Update README.md

README.md

@@ -17,15 +17,15 @@ size_categories:
 
 # Multi-Strategy Algorithmic Tasks
 
-Multi-Strategy Algorithmic Tasks is a synthetic benchmark of parseable
-algorithmic problems with multiple valid solution strategies. Each example
-contains a problem, a complete strategy-specific solution trace, and the
-strategy used to generate that trace.
+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](https://arxiv.org/abs/2607.17674).
-It provides controlled reference strategy labels for studying strategy
-recovery, representation, routing, and controllable generation.
+*[Uncovering Latent Reasoning Strategies in Language Models](https://arxiv.org/abs/2607.17674)*,
+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
 
@@ -56,8 +56,6 @@ sorting = dataset.filter(
 
 ## Fields
 
-All four fields are strings.
-
 | Field | Description |
 | --- | --- |
 | `task_name` | Algorithmic task family |
@@ -65,9 +63,6 @@ All four fields are strings.
 | `input_text` | Rendered problem instance |
 | `reasoning_trace` | Complete strategy-specific solution trace, including the final answer |
 
-The dataset intentionally contains text rather than token IDs. Users can apply
-the tokenizer and sequence framing appropriate for their own model.
-
 Example:
 
 ```python
@@ -97,70 +92,13 @@ samples a strategy uniformly within that task family.
 | `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 namespaced strategies in total.
-
-## Generation and reproducibility
-
-The published rows are a text rendering of the frozen aggregate dataset used
-for the accompanying paper. The release preserves the original examples,
-split membership, and row order. It does not resample or filter the data.
-
-The task family, input, and strategy are sampled as:
-
-\[
-T \sim \operatorname{Unif}(\text{task families}), \qquad
-X \sim \mathcal{D}_T, \qquad
-S \sim \operatorname{Unif}(\mathcal{S}_T), \qquad
-Y = \operatorname{Trace}_T(X,S).
-\]
-
-The repository includes `generation_config.yaml` with the resolved task
-parameters and random seeds. `release_manifest.json` records source and
-release checksums. `validation_report.json` records task and strategy counts,
-trace ambiguity, duplicates, and cross-split overlap.
-
-## Important characteristics
-
-### Sampled strategy labels
+`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. Such a trace is valid
-but its sampled strategy may not be uniquely recoverable from the trace alone.
+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.
 
-### Sampling with replacement
-
-Examples are sampled with replacement from finite task distributions.
-Repeated examples within a split and overlap between splits are therefore
-expected. The splits represent independent samples from the same controlled
-distribution; they are not disjoint-input generalization splits.
-
-## Intended use
-
-The dataset is intended for controlled research on:
-
-- learning and evaluating multiple reasoning strategies;
-- strategy recovery from generated traces or model representations;
-- latent-variable routing and controllable generation;
-- algorithmic sequence modeling with parseable outputs.
-
-The reference strategy labels are intended for evaluation and analysis. In the
-paper's main setup, they are not provided as supervision to the language model.
-
-## Limitations
-
-- The tasks are synthetic, symbolic, and deliberately small.
-- Difficulty ranges are fixed by the published generation configuration.
-- The traces implement a finite reference set of procedures, not every valid
-  solution method.
-- Strategy labels describe the generator procedure and should not be treated
-  as natural human reasoning annotations.
-- Results on this benchmark do not by themselves establish strategy recovery
-  on open-ended natural-language reasoning.
-
-The dataset contains no human-authored examples or personal information.
-
 ## License
 
 The dataset is released under the