The study, published in Natural Resources Research, was led by Changkui Lei and Qi Qiao of Taiyuan University of Technology, together with Jun Deng and Jingyu Zhao of Xi’an University of Science and Technology, Chuanbo Cui, and Weigang Wang of the Gansu Bureau of the National Mine Safety Administration. Their central innovation is a hybrid model that pairs support vector regression, a powerful but notoriously parameter-sensitive prediction method, with the grey wolf optimization algorithm, a metaheuristic that mimics the leadership hierarchy and hunting strategy of wolf packs. The result, dubbed GWO-SVR, achieved a coefficient of determination of 0.9902 on test samples, meaning the model explained more than 99 percent of the variance in coal temperature during oxidation experiments.

To understand why this matters, it helps to grasp the physics of spontaneous combustion. When coal is exposed to oxygen at ambient temperatures, it oxidizes slowly, releasing heat. In loose coal piles, goaf areas behind longwall mining faces, and stockpiles, that heat can accumulate faster than it dissipates. As temperature rises, oxidation accelerates exponentially, creating a runaway feedback loop. The researchers’ large-scale coal oxidation experiment, using samples from the Paner Coal Mine, traced this process in detail. They observed that the high-temperature point inside their experimental furnace did not stay put: it migrated dynamically from the middle-upper section toward the lower section, eventually settling at the air inlet position, where fresh oxygen continuously fed the reaction.

The experimental campaign also quantified the characteristic indicators that signal a developing fire. The oxygen consumption rate, the generation rates of indicator gases, and the exothermic strength of the oxidation reaction all followed exponential growth trends as temperature climbed. These gas signatures, including carbon monoxide and other products of low-temperature oxidation, are what mine safety engineers monitor in the field. The challenge has always been translating gas concentrations measured in a mine roadway back into an accurate estimate of the coal temperature deep inside a goaf, where no thermometer can reach. That inverse problem is exactly where machine learning excels, provided the model’s internal parameters are tuned correctly.

This is where the grey wolf enters the story. Support vector regression depends critically on hyperparameters, such as the penalty factor and kernel settings, that govern how it fits nonlinear relationships in data. Set poorly, the model underfits or overfits; set well, it generalizes beautifully. Traditionally, engineers have relied on trial and error or grid searches. The grey wolf optimization algorithm instead treats the parameter search as a simulated hunt: candidate solutions are ranked as alpha, beta, delta, and omega wolves, and the pack iteratively encircles and converges on the optimal parameter combination. The researchers found that this optimization step was decisive. On the experimental test set, the root mean square error of the GWO-SVR model dropped to 2.4725, while a grey wolf-optimized back propagation neural network achieved 2.9024, both substantially better than standalone SVR and BPNN models.

Laboratory results alone rarely convince mining engineers, so the team took a further step that distinguishes this work: field verification. They validated the GWO-SVR model against in situ monitoring data from the Sanhejian Coal Mine, a real underground environment with all the messiness that entails, including variable airflow, moisture, and heterogeneous coal distributions. Under these genuine field conditions, the GWO-SVR model posted a root mean square error of just 0.9310 degrees in its temperature predictions, compared with 1.7778 for standalone SVR, 2.5469 for the BPNN, and 1.2791 for the GWO-BPNN. The comparison underscores that the optimization algorithm, not merely the choice of base model, drives the performance gain.

The implications for mine safety are significant. Spontaneous combustion fires in goaf areas are extraordinarily difficult to detect early because they develop out of sight, behind sealed zones, and by the time smoke or elevated carbon monoxide reaches sensors, the fire may already be well established. A model that can convert routinely measured indicator gas concentrations into a reliable temperature estimate gives fire prevention teams a quantitative gauge of how close the coal is to critical stages, allowing targeted interventions such as nitrogen injection, grouting, or adjusted ventilation before conditions become dangerous. The authors’ earlier work, including comparisons of random forest and support vector machine approaches and studies of high-temperature point migration, laid the groundwork for this refined approach.

The study also situates itself within a broader scientific effort to tame coal’s reactivity. Recent research has explored biomass aerogels that inhibit combustion at the microstructural level, shape memory hydrogels and plastogels for fire prevention, and thermokinetic analyses of low-rank coals during low-temperature oxidation. Other teams have applied genetic algorithm-optimized SVR to predict coal temperature from carbon monoxide and used back propagation networks for spatio-temporal temperature prediction. What the new study adds is a rigorous, experimentally grounded pipeline: a large-scale oxidation experiment to characterize the indicators, a hybrid model whose hyperparameters are intelligently optimized, and validation at both laboratory and field scales, a combination that few previous studies have achieved end to end.

From a technical standpoint, the exponential growth patterns observed in oxygen consumption, gas generation, and exothermic strength explain why linear or simple empirical models have historically struggled. The relationship between indicator gases and coal temperature is strongly nonlinear, with different gases dominating different temperature ranges. Machine learning models can capture these nonlinearities, but only if trained on representative data and tuned with care. The GWO-SVR framework addresses both requirements, and its strong generalization performance on unseen test samples suggests it is not merely memorizing the experimental data but learning the underlying physics of coal oxidation.

For an industry still central to global energy supply, and for the hundreds of thousands of miners who work underground every day, tools like this represent a quiet but meaningful advance in the long fight against one of mining’s oldest hazards. The research was supported by the National Natural Science Foundation of China and the Basic Research Program of Shanxi Province. As mines push deeper and coal seams become more prone to self-heating, the ability to forecast the invisible heat before it becomes an inferno may prove not just scientifically elegant but genuinely lifesaving.

Subject of Research: Machine learning prediction of coal spontaneous combustion temperature using grey wolf optimization and support vector regression

Article Title: Enhanced Prediction of Coal Spontaneous Combustion Temperature via GWO-SVR with Experimental and Field Verification

Article References: Lei, C., Qiao, Q., Deng, J., Zhao, J., Cui, C., & Wang, W. (2026). Enhanced Prediction of Coal Spontaneous Combustion Temperature via GWO-SVR with Experimental and Field Verification. Natural Resources Research. https://doi.org/10.1007/s11053-026-10779-9

Image Credits: AI Generated

DOI: 10.1007/s11053-026-10779-9

Keywords: coal spontaneous combustion, grey wolf optimization, support vector regression, machine learning, mine safety, indicator gases, temperature prediction, coal oxidation, nonlinear prediction, field verification, Enhanced, Prediction

The work arrives at a moment when the AI industry is pouring enormous resources into so-called agentic systems—models that autonomously plan tasks and invoke software tools on the user’s behalf. Tool invocation sits at the heart of this vision. In the standard tool-learning pipeline, which researchers typically divide into task planning, tool selection, tool invocation and response generation, the invocation stage is the make-or-break moment. The model must extract parameters from a natural-language query, match them to a tool’s specification, and emit a request so precisely structured that a downstream server can parse it without error. Any stray character, a missing parenthesis, or a misplaced comma can cause the entire call to fail silently.

What the researchers discovered through systematic perturbation experiments is striking: the success of a tool call is far more sensitive to format standardization than to semantic accuracy. When they fine-tuned the Llama3.1-8B-Instruct model on the ToolACE dataset using LoRA, randomly altering numbers in the training labels left accuracy nearly untouched, and shuffling parameter strings produced only a modest decline. But when they changed the format itself—swapping bracket types, converting integers to floating-point numbers, or reordering parameters—performance collapsed. Simply changing bracket styles dragged live-task accuracy down to 42.51 percent from a much higher baseline. Converting numbers to floats proved most devastating of all, with one metric plunging to 25.39 percent, because the abstract syntax tree evaluation used by the benchmark flags data-type mismatches instantly.

The most dramatic result came from compounding these perturbations. In a double mixed-modification experiment that first randomized bracket usage and then converted numbers to floats, live accuracy cratered to just 3.02 percent—essentially total failure. The lesson, the authors argue, is that conventional fine-tuning creates what they call format fragility: models rigidly cling to whatever format patterns they saw in training data, and even when prompts explicitly specify an output format, fine-tuned models frequently ignore those instructions and emit unparsable output. Earlier studies have described this phenomenon as format specialization or task locking, where intense fine-tuning erodes a model’s general in-context learning ability on non-target tasks.

Recognizing that reasoning and formatting are fundamentally different skills, the team designed a division-of-labor architecture that separates them. In their collaborative framework, the large language model receives the user’s question and a list of available tools, then produces an intermediate output containing its thought process, the selected tool name and the parameter information. Crucially, this intermediate output need not follow any format at all—the large model is freed from worrying about syntax. That freedom is precisely what preserves its generalization. The intermediate result is then handed to a small, specialized format model whose sole job is to normalize it into a strict, predefined structure that can be parsed directly into a callable API request.

The experimental payoff was substantial. When the same perturbed models were paired with the formatting model, accuracy rebounded dramatically. The Random Mix Twice configuration, which had fallen to 3.02 percent, soared to 73.42 percent once the small model normalized the output. The formatting step effectively absorbs all the chaotic variations—missing parentheses, wrong number formats, unexpected parameter orders—that would otherwise doom the invocation. The authors also contrast their approach with in-context learning, noting that few-shot examples struggle to exhaustively cover complex, nested parameter schemas, whereas a dedicated format model explicitly models the output structure and separates tool selection from argument generation.

The study used the Berkeley Function Call Leaderboard, a benchmark of more than 1,700 instances spanning simple, multiple, parallel and parallel-multiple function calls in Python, as well as REST API, JavaScript and Java tasks. Evaluation relied on the benchmark’s abstract syntax tree methodology, which checks whether function names, required parameters and data types all conform to the function documentation. The experiments ran on a single RTX 4090 GPU, underscoring that the collaborative method is computationally modest: instead of retraining a giant model, it attaches a lightweight normalizer to the end of the pipeline.

The implications reach across the AI industry. Giants including IBM’s Granite-20B-FunctionCalling, ToolLLM, APIGen and ToolACE have all pursued better function-calling models through increasingly sophisticated fine-tuning and dataset synthesis. But the new study suggests a quiet vulnerability running through that entire paradigm: as long as a single model is asked to be both reasoner and formatter, it will remain brittle. A wrong parameter value may still pass parsing and merely yield an irrelevant result, but a wrong bracket is fatal. The finding that format correctness outranks content accuracy inverts a common assumption that semantic quality is the primary axis of model quality, and it offers a practical, modular fix that developers could retrofit onto existing systems without touching the underlying model weights.

The authors are candid about the limits of their approach. Their evaluation remains confined to static, single-turn settings on specific test sets, while real-world agents must handle multi-turn dialogues that demand consistent tracking of context and parameters across turns. Real APIs also evolve their schemas over time, and format learning grounded in fixed training data struggles to adapt. Yet the multi-model framework points toward a natural solution: the large model can continue to handle contextual reasoning and intent understanding while the small model operates as a lightweight, updatable formatter that maps intent to whatever the current API schema requires. The team plans to integrate schema-based validation checkers and explore online adaptation techniques, including few-shot in-context learning and parameter-efficient fine-tuning, to keep inference costs low while boosting robustness in dynamic deployments.

For a field fixated on scale, the takeaway is refreshingly counterintuitive. Sometimes the fastest way to make a giant AI smarter is to pair it with a tiny, single-minded helper obsessed with punctuation. By decoupling what a model knows from how it says it, the study reframes tool learning as a coordination problem rather than a capability problem—and suggests that the next leap in autonomous AI agents may come not from bigger brains, but from better teamwork.

Subject of Research: Enhanced tool invocation method through multi-model collaboration

Article Title: Enhanced tool invocation method through multi-model collaboration

Article References: Enhanced tool invocation method through multi-model collaboration. (n.d.). https://doi.org/10.1007/s44336-025-00028-7

Image Credits: AI Generated

DOI: 10.1007/s44336-025-00028-7

Keywords: Enhanced, tool, invocation, method, multi-model, collaboration, scientific research

Photodynamic therapy, or PDT, works by delivering a photosensitizer, a compound that is otherwise non-toxic, into tumor tissue and then activating it with light of a specific wavelength. Upon illumination, the photosensitizer transfers energy to surrounding oxygen molecules, generating reactive oxygen species and singlet oxygen that destroy cells. Two reaction types are recognized: a Type I pathway that produces peroxides, hydroxyl radicals and superoxide ions through electron transfer cascades, and a Type II pathway that generates highly reactive singlet oxygen through direct energy transfer to molecular oxygen. Beyond direct tumor cell killing, PDT also damages tumor vasculature and triggers an acute inflammatory response that can activate both innate and adaptive immunity, with immunogenic cell death releasing damage-associated molecular patterns that further stimulate the immune system.

Despite this attractive mechanism, PDT has struggled to become a standard treatment for bladder cancer. Earlier photosensitizers such as photofrin and hematoporphyrin derivatives caused unwanted skin photosensitization when given systemically, while local application led to bladder wall fibrosis because of the high light intensities required. Perhaps the most fundamental limitation has been optical: conventional photosensitizers are activated at wavelengths of 630 to 690 nanometers, where light penetrates tissue to a depth of only about five millimeters, making it difficult to treat thicker, muscle-invasive tumors. The Leipzig team turned to tetrahydroporphyrin-tetratosylate, or THPTS, a water-soluble, positively charged near-infrared photosensitizer with an absorption maximum at 760 nanometers, a wavelength that allows tissue penetration of up to fifteen millimeters and therefore reaches tumors that earlier agents could not.

Before moving into animals, the researchers characterized the cellular effects of the combined approach in vitro. They cultured AY-27 rat bladder carcinoma cells, a line originally derived from a carcinogen-induced bladder tumor in Fischer rats, both as flat two-dimensional cultures and as three-dimensional spheroids that mixed tumor cells with primary rat bladder fibroblasts. These spheroids self-organized into bladder-like structures with an outer tumor cell layer surrounding an inner fibroblast core, better mimicking the architecture of real tumors. Treatment involved a single 4 Gy dose of X-ray irradiation followed one hour later by incubation with THPTS at concentrations ranging from 6.25 to 50 micromolar and illumination at 760 nanometers with a light dose of 10 joules per square centimeter.

The two culture formats told strikingly different stories. In two-dimensional cultures, the combination reduced cell viability by up to 90 percent after 72 hours, but response additivity analysis revealed the effects were subadditive rather than synergistic. In the three-dimensional spheroids, however, the combined treatment produced genuine synergism at nearly all tested concentrations, reducing metabolic activity by 39 percent after 72 hours while single therapies alone showed no significant effect. Mechanistic staining showed that PDT alone drove a pronounced increase in 4-hydroxynonenal, a marker of oxidative stress, while radiation elevated phospho-histone H2A.X, a marker of DNA damage; the combination strongly elevated both markers, particularly in the tumor cell layer. Crucially, the treatment significantly thinned the malignant outer layer of the spheroids while leaving the non-malignant stromal core untouched, underscoring the tumor selectivity of the approach. The authors note that the greater resistance of three-dimensional cultures to therapy makes the observed synergism there especially meaningful for predicting in vivo behavior.

The in vivo work used the well-characterized orthotopic AY-27 model, in which bladder tumors are induced in female F344 Fischer rats by intravesical instillation of tumor cells, a technique that had previously achieved a 100 percent tumor induction rate in the group’s hands. A preliminary study delivered a sobering lesson: a single session of PDT unexpectedly shortened survival compared with untreated controls, likely because a single illumination rapidly depletes oxygen in the tumor, reducing reactive oxygen species generation and leaving residual tumor cells to activate survival pathways. This finding echoed literature showing that fractionated PDT regimens improve long-term tumor control, and it motivated the team to design a repeated-treatment protocol for the main trial.

In the main experiment, 40 rats with established bladder tumors were divided into four groups of ten: untreated controls, PDT alone, radiation alone, and radiation followed by PDT. Treatments were given three times at seven-day intervals, starting 14 days after tumor inoculation. Each session involved irradiation of the lower abdomen with 8.5 Gy from an orthovoltage X-ray source, immediately followed by transurethral instillation of 100 micromolar THPTS into the bladder for two hours and then laser illumination at 760 nanometers through a glass fiber. The radiation schedule was carefully calibrated so that its biologically effective dose on late-responding normal bladder tissue matched that of current human bladder cancer radiotherapy regimens, while leaving room for additional antitumor effect from the PDT component.

The survival results were remarkable. Rats receiving three cycles of PDT alone showed a doubling of median overall survival to 70 days compared with 36.5 days in untreated controls, a difference the authors believe is the first reported 100 percent survival increase from local repeated PDT monotherapy in bladder cancer in vivo. Even more striking, neither the radiation-only group nor the combination group reached the study’s termination criteria during the observation period, meaning every animal in those cohorts survived without clinical symptoms. Histopathology added crucial nuance: while radiation alone left three of eight bladders with viable tumors, the combination therapy produced a treatment response in every tumor examined, achieving complete regression in seven of eight rats, with only one residual tumor measuring 1817 micrometers at its base compared with tumors exceeding 3100 micrometers after radiation alone.

Immunohistochemistry for CD45-positive leukocytes revealed that PDT triggered a significant local immune response, with increased leukocyte accumulation at both the invasive tumor margin and the tumor center compared with untreated controls. In the combination-treated animals, large numbers of strongly CD45-positive immune cells were observed beneath tumor regression zones, though statistical comparison was impossible because so many tumors had regressed completely. The authors suggest that pairing this multimodal focal therapy with immune checkpoint inhibitors could further amplify its inherent immunogenic effects, and they point toward future refinements including systemic photosensitizer delivery to reach deeper tumor regions, alternative excitation sources such as X-ray-activated scintillating nanoparticles, and photosensitizers favoring less oxygen-dependent Type I reactions to overcome the rapid oxygen depletion that limits solid tumor treatment.

The study is not without limitations. The biologically effective dose calculations relied on reference alpha-beta ratios from the literature rather than values determined specifically for AY-27 cells or rat bladder tissue, and the trial lacked a control group of tumor-free rats treated with the combination to assess off-target toxicity. Roughly half of the bladders showing complete tumor regression displayed moderate histopathological changes, including urothelial damage, edema, fibrosis and inflammation. Nevertheless, the treatments were overall well tolerated, with 33 of 40 animals completing the protocol as planned. If the results translate to the clinic, the authors propose the approach could benefit patients with high-grade non-muscle-invasive bladder cancer, for whom bladder removal is currently recommended, and potentially even those with non-metastatic muscle-invasive tumors up to fifteen millimeters thick, offering a genuinely organ-preserving alternative built on two technologies, near-infrared light and fractionated radiation, that are already mainstays of modern oncology.

Subject of Research: Combined photodynamic therapy and ionizing radiation for organ-preserving treatment of bladder cancer in an orthotopic rat model

Article Title: Enhanced survival through repeated photodynamic therapy and almost complete tumor regression by prior radiation therapy in an orthotopic rat bladder cancer model

Article References: Berndt-Paetz, M., Nürnberger, S., Gonsior, S., Pączek-Hippe, E., Patties, I., Weimann, A., Michalik, R., Neuhaus, J., & Glasow, A. (2026). Enhanced survival through repeated photodynamic therapy and almost complete tumor regression by prior radiation therapy in an orthotopic rat bladder cancer model. Cancer Cell International, 26(1), Article 315. https://doi.org/10.1186/s12935-026-04465-2

Image Credits: AI Generated

DOI: 10.1186/s12935-026-04465-2

Keywords: photodynamic therapy, THPTS, bladder cancer, ionizing radiation, near-infrared photosensitizer, orthotopic rat model, tumor regression, synergistic cytotoxicity, organ-preserving therapy, immune cell recruitment, Enhanced, survival