A preprint reports that D-TAIA, a retrieval-assisted language-model system, matched or improved on a fine-tuned large language model and a recurrent neural network when it predicted both the next activity and the remaining time in an event log. Its clearest margins appeared with short prefixes, meaning only a small part of a case was available, and on the log with the highest process entropy, or the most varied possible paths. The finding comes from model benchmarking across event logs.
One model, two forecasts
D-TAIA is built around one model making both forecasts from an event-log prefix, or the sequence available so far. It combines parameter-efficient LLM fine-tuning using LoRA with domain-aware triplet-loss pre-training, FAISS nearest-neighbor retrieval for the time estimate, and TAIA inference intended to preserve sequential reasoning under distributional shift. In plain terms, it uses the model's learned patterns alongside similar prefixes retrieved from the data.
Four logs, one fixed timeline
The evaluation covered four publicly available BPI Challenge 3 event logs. BPI2012 had 13,087 cases, BPI2017 had 31,509, BPI2015_2 had 832 and BPI2020_DD had 10,500. Together, the logs span different sizes and levels of process entropy, or variation in the possible paths through a case.
Researchers reported three measures: Macro-F1 for next-activity prediction, mean absolute error, or MAE, for the size of the remaining-time mistake in days, and wall-clock runtime in hours. Results were averaged over five random seeds and presented with 95% confidence intervals, giving readers a sense of variation across runs.
To preserve the order in which cases began, the train, validation and test sets were split by case-start time at 65%, 15% and 20%. Cases remained separate across the split, and the division was fixed across seeds; changing the seed altered only model initialization and mini-batch order.
The clearest edge came early
The main comparator, FT-LLM, used direct LoRA fine-tuning with task-specific tokenization but left out D-TAIA's domain-aware pre-training, retrieval and TAIA inference. The MT-RNN baseline was trained from scratch. Against these baselines, D-TAIA generally produced lower remaining-time error at comparable runtime and a higher Macro-F1 score.
The margin was not equally decisive everywhere. Macro-F1 confidence intervals overlapped in individual comparisons. For MAE, the intervals did not overlap on BPI2012 and BPI2020_DD, but did overlap on BPI2017 and BPI2015_2. Overall, D-TAIA matched or improved on both baselines, with its clearest margins on short prefixes and the highest-entropy log.
Input length mattered. In the shortest prefix bucket, bucket 1, D-TAIA's Macro-F1 advantage over FT-LLM ranged from 3.0 to 6.0 percentage points. The gap narrowed toward bucket 5, the longest bucket, and remaining-time error followed the same pattern. Even at bucket 5, the residual gap was proportionally largest on BPI2015_2.
Data scarcity produced a similar but less dramatic picture. Across all four logs and every training-data budget reported, D-TAIA had the best point estimate for both Macro-F1 and remaining-time MAE. At the 20% budget, however, the confidence intervals for all three models overlapped on each log. The clearest advantage appeared at full data on high-entropy BPI2015_2.
Model size also mattered most on the highest-entropy log. The backbone test compared Tiny-LLM at 10 million parameters, Qwen2.5 at 500 million and Llama3.2 at 1 billion. On BPI2015_2, Macro-F1 rose by 13.7% from Tiny-LLM to Llama3.2, compared with relative gains of 3% to 6% on the other logs. The paper presents that pattern as a suggestion that capacity mattered most where process entropy was highest, not as a causal estimate.
One component stood out
The component tests offered a clue about the result. On BPI2015_2, full D-TAIA had a Macro-F1 of 0.445 and a remaining-time MAE of 53.85 days. Removing DATL lowered Macro-F1 to 0.352 and raised MAE to 68.45 days. Removing Domain ID produced 0.385 and 58.22 days; removing FAISS produced 0.418 and 61.85 days; removing TAIA produced 0.425 and 55.28 days. Every removal worsened the point estimates, with the largest losses after DATL was removed.
On BPI2020_DD, full D-TAIA scored 0.811 Macro-F1 with a 6.52-day MAE. The corresponding figures after removing DATL, Domain ID, FAISS and TAIA were 0.783 and 7.85 days, 0.790 and 6.92 days, 0.807 and 6.68 days, and 0.806 and 6.72 days. DATL removal was the only variant with a distinguishable 95% confidence interval on both high-entropy metrics and on low-entropy MAE.
The missing test
But the DATL result is not a clean isolation of that component. The ablation kept FAISS retrieval while using a randomly initialized encoder, so the retrieval was partly against noise. The authors therefore treat the result cautiously: DATL removal had the largest point-estimate drop, but the comparison is confounded by the retrieval setup. TAIA was designed for distributional shift, yet the study did not directly evaluate it under that condition.
The evidence remains a four-log benchmark. It reports results from the tested language-model and RNN baselines and backbones, with the strongest signal when the observed prefix is short. The manuscript is identified as an arXiv preprint, so the result is best read as an early model evaluation.
Paper data and sources
Original title: D-TAIA: Domain-Aware LLM Adaptation for Multi-Task Predictive Process Monitoring
Authors: Sjoerd van Straten, Christine Jacob, Marwan Hassani
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text