Safety management in the chemical process industry remains a critical challenge due to recurring high impact industrial accidents and the limited predictive capability of conventional threshold based safety systems. Traditional PLC-SCADA frameworks rely on static alarm limits and reactive shutdown logic, which often fail to detect early stage nonlinear deviations in complex, multivariate processes. This study presents ChemSafeAI+, a machine learning driven dynamic safety and optimization framework designed to augment existing industrial control architectures. The system integrates real-time anomaly detection using gradient-boosting models, predictive analytics, safety action processing, operator aware visualization dashboards, and traceable console logging within a unified, modular architecture. The framework is evaluated using a validated synthetic dataset derived from the Haber-Bosch ammonia synthesis process, capturing realistic thermodynamic, kinetic, and operational variability across 5000 operating scenarios. Experimental results demonstrate strong anomaly detection capability and consistent early warning behavior across multiple abnormal operating conditions. SHAP-based explainability provides both global and local interpretability, aligning model decisions with domain relevant process variables. By combining predictive intelligence with safety oriented decision logic and operator traceability, ChemSafeAI+ demonstrates the feasibility of ML driven supervisory safety systems for proactive risk mitigation and improved operational resilience in industrial chemical environments.
Language-conditioned autonomous agents rely on natural language to represent internal state, reason about goals, and select actions. Despite recent advances in reasoning and planning, such agents remain unreliable in long-horizon tasks. In this work, we identify state drift as a fundamental and underexplored failure mode, characterized by persistent divergence between an agent’s internal textual state and the true environment state over time. We study state drift through controlled experiments with language-driven agents operating in long-horizon settings. By comparing fact-level internal belief representations against ground-truth environment states across sequential interactions, we show that state drift can arise and persist even when individual reasoning steps are locally coherent and logically valid. This indicates that long-horizon failures cannot be explained solely by step-wise reasoning errors. Moreover, we find that increasing context capacity does not mitigate state drift in deterministic environments, suggesting that the phenomenon is not simply a consequence of limited memory or forgetting. Instead, our results point to a structural limitation of using natural language as an internal state representation. Ensuring semantic state consistency over extended horizons thus emerges as a distinct and unresolved challenge for language-conditioned autonomy, with important implications for the design and evaluation of reliable autonomous agents.
Safety management in the chemical process industry remains a critical challenge due to recurring high impact industrial accidents and the limited predictive capability of conventional threshold based safety systems. Traditional PLC–SCADA frameworks rely on static alarm limits and reactive shutdown logic, which often fail to detect early stage nonlinear deviations in complex, multivariate processes. This study presents ChemSafeAI+, a machine learning driven dynamic safety and optimization framework designed to augment existing industrial control architectures. The system integrates real-time anomaly detection using gradient-boosting models, predictive analytics, safety action processing, operator aware visualization dashboards, and traceable console logging within a unified, modular architecture. The framework is evaluated using a validated synthetic dataset derived from the Haber–Bosch ammonia synthesis process, capturing realistic thermodynamic, kinetic, and operational variability across 5000 operating scenarios. Experimental results demonstrate strong anomaly detection capability and consistent early warning behavior across multiple abnormal operating conditions. SHAP-based explainability provides both global and local interpretability, aligning model decisions with domain relevant process variables. By combining predictive intelligence with safety oriented decision logic and operator traceability, ChemSafeAI+ demonstrates the feasibility of ML driven supervisory safety systems for proactive risk mitigation and improved operational resilience in industrial chemical environments.
Recent work on neural scaling demonstrates consistent performance gains with increased data and model capacity, yet these improvements are typically assessed using surface-level metrics that do not capture factual reliability. In multi-document summarization (MDS), this limitation is particularly acute, as scaling has been shown to amplify hallucination and content distortion. In this paper, we investigate the empirical scaling behaviour of faithfulness-aware transformers under tightly controlled conditions, using LSHT as a fixed architectural and training baseline. Rather than proposing new scaling laws, we analyze how summarization quality, faithfulness and efficiency evolve as dataset size and model capacity are independently increased, while holding architecture, optimization, decoding and hardware constant. All experiments are conducted exclusively on the Multi-News benchmark to avoid cross-dataset confounds. Across ROUGE, coverage, repetition and faithfulness-oriented metrics, we show that lexical overlap and factual consistency follow distinct scaling dynamics. Faithfulness improves most rapidly during early data scaling (approximately 3–4% relative gain from 3k to 12k samples) but exhibits diminishing marginal returns at larger scales, whereas ROUGE continues to increase more smoothly. We further show that faithfulness is more sensitive to data diversity than to volume alone and identify practical scaling regimes that maximize faithfulness gains relative to computational cost. These results establish empirical expectations for scaling faithfulness-aware MDS systems and provide actionable guidance for reliable summarization under realistic resource constraints.
We present MINT (Multilingual Indic Neural Transformer), a compact 14.7M parameter encoder-decoder architecture for abstractive summarization across seven Indic languages. MINT is designed specifically to operate within the memory envelope of a single commodity NVIDIA T4 GPU (15 GB VRAM), addressing the paradox in which models serving the most resource-constrained communities are themselves the most resource-intensive to deploy. The architecture incorporates Rotary Position Embeddings (RoPE), SiLU feed-forward activations, DropPath regularization, weight tying, and a custom 32,000-token SentencePiece Unigram tokenizer trained over balanced Indic corpora. Training proceeds in two phases on the XL-Sum BBC dataset across Hindi, Bengali, Marathi, Tamil, Telugu, Punjabi, and Urdu: a fluency phase (epochs 1-15) using linear warmup with cosine decay, followed by a refinement phase (epochs 16-25) with a flat low learning rate and a combined coverage-attention entropy loss that jointly penalizes repetition and hallucination. We conduct the first identical-regime comparison in Indic summarization, fine-tuning both IndicBART (440M parameters) and mT5-small (556M parameters) under the same loss function, optimizer, decoding strategy, and data pipeline as MINT’s refinement phase. On the XL-Sum test set, MINT achieves an average ROUGE-1 of 0.1187 at epoch 15, rising to 0.1302 on validation after full refinement, reaching approximately 84.8% of IndicBART’s ROUGE-1 (0.1409) on the six overlapping languages while using only 3.3% of its parameters. A critical methodological contribution of this work is the demonstration that the standard Google rouge_score library returns zero for all Indic scripts due to English centric tokenization; we implement and advocate for whitespace-based ROUGE evaluation as the correct approach. MINT additionally benefits from BERTScore-F1 of 0.8497 (via XLM-RoBERTa-Large) and LaBSE embedding cosine similarity of 0.4306, confirming that generated summaries carry semantic meaning even when surface overlap metrics are modest. All code and checkpoints are publicly released.
Multi-document summarization (MDS) under strict context budgets is acutely vulnerable to hallucination, cross-document contradiction, entity drift, and redundant paraphrasing. Existing models address these issues only implicitly through training objectives, leaving decoding as an ad-hoc pipeline layered on maximum-likelihood generation. Arguing that faithfulness is fundamentally a constraint satisfaction problem rather than a fluency optimization problem, we introduce FADCO (Faithfulness-Aware Decoding via Constrained Optimization). FADCO is a model-agnostic inference-time framework that encodes evidence grounding, non-contradiction, and redundancy control as explicit constraints within a Lagrangian-relaxed objective. To support this framework, we formally diagnose and resolve beam collapse, a failure mode in which standard beam search degrades constrained selection to greedy decoding, by employing a mixed candidate pool that increases verifier support diversity. Furthermore, we resolve log-probability scale dominance through rank-based multi-objective aggregation and introduce a bounded local repair operator with provable termination and edit minimality guarantees under a strict retry budget of R=3. Evaluated on MultiNews using MiniCheck-FlanT5, QA-F1, and NLI entailment, preliminary stratified validation demonstrates that bounded self-healing improves MiniCheck scores by 65.7 percent and NLI entailment by 59.0 percent, while reducing contradictions by 5.8 percent. These gains incur only a negligible 0.69 percent ROUGE-1 trade-off, demonstrating a highly favorable faithfulness-fluency Pareto frontier for inference-time decoding interventions.