manulife/tam-benchmarks

Tasks over Application Manuals (TAM)

TAM is a benchmark for evaluating long-horizon procedural reasoning: the ability of a language-model system to follow a large application manual, resolve cross-references, apply interdependent constraints, and produce an exact answer. Unlike short-horizon multi-hop tasks, TAM requires systems to maintain consistency across dozens of decisions drawn from manuals containing tens of thousands of rules. An early missed exception or incorrect commitment can invalidate the final result even when individual intermediate steps appear plausible.

TAM contains two real-world tasks with human-validated labels:

Track	Dataset size	Authoritative manuals	Output	Availability
ICD-10-CM coding	1,000 cases sampled from 38,332 eligible encounters	2019 ICD-10-CM Guidelines, 1,304-page Alphabetic Index, and 1,942-page Tabular List	Diagnosis code set and primary diagnosis	Manuals included; cases reconstructed locally by authorized PhysioNet users
Federal sentencing	200 approved cases after human review	Five annual Title 18 archives and five annual USSG manuals, 2021-2025	Total offense level	Cases and manuals included

Each instance combines a case description, an application manual, and an exact target derived by applying the manual. Solving an instance may require retrieving rules, resolving cross-references, tracking active obligations, applying ordering constraints, revising earlier decisions, and backtracking before termination.

The legal track and the public application manuals for both tracks are distributed directly in this repository. MIMIC source data and reconstructed patient records are not redistributed; the ICD track is provided as a reproducible cohort builder, public curation inputs, and documentation. The paper's 1,000-case ICD dataset was sampled from the eligible cohort rather than distributed here.

Baselines

The paper evaluates four general-purpose approaches on GPT-5: single-pass retrieval-augmented generation, agentic RAG, ReAct-style tool use, and an agent-harness baseline. These are initial reference baselines rather than an exhaustive comparison of specialized methods. Exact match remains at or below 1% for ICD-10-CM coding and reaches at most 15.5% for federal sentencing, exposing a substantial gap in reliable end-to-end manual execution.

Reported Results

Baseline	ICD EM	ICD Precision	ICD Recall	ICD Prim. Dx Acc.	Legal EM	Legal MAE
Single-pass RAG	1%	52%	33%	42.7%	7.5%	2.87
Agentic RAG	1%	59%	37%	45.7%	11%	2.80
ReAct-style tool use	1%	58.88%	26.82%	20%	15.5%	3.04
Agent harness	0.7%	59.9%	27.4%	50.1%	14.5%	2.34

Metric Definitions

For ICD evaluation, predicted and reference codes are normalized and treated as sets for exact match, precision, and recall. These metrics ignore code order; ordering is evaluated separately through the primary diagnosis.

Metric	Meaning	Direction
ICD EM	Exact set match: the percentage of cases where the predicted ICD-10-CM code set is identical to the reference set. Any missing or extra code makes the case incorrect.	Higher is better
ICD Precision	For each case, the number of correctly predicted codes divided by the number of predicted codes, averaged across cases.	Higher is better
ICD Recall	For each case, the number of correctly predicted codes divided by the number of reference codes, averaged across cases.	Higher is better
ICD Prim. Dx Acc.	Primary diagnosis accuracy: the percentage of cases where the first normalized predicted code equals the first reference code in diagnosis sequence order.	Higher is better
Legal EM	Exact match: the percentage of cases where the predicted integer offense level equals the approved offense level.	Higher is better
Legal MAE	Mean absolute error: the average absolute difference between the predicted and approved offense levels, measured in offense-level points.	Lower is better

Included Manuals

The complete public manual bundle used by TAM is available under manuals/:

• ICD-10-CM: the 2019 Official Guidelines for Coding and Reporting, Alphabetic Index, Tabular List, External Cause Index, Table of Drugs and Chemicals, and Neoplasm Table. PDF and structured XML versions are included where published.
• Federal sentencing: annual Title 18 HTML archives and U.S. Sentencing Guidelines PDF manuals for 2021 through 2025.

The ICD cohort-filtering addenda for 2018-2020 are included with the reconstruction code under icd/curation/addenda/.

Legal Track

The hosted legal dataset contains 200 human-reviewed cases. Cases draw on federal sentencing materials from 2021-2025, including annual Title 18 archives and U.S. Sentencing Guidelines manuals.

Field	Type	Description
docket_id	string	Case identifier used by the evaluation pipeline.
input	string	Case-fact text presented to the evaluated system.
output	integer	Approved total offense-level target.

Targets range from 2 through 44. The dataset contains no missing inputs, missing targets, or duplicate docket identifiers. The primary metrics are exact-match accuracy and mean absolute error in offense-level points. Additional details are available in legal/README.md.

ICD Track

The ICD task maps a full discharge note to normalized ICD-10-CM diagnosis codes ordered by source seq_num. Evaluation uses exact code-set match regardless of order and reports primary-diagnosis accuracy separately. The benchmark contains 1,000 cases sampled from 38,332 eligible encounters. These are MIMIC admissions that have ICD-10 diagnoses and a discharge note, fall within the strict reconstructed 2017-2019 date window, and contain no diagnosis code changed by the 2018-2020 addenda. The task is grounded in the 2019 ICD-10-CM Guidelines, Alphabetic Index, and Tabular List.

Authorized users reconstruct one final CSV from local MIMIC-IV and MIMIC-IV-Note files with the standard-library builder in icd/curation/. The repository contains no MIMIC rows and intentionally exposes no hosted ICD split. The complete protocol and validation criteria are documented in icd/CURATION.md.

The ICD metrics are exact set match, mean per-case precision, mean per-case recall, and primary diagnosis accuracy.

Repository Contents

• data/legal/dataset.csv: hosted 200-case legal dataset.
• manuals/: public ICD-10-CM, Title 18, and U.S. Sentencing Guidelines manuals used by the benchmark.
• icd/curation/: standalone ICD reconstruction script and public 2018-2020 ICD-10-CM addenda.
• icd/CURATION.md: ICD access, reconstruction, schema, and validation documentation.
• results/paper_results.csv: combined machine-readable paper results.

Paper and Citation

Tasks over Application Manuals: Revealing Gaps in Long-Horizon Procedural Reasoning for Language Models
Utkarsh Soni, Syed Shariyar Murtaza, Yifan Nie, Sachin Chandrasekhar, and Eugene Wen
35th ACM International Conference on Information and Knowledge Management (CIKM '26), Rome, Italy

@inproceedings{soni2026tasks,
  author = {Soni, Utkarsh and Murtaza, Syed Shariyar and Nie, Yifan and Chandrasekhar, Sachin and Wen, Eugene},
  title = {Tasks over Application Manuals: Revealing Gaps in Long-Horizon Procedural Reasoning for Language Models},
  booktitle = {Proceedings of the 35th ACM International Conference on Information and Knowledge Management},
  year = {2026},
  address = {Rome, Italy},
  publisher = {Association for Computing Machinery},
  isbn = {979-8-4007-2539-5},
}

License

This work is licensed under the Creative Commons Attribution 4.0 International License. Copyright 2026 held by the owner/author(s).
ACM ISBN 979-8-4007-2539-5/2026/11