abstraction framework based on graph tangles of empirical transition graphs. The method constructs abstract states from consistently oriented low-order separations and represents shared interfaces through a membership kernel rather than a hard partition. We give value-preservation guarantees for the induced overlapping abstract MDP under an explicit action-consistency condition, identify an interior-homogeneity/boundary-leakage error decomposition, and prove a quantitative interface-overlap result showing when hard partitions incur an avoidable boundary error. Empirically, tangle-core abstractions achieve favorable compression--return tradeoffs against reward-aware, learned, topological-map, and graph-partitioning baselines across bottlenecked tabular domains, procedurally generated mazes, and MiniGrid representations. We also identify a clear failure regime in which transition topology is uninformative, where tangles predictably offer little benefit. These results position graph tangles as an effective topology-aware abstraction prior for decision problems with shared interface structure.

Ibne Farabi Shihab, Sanjeda Akter, Anuj Sharma

Document layout understanding remains data-intensive despite advances in semi-supervised learning. We present a framework that enhances semi-supervised detection by fusing visual predictions with structural priors from text-pretrained LLMs via principled probabilistic weighting. Given unlabeled documents, an OCR-LLM pipeline infers hierarchical regions which are combined with teacher detector outputs through inverse-variance fusion to generate refined pseudo-labels.Our method demonstrates consistent gains across model scales. With a lightweight SwiftFormer backbone (26M params), we achieve 88.2$\pm$0.3 AP using only 5\% labels on PubLayNet. When applied to document-pretrained LayoutLMv3 (133M params), our fusion framework reaches 89.7$\pm$0.4 AP, surpassing both LayoutLMv3 with standard semi-supervised learning (89.1$\pm$0.4 AP, p=0.02) and matching UDOP~\cite{udop} (89.8 AP) which requires 100M+ pages of multimodal pretraining. This demonstrates that LLM structural priors are complementary to both lightweight and pretrained architectures. Key findings include: (1) learned instance-adaptive gating improves over fixed weights by +0.9 AP with data-dependent PAC bounds correctly predicting convergence; (2) open-source LLMs enable privacy-preserving deployment with minimal loss (Llama-3-70B: 87.1 AP lightweight, 89.4 AP with LayoutLMv3); (3) LLMs provide targeted semantic disambiguation (18.7\% of cases, +3.8 AP gain) beyond simple text heuristics.Total system cost includes \$12 for GPT-4o-mini API or 17 GPU-hours for local Llama-3-70B per 50K pages, amortized across training runs.

Ibne Farabi Shihab, Sanjeda Akter, Anuj Sharma

Chromatic Correlation Clustering (CCC) extends Correlation Clustering by assigning semantic colors to edges and requiring each cluster to receive a single color label. Unlike standard CC, whose LP relaxation has integrality gap 2 on complete graphs and admits a 2.06-approximation, the analogous LP for CCC has a strict lower bound of 2.11, and the best known LP-rounding algorithm achieves 2.15. We explain this gap by isolating the source of difficulty: cross-edge chromatic interference. Neutral edges, whose color does not match the candidate cluster color, create an irreducible cost absent from standard CC and force any color-independent rounding scheme to pay an additional mismatch penalty. We make four contributions. First, we prove a Global Integrality Gap Decomposition Theorem showing that the gap of any color-independent CCC rounding algorithm equals the standard CC gap plus an irreducible chromatic penalty Delta(L) > 0. Second, we solve the associated min-max problem and derive the staircase formula Delta(L) = ((L-1)/L) Delta_infinity, where Delta_infinity is approximately 0.0734. In particular, the two-color gap is 2.0967, separating CCC from standard CC already at L = 2. Third, we introduce Color-Coupled Correlation Clustering (C4). Adding the valid global constraint sum_c x_uv^c >= L-1 and a correlated interval-packing rounding scheme makes neutral edges behave like classical negative edges, recovering the optimal 2.06 approximation and bypassing the 2.11 lower bound for the uncoupled LP. Fourth, experiments on extremal instances, real multi-relational networks, and fairness benchmarks validate the theory: empirical LP gaps follow the predicted staircase, and C4 matches the unconstrained approximation ratio under fairness constraints.

Ibne Farabi Shihab, Sanjeda Akter, Anuj Sharma

Majority-vote ensembles achieve variance reduction by averaging over diverse, approximately independent base learners. When training data exhibits Markov dependence, as in time-series forecasting, reinforcement learning (RL) replay buffers, and spatial grids, this classical guarantee degrades in ways that existing theory does not fully quantify. We provide a minimax characterization of this phenomenon for discrete classification in a fixed-dimensional Markov setting, together with an adaptive algorithm that matches the rate on a graph-regular subclass. We first establish an information-theoretic lower bound for stationary, reversible, geometrically ergodic chains in fixed ambient dimension, showing that no measurable estimator can achieve excess classification risk better than $Ω(\sqrt{\Tmix/n})$. We then prove that, on the AR(1) witness subclass underlying the lower-bound construction, dependence-agnostic uniform bagging is provably suboptimal with excess risk bounded below by $Ω(\Tmix/\sqrt{n})$, exhibiting a $\sqrt{\Tmix}$ algorithmic gap. Finally, we propose \emph{adaptive spectral routing}, which partitions the training data via the empirical Fiedler eigenvector of a dependency graph and achieves the minimax rate $\mathcal{O}(\sqrt{\Tmix/n})$ up to a lower-order geometric cut term on a graph-regular subclass, without knowledge of $\Tmix$. Experiments on synthetic Markov chains, 2D spatial grids, the 128-dataset UCR archive, and Atari DQN ensembles validate the theoretical predictions. Consequences for deep RL target variance, scalability via Nyström approximation, and bounded non-stationarity are developed as supporting material in the appendix.

Ibne Farabi Shihab

We study the communication complexity of distributed offline dynamic programming, where a fixed batch dataset is partitioned across (M) machines connected by the data-induced dependency graph. We compare two paradigms: direct boundary-value propagation, which follows Bellman dependencies, and gossip averaging, which mixes local estimates. Our results show that **locality** is the fundamental driver of round complexity. In particular, we prove that no method can achieve (\varepsilon)-accuracy in fewer than (L_\varepsilon = \left\lfloor \log(1/2\varepsilon) / \log(1/γ) \right\rfloor) rounds on graphs of diameter at least (L_\varepsilon), and we show that direct propagation matches this scaling up to constants, attaining error (O(γ^T/(1-γ) + δ/(1-γ))) after (T) rounds. In contrast, gossip-style fitted value iteration incurs an additional (1/\mathrm{gap}(W)) dependence in both convergence rate and asymptotic error. We also prove bandwidth-sensitive lower bounds on path topologies and extend the analysis to asynchronous systems with bounded delays. Together, these results show that spectral dependence is an artifact of gossip-based algorithms, whereas locality is the intrinsic barrier in distributed offline dynamic programming.

Ibne Farabi Shihab, Sanjeda Akter, Anuj Sharma

Large language models increasingly require structured inference, from JSON schema enforcement to multi-lingual parsing, where outputs must satisfy complex constraints. We introduce MetaJuLS, a meta-reinforcement learning approach that learns universal constraint propagation policies applicable across languages and tasks without task-specific retraining. By formulating structured inference as adaptive constraint propagation and training a Graph Attention Network with meta-learning, MetaJuLS achieves 1.5--2.0$\times$ speedups over GPU-optimized baselines while maintaining within 0.2\% accuracy of state-of-the-art parsers. On Universal Dependencies across 10 languages and LLM-constrained generation (LogicBench, GSM8K-Constrained), MetaJuLS demonstrates rapid cross-domain adaptation: a policy trained on English parsing adapts to new languages and tasks with 5--10 gradient steps (5--15 seconds) rather than requiring hours of task-specific training. Mechanistic analysis reveals the policy discovers human-like parsing strategies (easy-first) and novel non-intuitive heuristics. By reducing propagation steps in LLM deployments, MetaJuLS contributes to Green AI by directly reducing inference carbon footprint.

Ibne Farabi Shihab, Sanjeda Akter, Anuj Sharma

Crash classification models in transportation safety are typically evaluated using accuracy, F1, or AUC, metrics that cannot reveal whether a model is silently overfitting. We introduce a spectral diagnostic framework grounded in Random Matrix Theory (RMT) and Heavy-Tailed Self-Regularization (HTSR) that spans the ML taxonomy: weight matrices for BERT/ALBERT/Qwen2.5, out-of-fold increment matrices for XGBoost/Random Forest, empirical Hessians for Logistic Regression, induced affinity matrices for Decision Trees, and Graph Laplacians for KNN. Evaluating nine model families on two Iowa DOT crash classification tasks (173,512 and 371,062 records respectively), we find that the power-law exponent $α$ provides a structural quality signal: well-regularized models consistently yield $α$ within $[2, 4]$ (mean $2.87 \pm 0.34$), while overfit variants show $α< 2$ or spectral collapse. We observe a strong rank correlation between $α$ and expert agreement (Spearman $ρ= 0.89$, $p < 0.001$), suggesting spectral quality captures model behaviors aligned with expert reasoning. We propose an $α$-based early stopping criterion and a spectral model selection protocol, and validate both against cross-validated F1 baselines. Sparse Lanczos approximations make the framework scalable to large datasets.

Ibne Farabi Shihab, Sanjeda Akter, Anuj Sharma

Deep protein structure predictors such as AlphaFold provide confidence estimates (e.g., pLDDT) that are often miscalibrated and degrade under distribution shifts across experimental modalities, temporal changes, and intrinsically disordered regions. We introduce CalPro, a prior-aware evidential-conformal framework for shift-robust uncertainty quantification. CalPro combines (i) a geometric evidential head that outputs Normal-Inverse-Gamma predictive distributions via a graph-based architecture; (ii) a differentiable conformal layer that enables end-to-end training with finite-sample coverage guarantees; and (iii) domain priors (disorder, flexibility) encoded as soft constraints. We derive structure-aware coverage guarantees under distribution shift using PAC-Bayesian bounds over ambiguity sets, and show that CalPro maintains near-nominal coverage while producing tighter intervals than standard conformal methods in regions where priors are informative. Empirically, CalPro exhibits at most 5% coverage degradation across modalities (vs. 15-25% for baselines), reduces calibration error by 30-50%, and improves downstream ligand-docking success by 25%. Beyond proteins, CalPro applies to structured regression tasks in which priors encode local reliability, validated on non-biological benchmarks.

Ibne Farabi Shihab, Sanjeda Akter, Anuj Sharma

Post-training activation compression is essential for deploying Large Language Models (LLMs) on resource-constrained hardware. However, standard methods like Singular Value Decomposition (SVD) are gradient-blind: they preserve high-variance dimensions regardless of their impact on factual knowledge preservation. We introduce Fisher-Aligned Subspace Compression (FASC), a knowledge-aware compression framework that selects subspaces by directly modeling activation-gradient coupling, minimizing a second-order surrogate of the loss function. FASC leverages the Fisher Information Matrix to identify dimensions critical for factual knowledge, which often reside in low-variance but high-gradient-sensitivity subspaces. We propose the Dependence Violation Score (\r{ho}) as a general-purpose diagnostic metric that quantifies activation-gradient coupling, revealing where factual knowledge is stored within transformer architectures. Extensive experiments on Mistral-7B and Llama-3-8B demonstrate that FASC preserves 6-8% more accuracy on knowledge-intensive benchmarks (MMLU, LAMA) compared to variance-based methods at 50% rank reduction, effectively enabling a 7B model to match the factual recall of a 13B uncompressed model. Our analysis reveals that \r{ho} serves as a fundamental signal of stored knowledge, with high-\r{ho} layers emerging only when models internalize factual associations during training.

Ibne Farabi Shihab, Sanjeda Akter, Anuj Sharma

We study black-box optimization of Lipschitz functions under noisy evaluations. Existing adaptive discretization methods implicitly avoid suboptimal regions but do not provide explicit certificates of optimality or measurable progress guarantees. We introduce \textbf{Certificate-Guided Pruning (CGP)}, which maintains an explicit \emph{active set} $A_t$ of potentially optimal points via confidence-adjusted Lipschitz envelopes. Any point outside $A_t$ is certifiably suboptimal with high probability, and under a margin condition with near-optimality dimension $α$, we prove $\Vol(A_t)$ shrinks at a controlled rate yielding sample complexity $\tildeO(\varepsilon^{-(2+α)})$. We develop three extensions: CGP-Adaptive learns $L$ online with $O(\log T)$ overhead; CGP-TR scales to $d > 50$ via trust regions with local certificates; and CGP-Hybrid switches to GP refinement when local smoothness is detected. Experiments on 12 benchmarks ($d \in [2, 100]$) show CGP variants match or exceed strong baselines while providing principled stopping criteria via certificate volume.

Ibne Farabi Shihab, Sanjeda Akter, Anuj Sharma

Hybrid architectures combining state-space models with attention have achieved strong efficiency-quality tradeoffs, yet existing approaches either apply attention uniformly or learn static sparse patterns. This misses a key opportunity: \emph{attention demand should decrease over time as recurring patterns become familiar}. We present a surprising finding from analyzing GPT-2 models: \textbf{88\%} of attention operations retrieve information already predictable from the model's hidden state, and this redundancy does \emph{not} decrease during training. Motivated by this observation, we introduce \textbf{\ours{}} (\textbf{C}onsolidation-based \textbf{R}outing for \textbf{A}daptive \textbf{M}emory), a biologically inspired memory consolidation mechanism that gradually distills episodic retrievals into parametric semantic memory. Unlike prior sparse attention methods, \ours{} exhibits \emph{decreasing attention utilization} over training, achieving a \textbf{37.8$\times$} reduction through a sharp phase transition at approximately 3K steps. We prove that this capability is \emph{impossible} without consolidation: any static routing scheme requires $Ω(f \cdot n)$ attention for tasks with recurring patterns of frequency $f$. On our proposed SRCD benchmark, \ours{} achieves \textbf{100\% retrieval accuracy} at 1.6\% attention compute (vs.\ 68\% for baselines), and consolidated patterns transfer to unseen tasks with \textbf{48--52\%} attention reduction without retraining. Remarkably, the learned consolidation dynamics quantitatively match human episodic-to-semantic memory transition curves from cognitive psychology ($γ= 0.43$ vs.\ $γ_{\text{human}} \approx 0.4$--$0.5$). Code and benchmarks are available at [anonymized].

Ibne Farabi Shihab, Sanjeda Akter, Anuj Sharma

Temporal action localization requires both precise boundary detection and computational efficiency. Current methods apply uniform computation across all temporal positions, wasting resources on easy boundaries while struggling with ambiguous ones. We address this through two complementary innovations: Boundary Distance Regression (BDR), which replaces classification-based boundary detection with signed-distance regression achieving 3.3--16.7$\times$ lower variance; and Adaptive Temporal Refinement (ATR), which allocates transformer depth continuously ($τ\in[0,1]$) to concentrate computation near difficult boundaries. On THUMOS14, our method achieves 56.5\% mAP@0.7 and 58.2\% average mAP@[0.3:0.7] with 151G FLOPs, using 36\% fewer FLOPs than ActionFormer++ (55.7\% mAP@0.7 at 235G). Compared to uniform baselines, we achieve +2.9\% mAP@0.7 (+1.8\% avg mAP, 5.4\% relative) with 24\% fewer FLOPs and 29\% lower latency, with particularly strong gains on short actions (+4.2\%, 8.6\% relative). Training requires 1.29$\times$ baseline FLOPs, but this one-time cost is amortized over many inference runs; knowledge distillation further reduces this to 1.1$\times$ while retaining 99.5\% accuracy. Our contributions include: (i) a theoretically-grounded distance formulation with information-theoretic analysis showing optimal variance scaling; (ii) a continuous depth allocation mechanism avoiding discrete routing complexity; and (iii) consistent improvements across four datasets with gains correlating with boundary heterogeneity.

Ibne Farabi Shihab, Sanjeda Akter, Anuj Sharma

Post-Training Quantization (PTQ) compresses large language models to low bit-widths using a small calibration set, and its quality depends strongly on which samples are chosen. We identify a failure mode in which calibration samples fail to activate outlier channels, hidden dimensions with unusually large activations, causing the quantizer to underestimate their dynamic range and producing per-channel reconstruction errors that dominate layer-wise loss. Motivated by this observation, we argue that PTQ calibration quality is governed more by weighted outlier-channel coverage than by generic sample representativeness, and formulate calibration selection as a weighted set cover problem over outlier channels. The objective is monotone submodular, and the greedy algorithm, COVERCAL, operates on pre-computed activation statistics and requires no GPU time at selection. We further show that the weight choice is internally consistent: under a stylized clipping model, missed weighted coverage upper-bounds surrogate loss, justifying the weighted coverage objective as principled rather than purely empirical. Across LLaMA-2, LLaMA-3, and Mistral, under AWQ and GPTQ backends and five downstream evaluations, COVERCAL improves over random, max-perplexity, max-activation-variance, and stratified baselines, with the largest gains at small calibration budgets. At INT4 with 128 samples, COVERCAL improves MMLU by 1.2 to 1.5 points over random calibration and reduces perplexity degradation by 15 to 30\%; with 64 samples, it matches or exceeds random calibration at 256. The contribution is not a new PTQ backend but a formulation of calibration selection as weighted outlier coverage, with a simple, efficient algorithm and a surrogate-based justification.

Ibne Farabi Shihab, Sanjeda Akter, Anuj Sharma

Continual learning for large language models is typically evaluated through accuracy retention under sequential fine-tuning. We argue that this perspective is incomplete, because uncertainty reliability can degrade earlier and more sharply than top-1 performance. We study this empirically by measuring conformal coverage and calibration error on sequentially fine-tuned models across three model families and eight task sequences drawn primarily from classification and multiple-choice benchmarks. Across the classification-style settings we study, coverage loss exceeds accuracy loss by a factor of roughly \(3.4\times \pm 0.5\times\) on average across seeds; in the most pronounced case, coverage drops from \(0.92\) to \(0.61\), while accuracy remains within three points of baseline. Standard continual-learning methods that preserve accuracy do not automatically preserve coverage, and naive calibration baselines recover only part of the gap. We propose calibration replay, a lightweight post-hoc procedure that maintains a task-specific held-out buffer and refits a task-specific conformal threshold under the current model after each update. It adds no training-time gradient cost, uses less than one percent of the memory of ordinary experience replay, and typically restores coverage to within two points of nominal at buffer size \(m = 200\). We accompany the empirical study with a drift decomposition, a finite-sample recovery theorem showing exact conformal validity under exchangeability, and a mixture-validity proposition explaining why pooled thresholds do not suffice. Our guarantees are stated for classification-style tasks with task-specific buffers; extensions to open-ended generation are exploratory.

Ibne Farabi Shihab, Sanjeda Akter, Anuj Sharma

Reward hacking in Reinforcement Learning (RL) systems poses a critical threat to the deployment of autonomous agents, where agents exploit flaws in reward functions to achieve high scores without fulfilling intended objectives. Despite growing awareness of this problem, systematic detection and mitigation approaches remain limited. This paper presents a large-scale empirical study of reward hacking across diverse RL environments and algorithms. We analyze 15,247 training episodes across 15 RL environments (Atari, MuJoCo, custom domains) and 5 algorithms (PPO, SAC, DQN, A3C, Rainbow), implementing automated detection algorithms for six categories of reward hacking: specification gaming, reward tampering, proxy optimization, objective misalignment, exploitation patterns, and wireheading. Our detection framework achieves 78.4% precision and 81.7% recall across environments, with computational overhead under 5%. Through controlled experiments varying reward function properties, we demonstrate that reward density and alignment with true objectives significantly impact hacking frequency ($p < 0.001$, Cohen's $d = 1.24$). We validate our approach through three simulated application studies representing recommendation systems, competitive gaming, and robotic control scenarios. Our mitigation techniques reduce hacking frequency by up to 54.6% in controlled scenarios, though we find these trade-offs are more challenging in practice due to concept drift, false positive costs, and adversarial adaptation. All detection algorithms, datasets, and experimental protocols are publicly available to support reproducible research in RL safety.

Mohammed Shaiqur Rahman, Ibne Farabi Shihab, Lynna Chu et al.

In this study, we introduce DeepLocalization, an innovative framework devised for the real-time localization of actions tailored explicitly for monitoring driver behavior. Utilizing the power of advanced deep learning methodologies, our objective is to tackle the critical issue of distracted driving-a significant factor contributing to road accidents. Our strategy employs a dual approach: leveraging Graph-Based Change-Point Detection for pinpointing actions in time alongside a Video Large Language Model (Video-LLM) for precisely categorizing activities. Through careful prompt engineering, we customize the Video-LLM to adeptly handle driving activities' nuances, ensuring its classification efficacy even with sparse data. Engineered to be lightweight, our framework is optimized for consumer-grade GPUs, making it vastly applicable in practical scenarios. We subjected our method to rigorous testing on the SynDD2 dataset, a complex benchmark for distracted driving behaviors, where it demonstrated commendable performance-achieving 57.5% accuracy in event classification and 51% in event detection. These outcomes underscore the substantial promise of DeepLocalization in accurately identifying diverse driver behaviors and their temporal occurrences, all within the bounds of limited computational resources.

Sanjeda Akter, Ibne Farabi Shihab, Anuj Sharma

Crash detection from video feeds is a critical problem in intelligent transportation systems. Recent developments in large language models (LLMs) and vision-language models (VLMs) have transformed how we process, reason about, and summarize multimodal information. This paper surveys recent methods leveraging LLMs for crash detection from video data. We present a structured taxonomy of fusion strategies, summarize key datasets, analyze model architectures, compare performance benchmarks, and discuss ongoing challenges and opportunities. Our review provides a foundation for future research in this fast-growing intersection of video understanding and foundation models.

Sanjeda Akter, Ibne Farabi Shihab, Anuj Sharma

The integration of Large Language Models (LLMs) with computer vision is profoundly transforming perception tasks like image segmentation. For intelligent transportation systems (ITS), where accurate scene understanding is critical for safety and efficiency, this new paradigm offers unprecedented capabilities. This survey systematically reviews the emerging field of LLM-augmented image segmentation, focusing on its applications, challenges, and future directions within ITS. We provide a taxonomy of current approaches based on their prompting mechanisms and core architectures, and we highlight how these innovations can enhance road scene understanding for autonomous driving, traffic monitoring, and infrastructure maintenance. Finally, we identify key challenges, including real-time performance and safety-critical reliability, and outline a perspective centered on explainable, human-centric AI as a prerequisite for the successful deployment of this technology in next-generation transportation systems.

Anush Lakshman Sivaraman, Kojo Adu-Gyamfi, Ibne Farabi Shihab et al.

Adverse weather conditions challenge safe transportation, necessitating robust real-time weather detection from traffic camera imagery. We propose a novel framework combining CycleGAN-based domain adaptation with efficient contrastive learning to enhance weather classification, particularly in low-light nighttime conditions. Our approach leverages the lightweight SigLIP-2 model, which employs pairwise sigmoid loss to reduce computational demands, integrated with CycleGAN to transform nighttime images into day-like representations while preserving weather cues. Evaluated on an Iowa Department of Transportation dataset, the baseline EVA-02 model with CLIP achieves a per-class overall accuracy of 96.55\% across three weather conditions (No Precipitation, Rain, Snow) and a day/night overall accuracy of 96.55\%, but shows a significant day-night gap (97.21\% day vs.\ 63.40\% night). With CycleGAN, EVA-02 improves to 97.01\% per-class accuracy and 96.85\% day/night accuracy, boosting nighttime performance to 82.45\%. Our Vision-SigLIP-2 + Text-SigLIP-2 + CycleGAN + Contrastive configuration excels in nighttime scenarios, achieving the highest nighttime accuracy of 85.90\%, with 94.00\% per-class accuracy and 93.35\% day/night accuracy. This model reduces training time by 89\% (from 6 hours to 40 minutes) and inference time by 80\% (from 15 seconds to 3 seconds) compared to EVA-02. By narrowing the day-night performance gap from 33.81 to 8.90 percentage points, our framework provides a scalable, efficient solution for all-weather classification using existing camera infrastructure.

Ibne Farabi Shihab, Sudesh Ramesh Bhagat, Anuj Sharma

This study aims to compare the effectiveness of a robust ensemble model with the state-of-the-art ONE-PEACE Large Language Model (LLM) for accurate detection of sidewalks. Accurate sidewalk detection is crucial in improving road safety and urban planning. The study evaluated the model's performance on Cityscapes, Ade20k, and the Boston Dataset. The results showed that the ensemble model performed better than the individual models, achieving mean Intersection Over Union (mIOU) scores of 93.1\%, 90.3\%, and 90.6\% on these datasets under ideal conditions. Additionally, the ensemble model maintained a consistent level of performance even in challenging conditions such as Salt-and-Pepper and Speckle noise, with only a gradual decrease in efficiency observed. On the other hand, the ONE-PEACE LLM performed slightly better than the ensemble model in ideal scenarios but experienced a significant decline in performance under noisy conditions. These findings demonstrate the robustness and reliability of the ensemble model, making it a valuable asset for improving urban infrastructure related to road safety and curb space management. This study contributes positively to the broader context of urban health and mobility.

Ibne Farabi Shihab, Sanjeda Akter, Anuj Sharma

State-space models (SSMs), particularly the Mamba architecture, have emerged as powerful alternatives to Transformers for sequence modeling, offering linear-time complexity and competitive performance across diverse tasks. However, their large parameter counts pose significant challenges for deployment in resource-constrained environments. We propose a novel unstructured pruning framework tailored for Mamba models that achieves up to 70\% parameter reduction while retaining over 95\% of the original performance. Our approach integrates three key innovations: (1) a gradient-aware magnitude pruning technique that combines weight magnitude and gradient information to identify less critical parameters, (2) an iterative pruning schedule that gradually increases sparsity to maintain model stability, and (3) a global pruning strategy that optimizes parameter allocation across the entire model. Through extensive experiments on WikiText-103, Long Range Arena, and ETT time-series benchmarks, we demonstrate significant efficiency gains with minimal performance degradation. Our analysis of pruning effects on Mamba's components reveals critical insights into the architecture's redundancy and robustness, enabling practical deployment in resource-constrained settings while broadening Mamba's applicability.

Ibne Farabi Shihab, Sanjeda Akter, Anuj Sharma

Traffic crash detection in long-form surveillance videos is essential for improving emergency response and infrastructure planning, yet remains difficult due to the brief and infrequent nature of crash events. We present \textbf{HybridMamba}, a novel architecture integrating visual transformers with state-space temporal modeling to achieve high-precision crash time localization. Our approach introduces multi-level token compression and hierarchical temporal processing to maintain computational efficiency without sacrificing temporal resolution. Evaluated on a large-scale dataset from the Iowa Department of Transportation, HybridMamba achieves a mean absolute error of \textbf{1.50 seconds} for 2-minute videos ($p<0.01$ compared to baselines), with \textbf{65.2%} of predictions falling within one second of the ground truth. It outperforms recent video-language models (e.g., TimeChat, VideoLLaMA-2) by up to 3.95 seconds while using significantly fewer parameters (3B vs. 13--72B). Our results demonstrate effective temporal localization across various video durations (2--40 minutes) and diverse environmental conditions, highlighting HybridMamba's potential for fine-grained temporal localization in traffic surveillance while identifying challenges that remain for extended deployment.

Ibne Farabi Shihab, Sanjeda Akter, Anuj Sharma

Integrating large language models (LLMs) as priors in reinforcement learning (RL) offers significant advantages but comes with substantial computational costs. We present a principled cache-efficient framework for posterior sampling with LLM-derived priors that dramatically reduces these costs while maintaining high performance. At the core of our approach is an adaptive caching mechanism, where cache parameters are meta-optimized using surrogate gradients derived from policy performance. This design enables efficient inference across both discrete text environments (e.g., TextWorld, ALFWorld) and continuous control domains (e.g., MuJoCo), achieving a 3.8--4.7$\times$ reduction in LLM queries and 4.0--12.0$\times$ lower median latencies (85--93\,ms on a consumer GPU) while retaining 96--98\% of uncached performance. Our theoretical analysis provides KL divergence bounds on approximation quality, validated empirically. The framework extends to offline RL, where our CQL-Prior variant improves performance by 14--29\% and reduces training time by 38--40\%. Extensive evaluations across a diverse suite of eight tasks demonstrate the generalizability and practical viability of LLM-guided RL in resource-constrained settings.

Sudesh Ramesh Bhagat, Ibne Farabi Shihab, Anuj Sharma

This study investigates the relationship between deep learning (DL) model accuracy and expert agreement in classifying crash narratives. We evaluate five DL models -- including BERT variants, USE, and a zero-shot classifier -- against expert labels and narratives, and extend the analysis to four large language models (LLMs): GPT-4, LLaMA 3, Qwen, and Claude. Our findings reveal an inverse relationship: models with higher technical accuracy often show lower agreement with human experts, while LLMs demonstrate stronger expert alignment despite lower accuracy. We use Cohen's Kappa and Principal Component Analysis (PCA) to quantify and visualize model-expert agreement, and employ SHAP analysis to explain misclassifications. Results show that expert-aligned models rely more on contextual and temporal cues than location-specific keywords. These findings suggest that accuracy alone is insufficient for safety-critical NLP tasks. We argue for incorporating expert agreement into model evaluation frameworks and highlight the potential of LLMs as interpretable tools in crash analysis pipelines.

Sanjeda Akter, Ibne Farabi Shihab, Anuj Sharma

We introduce Sequential-EDFL (Empirical Dynamic Formal Lift), applying anytime-valid sequential testing to language model generation stopping. Our approach tracks information lift -- the log-likelihood ratio between full models and deliberately weakened "skeleton" baselines -- using self-normalized empirical-Bernstein e-processes that provide formal delta-level error control regardless of stopping time. We handle unknown centering through online mean estimation, combine multiple parameters via mixture e-processes, and support adaptive resets under distributional drift. On six benchmarks, Sequential-EDFL reduces generation by 22-28% vs. sequential baselines while maintaining delta-level control with 12% computational overhead. We introduce automated skeletons (distilled submodels, randomized logits) and show robustness across skeleton families. Composing EDFL with a lightweight correctness gate (sentence boundaries + verifier) improves end-task correctness while preserving anytime-valid guarantees by only delaying stopping. Our certificates control information sufficiency, not factual correctness -- 10.9% of stopped sequences remain incorrect even with the gate (13.2-22.7% without it). EDFL serves as a first-stage filter reducing verification burden by 83%, not as a standalone solution for safety-critical domains.

Ibne Farabi Shihab, Weiheng Chai, Jiyang Wang et al.

Driver monitoring systems require not just high accuracy but reliable, well-calibrated confidence scores for safety-critical deployment. While direct low-resolution training yields high overall accuracy, it produces poorly calibrated predictions that can be dangerous in safety-critical scenarios. We propose a resource-aware adaptive super-resolution framework that optimizes for model calibration and high precision-recall on critical events. Our approach achieves state-of-the-art performance on safety-centric metrics: best calibration (ECE of 5.8\% vs 6.2\% for LR-trained baselines), highest AUPR for drowsiness detection (0.78 vs 0.74), and superior precision-recall for phone use detection (0.74 vs 0.71). A lightweight artifact detector (0.3M parameters, 5.2ms overhead) provides additional safety by filtering SR-induced hallucinations. While LR-trained video models serve as strong general-purpose baselines, our adaptive framework represents the state-of-the-art solution for safety-critical applications where reliability is paramount.

Sanjeda Akter, Ibne Farabi Shihab, Anuj Sharma

Large language models often produce confident but incorrect outputs, creating a critical need for reliable uncertainty quantification with formal abstention guarantees. We introduce information-lift certificates that compare model probabilities to a skeleton baseline, accumulating evidence through sub-gamma PAC-Bayes bounds that remain valid under heavy-tailed distributions where standard concentration inequalities fail. On eight diverse datasets, our method achieves 77.0\% coverage at 2\% risk, outperforming recent baselines by 10.0 percentage points on average. In high-stakes scenarios, we block 96\% of critical errors compared to 18-31\% for entropy-based methods. While our frequency-based certification does not guarantee severity-weighted safety and depends on skeleton quality, performance degrades gracefully under distributional shifts, making the approach practical for real-world deployment.

Ibne Farabi Shihab, Sanjeda Akter, Anuj Sharma

Sparse-reward reinforcement learning (RL) remains fundamentally hard: without structure, any agent needs $Ω(|\mathcal{S}||\mathcal{A}|/p)$ samples to recover rewards. We introduce Policy-Aware Matrix Completion (PAMC) as a first concrete step toward a structural reward learning framework. Our key idea is to exploit approximate low-rank + sparse structure in the reward matrix, under policy-biased (MNAR) sampling. We prove recovery guarantees with inverse-propensity weighting, and establish a visitation-weighted error-to-regret bound linking completion error to control performance. Importantly, when assumptions weaken, PAMC degrades gracefully: confidence intervals widen and the algorithm abstains, ensuring safe fallback to exploration. Empirically, PAMC improves sample efficiency across Atari-26 (10M steps), DM Control, MetaWorld MT50, D4RL offline RL, and preference-based RL benchmarks, outperforming DrQ-v2, DreamerV3, Agent57, T-REX/D-REX, and PrefPPO under compute-normalized comparisons. Our results highlight PAMC as a practical and principled tool when structural rewards exist, and as a concrete first instantiation of a broader structural reward learning perspective.

Ibne Farabi Shihab, Sanjeda Akter, Anuj Sharma

We introduce the first differentiable approximation of range-partition entropy, a complexity measure from computational geometry that directly bounds algorithmic runtime. Unlike architectural modifications, our method is a complementary regularizer that provides orthogonal efficiency gains when combined with existing optimizations. We establish theoretical guarantees in computational geometry, achieving 4--5$\times$ provable speedups on convex hull and triangulation with $<$0.2\% error. On ImageNet-1K with ViT-Base, entropy regularization achieves 80.1\% top-1 accuracy at 80\% sparsity (1.60$\times$ standalone speedup), and when combined with FlashAttention yields 2.07$\times$ speedup versus 1.63$\times$ for FlashAttention alone. On large language models (LLaMA-2 7B, Mistral-7B, Phi-2), we achieve 1.48--1.60$\times$ inference speedups at 70--75\% sparsity with minimal quality degradation (ROUGE-L drops of 0.3--0.4 points, perplexity increase of 0.9). Unlike prior regularization methods that target output distributions, we directly minimize representation complexity, yielding both efficiency gains and improved robustness through semantically structured sparsity patterns (IoU 0.73 vs 0.41 for magnitude pruning, CIFAR-100-C mCE 48.7 vs 55.4). Benefits are strongest for geometry and vision transformers, with more modest but measurable gains on LLMs, demonstrating that complexity regularization offers a principled pathway to joint efficiency-robustness optimization.

Sanjeda Akter, Ibne Farabi Shihab, Anuj Sharma

The reasoning processes of large language models often lack faithfulness; a model may generate a correct answer while relying on a flawed or irrelevant reasoning trace. This behavior, a direct consequence of training objectives that solely reward final-answer correctness, severely undermines the trustworthiness of these models in high-stakes domains. This paper introduces \textbf{Counterfactual Sensitivity Regularization (CSR)}, a novel training objective designed to forge a strong, causal-like dependence between a model's output and its intermediate reasoning steps. During training, CSR performs automated, operator-level interventions on the generated reasoning trace (e.g., swapping ``+'' with ``-'') to create a minimally-perturbed counterfactual. A regularization term then penalizes the model if this logically flawed trace still yields the original answer. Our efficient implementation adds only 8.7\% training overhead through warm-start curriculum and token-subset optimization. We evaluate faithfulness using \textbf{Counterfactual Outcome Sensitivity (COS)}, a metric quantifying how sensitive the final answer is to such logical perturbations. Across diverse structured reasoning benchmarks -- arithmetic (GSM8K), logical deduction (ProofWriter), multi-hop QA (HotpotQA), and code generation (MBPP) -- models trained with CSR demonstrate a vastly superior trade-off between accuracy and faithfulness. CSR improves faithfulness over standard fine-tuning and process supervision by up to 70 percentage points, with this learned sensitivity generalizing to larger models and enhancing the performance of inference-time techniques like self-consistency.

Sudesh Bhagat, Ibne Farabi Shihab, Jonathan Wood

This research investigates the efficacy of machine learning (ML) and deep learning (DL) methods in detecting misclassified intersection-related crashes in police-reported narratives. Using 2019 crash data from the Iowa Department of Transportation, we implemented and compared a comprehensive set of models, including Support Vector Machine (SVM), XGBoost, BERT Sentence Embeddings, BERT Word Embeddings, and Albert Model. Model performance was systematically validated against expert reviews of potentially misclassified narratives, providing a rigorous assessment of classification accuracy. Results demonstrated that while traditional ML methods exhibited superior overall performance compared to some DL approaches, the Albert Model achieved the highest agreement with expert classifications (73% with Expert 1) and original tabular data (58%). Statistical analysis revealed that the Albert Model maintained performance levels similar to inter-expert consistency rates, significantly outperforming other approaches, particularly on ambiguous narratives. This work addresses a critical gap in transportation safety research through multi-modal integration analysis, which achieved a 54.2% reduction in error rates by combining narrative text with structured crash data. We conclude that hybrid approaches combining automated classification with targeted expert review offer a practical methodology for improving crash data quality, with substantial implications for transportation safety management and policy development.

Sudesh Bhagat, Raghupathi Kandiboina, Ibne Farabi Shihab et al.

Road traffic crashes are a significant global cause of fatalities, emphasizing the urgent need for accurate crash data to enhance prevention strategies and inform policy development. This study addresses the challenge of alcohol inference mismatch (AIM) by employing database narrative alignment to identify AIM in crash data. A framework was developed to improve data quality in crash management systems and reduce the percentage of AIM crashes. Utilizing the BERT model, the analysis of 371,062 crash records from Iowa (2016-2022) revealed 2,767 AIM incidents, resulting in an overall AIM percentage of 24.03%. Statistical tools, including the Probit Logit model, were used to explore the crash characteristics affecting AIM patterns. The findings indicate that alcohol-related fatal crashes and nighttime incidents have a lower percentage of the mismatch, while crashes involving unknown vehicle types and older drivers are more susceptible to mismatch. The geospatial cluster as part of this study can identify the regions which have an increased need for education and training. These insights highlight the necessity for targeted training programs and data management teams to improve the accuracy of crash reporting and support evidence-based policymaking.

Ibne Farabi Shihab

Using mobile phone video of the fingertip as a data source for estimating vital signs such as heart rate (HR) and respiratory rate (RR) during daily life has long been suggested. While existing literature indicates that these estimates are accurate to within several beats or breaths per minute, the data used to draw these conclusions are typically collected in laboratory environments under careful experimental control, and yet the results are assumed to generalize to daily life. In an effort to test it, a team of researchers collected a large dataset of mobile phone video recordings made during daily life and annotated with ground truth HR and RR labels from N=111 participants. They found that traditional algorithm performance on the fingerprint videos is worse than previously reported (7 times and 13 times worse for RR and HR, respectively). Fortunately, recent advancements in deep learning, especially in convolutional neural networks (CNNs), offer a promising solution to improve this performance. This study proposes a new method for estimating HR and RR using a novel 3D deep CNN, demonstrating a reduced error in estimated HR by 68% and RR by 75%. These promising results suggest that regressor-based deep learning approaches should be used in estimating HR and RR.

Ibne Farabi Shihab, Sanjeda Akter, Anuj Sharma

Quantum machine learning for spin and molecular systems faces critical challenges of scarce labeled data and computationally expensive simulations. To address these limitations, we introduce Hamiltonian-Masked Autoencoding (HMAE), a novel self-supervised framework that pre-trains transformers on unlabeled quantum Hamiltonians, enabling efficient few-shot transfer learning. Unlike random masking approaches, HMAE employs a physics-informed strategy based on quantum information theory to selectively mask Hamiltonian terms based on their physical significance. Experiments on 12,500 quantum Hamiltonians (60% real-world, 40% synthetic) demonstrate that HMAE achieves 85.3% $\pm$ 1.5% accuracy in phase classification and 0.15 $\pm$ 0.02 eV MAE in ground state energy prediction with merely 10 labeled examples - a statistically significant improvement (p < 0.01) over classical graph neural networks (78.1% $\pm$ 2.1%) and quantum neural networks (76.8% $\pm$ 2.3%). Our method's primary advantage is exceptional sample efficiency - reducing required labeled examples by 3-5x compared to baseline methods - though we emphasize that ground truth values for fine-tuning and evaluation still require exact diagonalization or tensor networks. We explicitly acknowledge that our current approach is limited to small quantum systems (specifically limited to 12 qubits during training, with limited extension to 16-20 qubits in testing) and that, while promising within this regime, this size restriction prevents immediate application to larger systems of practical interest in materials science and quantum chemistry.