TRACE
A tree-structured relational augmentation framework for oncology large language models. It constructs updatable tree-structured relational representations offline and retrieves compact prompt evidence at inference time to improve interpretability and accuracy in cancer classification and medical question-answering tasks.
Timeline (last 90 days)
TRACE is a reproducible electricity price forecasting benchmark comprising 7,300 zone-day instances.
The TRACE benchmark data covers five zones in a major U.S. electricity market, pairing each zone's electricity prices with official operational texts available at the forecast cutoff time.
TRACE reconstructs official operational texts at each cutoff time, mechanistically preventing post-cutoff information leakage.
The research team evaluates TRACE along two dimensions: semantic alignment and forecasting value.
Semantic evaluation reveals that textual information is aligned with central shifts and bilateral tail risks in actual prices, with the alignment between text and upper-tail price risk being the most stable.
After incorporating textual information, the median upper-tail pinball loss across multiple time-series foundation models decreased by 7.4%.
In the cross-day text-mismatch ablation experiment, pairing texts with prices from incorrect dates caused the performance gains to disappear and forecasting performance to fall below the no-text baseline.
The cross-day text-mismatch ablation experiment rules out the possibility that 'text merely acts as regularization noise,' demonstrating that text only translates into forecasting value when correctly matched with prices at the corresponding time point.
Official operational texts should be treated as effective inputs for improving electricity price tail-risk forecasting capability, rather than optional supplementary information.
TRACE can produce interpretable evidence paths that align with clinical reasoning logic.
10 more timeline events
All Facts (20)
TRACE is a reproducible electricity price forecasting benchmark comprising 7,300 zone-day instances.
50%UnverifiedThe TRACE benchmark data covers five zones in a major U.S. electricity market, pairing each zone's electricity prices with official operational texts available at the forecast cutoff time.
50%UnverifiedTRACE reconstructs official operational texts at each cutoff time, mechanistically preventing post-cutoff information leakage.
50%UnverifiedThe research team evaluates TRACE along two dimensions: semantic alignment and forecasting value.
50%UnverifiedSemantic evaluation reveals that textual information is aligned with central shifts and bilateral tail risks in actual prices, with the alignment between text and upper-tail price risk being the most stable.
50%UnverifiedAfter incorporating textual information, the median upper-tail pinball loss across multiple time-series foundation models decreased by 7.4%.
50%UnverifiedIn the cross-day text-mismatch ablation experiment, pairing texts with prices from incorrect dates caused the performance gains to disappear and forecasting performance to fall below the no-text baseline.
50%UnverifiedThe cross-day text-mismatch ablation experiment rules out the possibility that 'text merely acts as regularization noise,' demonstrating that text only translates into forecasting value when correctly matched with prices at the corresponding time point.
50%UnverifiedOfficial operational texts should be treated as effective inputs for improving electricity price tail-risk forecasting capability, rather than optional supplementary information.
50%UnverifiedTRACE can produce interpretable evidence paths that align with clinical reasoning logic.
50%UnverifiedTRACE (Tree-Relational Structure Enhancement) is a framework for enhancing oncology large language models by supplementing structured medical support through a tree-relational structure.
50%UnverifiedAs a newly released arXiv preprint, TRACE's conclusions still require broader clinical validation and peer review.
50%UnverifiedThe core design philosophy of TRACE is to separate expensive offline structure learning from lightweight online inference.
50%UnverifiedTRACE organizes oncology concepts and relationships into an updatable tree-relational structure, and refines it using LM-loss-derived evidence.
50%UnverifiedDuring the inference stage, TRACE retrieves compact evidential prompts from the structure and feeds them to the model as additional information.
50%UnverifiedTRACE's evidence selection mechanism supports task adaptation and requires no supervision labels in a zero-shot setting.
50%UnverifiedTRACE was evaluated on ten oncology classification tasks and one MedQuAD CancerGov question-answering benchmark.
50%UnverifiedTRACE brought performance improvements in both unlabeled evaluation and supervised fine-tuning modes.
50%UnverifiedTRACE outperforms vanilla RAG (Retrieval-Augmented Generation) and general-purpose GraphRAG.
50%UnverifiedUnder leakage-controlled METABRIC input conditions, TRACE remains effective.
50%