pythonformer/agents-learn-runtime-tasks

Agents Learn Their Runtime -- Task Definitions

Paper: Agents Learn Their Runtime: Interpreter Persistence as Training-Time Semantics

200 procedurally generated Opaque Knapsack tasks, split into 100 Easy and 100 Hard instances. These are the shared evaluation problems solved by all models in the paper's experiments.

The Opaque Knapsack Task

A partially observable constrained optimization problem. An agent is given a set of items identified only by opaque IDs and must:

1. Call inspect(item_id) to reveal an item's weight, value, and class (costs one unit of a limited inspection budget)
2. Call take_item(item_id) to select items, maximizing total value without exceeding a weight capacity
3. Respect class-validity constraints (only certain item classes are allowed)

Item properties are hidden behind random IDs, so the task is unsolvable by memorization. The agent must track running weight totals, budget usage, and candidate rankings across multiple steps.

Related Datasets

Dataset	What it contains
This dataset	Task definitions (the problems)
Training traces	2,000 teacher solutions by Gemini 3 Flash (1K persistent + 1K stateless), used to fine-tune LoRA adapters
Benchmark traces	1,200 inference traces from Qwen3-8B (base + 2 LoRA adapters) solving these exact tasks across 12 conditions

Structure

tasks/
├── easy/knapsack/
│   └── knapsack-0000000000.json ... knapsack-0000000099.json
└── hard/knapsack/
    └── knapsack-0000000000.json ... knapsack-0000000099.json

File Schema

Each JSON file fully specifies a single task instance:

{
  "task_id":   "unique identifier",
  "family":    "knapsack",
  "seed":      12345,
  "difficulty": {
    "n_items":             36,
    "capacity":            34,
    "budget_coverage":     0.58,
    "p_valid":             0.2,
    "optimal_set_size":    3,
    "max_item_dominance":  0.38
  },
  "public":    { "capacity": 34, "budget": 21, "valid_classes": ["A", "C"] },
  "private":   { "items": { "item_abc123": {"weight": 5, "value": 12, "class": "A"} } },
  "reference": { "optimal_value": 47, "optimal_items": ["item_abc123", "..."] },
  "nl":        { "title": "...", "instructions": "...", "output_format": "..." }
}

Field	Description
public	Parameters revealed to the agent (capacity, budget, valid classes)
private	Ground-truth item properties, hidden behind inspect() at runtime
reference	Optimal solution for scoring
nl	Natural-language prompt given to the agent
difficulty	Generation parameters controlling problem hardness

Reproduction

Tasks are generated via make tasks in the source repo, which calls pythonformer.cli with config files from pythonformer/task_configs/.

License

Apache License 2.0

Citation

@article{may2026agents,
  title={Agents Learn Their Runtime: Interpreter Persistence as Training-Time Semantics},
  author={May, Victor and Salgarkar, Aaditya and Wang, Yishan and Misra, Diganta and Nguyen, Huu},
  journal={arXiv preprint arXiv:2603.01209},
  year={2026}
}