awni00/multi-strategy-algorithmic-tasks

Multi-Strategy Algorithmic Tasks

A synthetic benchmark of parseable algorithmic problems with multiple valid solution strategies for each task. Each example contains a problem,a strategy-specific solution trace, and the strategy used to generate that trace.

The benchmark accompanies Uncovering Latent Reasoning Strategies in Language Models, which studies the problem of recovering mixtures of strategies implicitly represented in language models. The benchmark provides a controlled setting for studying strategy recovery, representation, routing, and controllable generation.

Load the dataset

from datasets import load_dataset

dataset = load_dataset("awni00/multi-strategy-algorithmic-tasks")
example = dataset["train"][0]

The release contains:

Split	Rows
train	1,000,000
validation	10,000
test	10,000

Release version: v1.0.0.

To select one task family:

sorting = dataset.filter(
    lambda example: example["task_name"] == "sorting_algorithms"
)

Fields

Field	Description
task_name	Algorithmic task family
strategy_id	Namespaced strategy sampled to generate the trace
input_text	Rendered problem instance
reasoning_trace	Complete strategy-specific solution trace, including the final answer

Example:

{
    "task_name": "multidigit_addition",
    "strategy_id": "multidigit_addition:left-to-right-partials",
    "input_text": "560+342",
    "reasoning_trace": (
        "p100:500+300=800 ; p10:60+40=100 ; "
        "p1:0+2=2 ; sum=800+100+2=902"
    ),
}

Tasks and strategies

The generator first samples one of the six task families uniformly. It then samples a strategy uniformly within that task family.

Task	Problem	Strategies
list_summation	Sum four integers from 0 to 9	left-to-right, right-to-left, pairwise
sorting_algorithms	Sort five integers from 0 to 9	bubble-sort, selection-sort, insertion-sort, merge-sort, heap-sort
grid_pathfinding	Monotone shortest paths on a 6×6 grid	right-first, down-first, alternating
linear_equation_solving	Solve integer equations of the form ax+b=c	subtract-then-divide, divide-then-subtract, inverse-ops
base_conversion	Convert integers from 1 to 255 to base 2, 4, 8, or 16	repeated-division, via-binary, decomposition
multidigit_addition	Add two three-digit nonnegative integers	right-to-left-carry, left-to-right-partials, rounding-decomposition

The public strategy_id includes the task namespace, for example grid_pathfinding:alternating. There are 20 strategies in total across the six tasks.

strategy_id is the strategy selected by the generator. On some inputs, multiple strategies produce the same observable trace. This is most common when a problem requires only a few steps or different algorithms happen to traverse identical intermediate states.

License

The dataset is released under the Creative Commons Attribution 4.0 International License.

Citation

@misc{altabaa2026uncovering,
  title         = {Uncovering Latent Reasoning Strategies in Language Models},
  author        = {Awni Altabaa and John Lafferty},
  year          = {2026},
  eprint        = {2607.17674},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
  url           = {https://arxiv.org/abs/2607.17674}
}

Code associated with the paper is available at Awni00/latent-strategies-in-lms.