QSPR modeling has long been a cornerstone of property-oriented fuel design. The approach establishes quantitative relationships between molecular structures and physicochemical properties, allowing researchers to screen candidate fuels for clean energy transition and high-efficiency engine design before synthesis. Yet conventional QSPR workflows carry well-known burdens. Models generally rely on manually engineered molecular descriptors, are usually developed to predict a single property, and often cannot readily incorporate newly available data. Adding fresh measurements may require descriptor reselection, model updates, or full retraining, limiting scalability and long-term applicability. FuelProp-LM was designed to bypass these constraints by unifying multi-property prediction within a single adaptable model.

To construct the framework, the researchers fine-tuned four compact open-source language models, each with fewer than 10 billion parameters: Qwen2.5-7B-Instruct, Llama-3.1-8B-Instruct, Phi-3.5-mini-instruct, and DeepSeek-R1-Distill-Qwen-7B. Training drew on 100,000 data entries obtained from the PubChem database. Notably, no task-specific fuel property data was used during the instruction fine-tuning stage. Instead, the models learned transferable structural reasoning capabilities from general molecular data, a design choice that proved essential to generalizability across downstream fuel-property tasks.

Prediction itself relies on dynamic in-context learning. The team integrated a retrieval algorithm built on 166-bit MACCS molecular fingerprints and Tanimoto similarity search, executed efficiently through the Hierarchical Navigable Small World, or HNSW, algorithm. When a target molecule is submitted, the system automatically retrieves structurally similar reference molecules from updated property databases and weaves them into informative in-context prompts. This enables real-time database updating without any modification of model weights, so newly measured properties can be exploited immediately.

Evaluations covered 10 critical fuel property datasets spanning ignitability, sooting tendency, volatility, and thermodynamics. The fine-tuned 7-billion-parameter FuelProp-LM consistently outperformed the 685-billion-parameter industry-leading DeepSeek-V3.2 baseline and compared favorably with several conventional machine learning approaches. The result is a striking demonstration of parameter efficiency: a model roughly one-hundredth the size of the leading general-purpose LLM delivered competitive or superior predictive performance across all downstream tasks, without additional fine-tuning on the evaluated datasets.

Beyond raw accuracy, the framework eliminates the need for multiple independent QSPR models. Ten essential fuel properties are predicted within one language model using SMILES strings and retrieved in-context examples, removing the tedious process of manual descriptor construction and task-specific model retraining. The accuracy, generalizability, data scalability, and ease of use position FuelProp-LM as a flexible tool for future AI-assisted sustainable fuel design.

To explain why instruction tuning improves QSPR accuracy so dramatically, the researchers conducted an in-depth attention layer analysis, probing the internal workings of the Transformer architecture. The results show that fine-tuning fundamentally reallocates attention patterns within the intermediate layers responsible for semantic integration and reasoning. Attention shifted heavily toward the property description and the first in-context example, the structurally most similar molecule, rather than distributing uniformly across the prompt. Attention to the target molecule’s SMILES string also increased, strengthening the intrinsic mapping between chemical structure and target properties, while attention scores assigned to non-essential prompt components, such as repetitive question text or less relevant context examples, decreased significantly.

These mechanistic insights reveal how instruction tuning and in-context learning enhance molecular property prediction by sharpening chemically relevant attention patterns, offering a window into what is often treated as a black box. The findings suggest that domain-adapted smaller models can outperform far larger general-purpose systems on specialized scientific tasks when trained with the right instructional structure and contextual retrieval.

Professor Bin Yang, corresponding author of the study, said FuelProp-LM bypasses the tedious process of manual descriptor construction and task-specific model retraining. By offering high accuracy, data scalability, and deployment ease, he noted, the framework provides an accessible and flexible tool for future AI-assisted sustainable fuel design. As the clean energy transition accelerates and engine designers seek fuels tailored for efficiency and low emissions, language models adapted through instruction tuning and in-context learning may play a significant role in the next generation of fuel discovery.

Subject of Research: Tsinghua University researchers introduce FuelProp-LM, a language model framework combining instruction tuning and dynamic in-context learning to predict multiple fuel properties directly from molecular SMILES strings.

Article Title: Adapting language models for fuel property prediction

Article References: Adapting language models for fuel property prediction. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: large language models, fuel property prediction, QSPR, instruction tuning, in-context learning, SMILES, Tsinghua University, ENGINEERING Energy, AI chemistry, sustainable fuels, Adapting, language

The core insight behind SENNA, short for Semantic ENcoding for fair tabular data geNerAtion, is that bias enters the pipeline at multiple points, so control must be applied at multiple points as well. The framework operates on tabular records, the rows of numbers and categories that dominate real-world datasets, from census tables to loan applications. Each record is first converted into a natural-language sentence using a fixed template, so that a row describing age, occupation, and income becomes a readable description. This conversion is purely representational and does not alter the underlying distribution, but it unlocks something powerful: pretrained language models can now encode the semantic relationships between attributes, rather than treating every column as an independent dimension the way conventional tabular encoders often do.

The technical heart of the method is disentangled representation learning. A frozen BERT encoder produces contextualized embeddings of each sentence, and two trainable projection heads then split this representation into two separate latent variables. The first, called the task-relevant variable, is trained to concentrate exactly the information needed to predict the target label, such as whether a loan will default. The second, the task-independent variable, is trained to forget the label entirely, using an adversarial classifier that tries to recover the label from it while a gradient reversal layer teaches the representation to confuse that classifier. The two variables are additionally pushed toward statistical independence with the Hilbert-Schmidt Independence Criterion, and a reconstruction objective prevents the latent space from collapsing into degenerate solutions. Variational encoders with Kullback-Leibler regularization keep the whole optimization smooth. The result is a latent space in which what matters for prediction and what does not are, at least approximately, separated.

This separation is then exploited at the generation stage, which is where large language models come in. Recent LLM-based tabular generators transform records into text and use in-context learning, presenting a handful of example records in the prompt and asking the model to produce new ones. The approach is flexible and expressive, but it is notoriously sensitive to which examples are chosen. When the in-context exemplars reflect demographic imbalance, the generated samples inherit the same skew, a phenomenon the authors describe as bias propagation. SENNA counters this by selecting exemplars not at random but through k-means clustering in the task-independent latent space. Because that space was trained to exclude label information, the selected exemplars cover diverse semantic territory without reinforcing spurious correlations between sensitive attributes and outcomes. The model is then prompted with a fixed template shared by training data, exemplars, and generated samples, and asked to produce candidate records that are converted back into tabular form.

Generation alone is not trusted. A two-stage post-generation process screens every candidate. First, individual fairness is enforced through nearest-neighbor validation: each generated sentence is encoded into the task-relevant space, its k nearest real neighbors are found, and a majority vote estimates the expected label. A candidate is retained only if its extracted label matches this local estimate, filtering out semantically inconsistent samples that would distort the decision boundary. Second, group fairness is enforced through balancing: the validated dataset is partitioned by protected-group membership and outcome, and subgroups are subsampled toward approximately uniform sizes, reducing residual correlations between sensitive attributes and predictions. Crucially, protected attributes are never used during representation learning or exemplar selection; they enter only at balancing and at evaluation, a design choice that limits how much sensitive information the pipeline itself consumes.

The empirical evaluation is unusually thorough for this area. The team compared SENNA against five established baselines, including the GAN-based CTGAN and TabFairGAN and the LLM-based GReaT, Curated LLM, and TabGen-ICL, on three standard fairness benchmarks: Adult, Credit scoring, and German credit. They also instantiated the framework with four different large language models, Claude 3.5 Haiku, GPT-4o-mini, Gemini 3 Flash, and Gemini 3 Pro, to test whether the approach generalizes across backbone architectures. Data quality was measured with density, coverage, and distance to closest record; utility with F1-scores of downstream classifiers trained on synthetic data and tested on real data; and fairness with a battery of metrics spanning statistical parity, equalized odds difference, predictive parity, ABROCA, and the Theil Index for individual-level inequality.

The results paint a nuanced but encouraging picture. On the larger and more imbalanced datasets, SENNA variants achieved substantially higher density and coverage than the baselines, and the strongest backbones produced the most locally realistic samples. On utility, SENNA did not dominate every column, which the authors present as a feature rather than a flaw: it remained competitive in F1-score while improving several disparity metrics, offering a balanced trade-off rather than raw performance maximization. Notably, the fairness gains tracked the control mechanisms rather than backbone scale. Stronger models improved fidelity and predictive performance, but the reduction in group and individual disparities came from the disentangled exemplar selection, validation, and balancing stages, suggesting the framework ports well across models.

The study is also candid about its own weaknesses, which lends it credibility. A per-group analysis of the validator revealed that acceptance rates were not uniform: on German credit, female positive-class candidates were accepted at only 0.400 versus 0.756 for the male majority cell, and the smallest post-balancing cells fell far below the nominal 25 percent target. Validation-induced filtering, not just balancing, drives residual imbalance, and the authors propose monitoring acceptance rates as a diagnostic. Quantitative probing showed that disentanglement is approximate rather than complete: on the tiny German credit dataset, a linear probe could still recover label information from the nominally task-independent space, while Credit scoring showed the cleanest separation. The authors also flag that no formal privacy audit, such as a membership-inference attack, has been performed, and that verbatim reproduction of in-context exemplars remains a risk inherent to prompting-based generation.

The broader implications reach beyond one framework. Prior work has shown that repeatedly training on synthetic data can amplify bias through fairness-feedback loops, that fairness measured on synthetic distributions does not always transfer to models trained on them, and that recursive reuse of generated data can cause model collapse. SENNA’s central lesson is that fairness in LLM-based synthesis should be treated as a pipeline-level property, engineered at every stage from representation to post-generation screening, rather than bolted onto a single model. The authors point toward future work on intersectional and multi-valued sensitive attributes, causal fairness guarantees, robustness to distribution shift, and richer serialization beyond templates. For now, the framework, released with full code and configuration, offers practitioners in regulated domains something they have largely lacked: a controllable dial for navigating the unavoidable tension between predictive accuracy and equitable outcomes in synthetic data.

Subject of Research: Fairness-aware synthetic tabular data generation using disentangled semantic encoding and large language models

Article Title: Fair synthetic tabular data generation via disentangled semantic encoding and large language models

Article References: Giang Thi Thu, H., Doan Viet, T., Nguyen Thanh, T., Ta Quoc, T., Le Quy, T., Nguyen Long, G., & Ban, H.-B. (2026). Fair synthetic tabular data generation via disentangled semantic encoding and large language models. Discover Artificial Intelligence, 6(1), Article 1468. https://doi.org/10.1007/s44163-026-02383-w

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02383-w

Keywords: synthetic data, tabular data generation, large language models, algorithmic fairness, disentangled representations, in-context learning, machine learning, group fairness, individual fairness, data privacy, BERT, bias mitigation

The research team, led by Amr A. Elgharib and colleagues, assembled a cohort of 162 patients aged 15 and over who presented with abdominal pain and were clinically suspected of having appendicitis between March 2024 and March 2025. All of them underwent appendectomy, and the final word on their diagnosis came from histopathological examination of the removed appendix, the definitive reference standard. The cohort was deliberately enriched: roughly 70 percent of patients turned out to have confirmed appendicitis, a prevalence far higher than in a general emergency department, which shapes how the results must be interpreted.

The researchers fed ChatGPT-5 structured clinical narratives for each patient, including age, sex, symptoms such as fever, nausea, vomiting and right lower quadrant pain, physical examination findings like rebound tenderness, and laboratory values including white blood cell count and neutrophil percentage. The prompt cast the model as a general surgery physician in an emergency department and, crucially, forbade it from using any established appendicitis scoring system such as Alvarado, AIR, RIPASA or AAS. Instead, the model had to rely on its own clinical reasoning, outputting either a diagnosis of acute or non-acute appendicitis and, for positive cases, a percentage probability stratified into low, intermediate or high risk.

The study was designed in two phases following the classic machine learning convention of a 70-30 split. In Phase 1, 114 cases were presented without any prior exposure to labeled examples, simulating how a clinician might use a readily available chatbot off the shelf. After the model’s predictions were recorded, the true diagnoses were revealed, and in Phase 2 the remaining 48 cases were used to test whether this in-context learning improved performance. The prompt format was fixed across all cases, and the model’s memory was cleared before each evaluation round to test its default behavior.

The headline numbers are eye-catching. In Phase 1 without randomization, ChatGPT-5 achieved 100 percent sensitivity, meaning it caught every single case of appendicitis with zero false negatives, alongside 81.6 percent specificity, a positive predictive value of 91.6 percent, a negative predictive value of 100 percent, and overall accuracy of 93.9 percent. The Alvarado score, by contrast, managed only 67.1 percent sensitivity, missing 25 true cases, though its specificity was higher at 92.1 percent, and its overall accuracy came to 75.4 percent. Statistical comparison using McNemar’s test for paired proportions confirmed that the model’s sensitivity and accuracy advantages were highly significant, with p-values below 0.001.

But the story took a twist when the researchers examined a subtle methodological trap known as order leakage. When cases are presented in a sequence, a language model can inadvertently exploit the ordering of the data rather than genuine clinical reasoning, artificially inflating its apparent accuracy. To probe this, the team re-ran the evaluations with the case order randomized. Sensitivity barely budged, remaining at 98.7 percent in Phase 1, but specificity collapsed dramatically from 81.6 percent to 36.8 percent, a difference that was highly statistically significant. In other words, part of the model’s apparent precision in ruling appendicitis in may have been an artifact of how the cases were sequenced in the prompt.

Phase 2 told a more balanced story. After in-context learning, the non-randomized model reached 100 percent sensitivity, 93.8 percent specificity, 97.0 percent positive predictive value, 100 percent negative predictive value and 97.9 percent accuracy, statistically indistinguishable from the Alvarado score’s 96.9 percent sensitivity and 93.8 percent specificity. Agreement between the two methods, measured with Cohen’s kappa, climbed from a moderate 0.503 in Phase 1 to an almost perfect 0.952 in Phase 2. The authors note that in-context learning produced a modest, non-significant gain in specificity, while sensitivity was largely unaffected, suggesting the model’s core ability to exclude appendicitis was robust from the start.

The findings sit within a growing but cautionary literature. Previous evaluations have found that ChatGPT’s answers to surgeon-designed appendicitis management questions were clinically pertinent but inconsistent, and that ChatGPT-3.5 performed significantly worse than clinicians for some acute abdominal conditions such as cholecystitis and diverticulitis. A systematic review of 29 studies on artificial intelligence in appendicitis highlighted wide heterogeneity in inputs, validation strategies and metrics. Imaging-based machine learning models reading CT scans have achieved sensitivities around 77 percent and specificities around 86 percent. The new study’s authors also point to broader evidence that data leakage has inflated reported performance in fields ranging from Parkinson’s disease detection to brain MRI classification, reinforcing their insistence on randomization as a guard against over-optimistic estimates.

The authors are careful about what their results do and do not license. Because the cohort consisted exclusively of surgically confirmed patients with high pre-test probability, the estimates cannot simply be generalized to the unselected emergency department population in whom appendicitis must be ruled in or out, and the study was limited to a single center, a single model and a single comparator, with newer scores like AIR and RIPASA omitted because the required data were not consistently available. ChatGPT-5, they conclude, cannot yet be relied upon to confirm acute appendicitis, given the Alvarado score’s higher specificity, and the findings should be regarded as hypothesis-generating pending prospective validation.

What the study does establish is a proof of concept with real clinical texture: a general-purpose language model, guided by a carefully engineered prompt and given nothing more than the routine clinical data already sitting in a patient’s chart, can match or exceed a century-old scoring heuristic at the task that matters most for safety, excluding the disease. The authors envision ChatGPT-5 as a clinical decision-support tool to assist professionals during early assessment, particularly in resource-limited settings, always under physician supervision and never as an autonomous diagnostician. They also warn explicitly against patient self-diagnosis, noting that false reassurance from a chatbot could delay timely medical evaluation. As large language models continue their march into medicine, this Egyptian study offers both an encouraging data point and a methodological warning: before trusting an AI’s diagnostic brilliance, check whether the cases were shuffled.

Subject of Research: Diagnostic accuracy of ChatGPT-5 compared with the Alvarado score for acute appendicitis

Article Title: Diagnostic performance of ChatGPT for acute appendicitis compared with the Alvarado score: a retrospective diagnostic accuracy study

Article References: Elgharib, A. A., Ghanem, M. I., Ibrahim, R. S., Elwakeel, N., & Shemes, A. (2026). Diagnostic performance of ChatGPT for acute appendicitis compared with the Alvarado score: a retrospective diagnostic accuracy study. Discover Artificial Intelligence, 6(1), Article 1352. https://doi.org/10.1007/s44163-026-02332-7

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02332-7

Keywords: ChatGPT-5, large language models, acute appendicitis, Alvarado score, diagnostic accuracy, artificial intelligence, clinical decision support, histopathology, order leakage, in-context learning, emergency medicine, prompt engineering