ContextBench

ContextBench measures what an agent keeps, updates, and offloads when its input does not fit in its context window. Each task streams a sequence of operations as ordinary user turns, more than the context budget holds, and the reward is read from the agent's own messages: what survived in its context, not what it wrote to disk.

Paper: Context Language Models (paper page)

Families

config	task	reward
kv_store	batches of SET key = value, then GET key queries	fraction of GETs answered with the right value
log_triage	batches of service log lines, then count and lookup questions	fraction of questions answered correctly
needle_retention	chunks with a few needle lines inside a large filler block	fraction of needle lines still present verbatim
sudoku_sketchpad	one fully specified move per turn on a 16x16 board	fraction of gold board versions the agent displayed

The all config holds every instance. This release has 256 instances: 64 kv_store, 56 log_triage, 80 needle_retention, 56 sudoku_sketchpad (32 levels, seeds 42-49), with a context budget of 32768 tokens and a reserve of 2048.

Columns

column	type	meaning
task_id	string	<family>-<level>-seed<seed>, also the Harbor task directory name
family	string	task family
level	string	difficulty level
seed	int	generator seed
budget	int	context budget in tokens
reserve	int	tokens reserved for the reply; the agent may use budget - reserve
n_ops	int	number of streamed operations
streamed_tokens	int	tokens streamed over the episode (o200k_base)
pressure	float	streamed_tokens / budget
chunks	list of {index, header, body}	the operations, one user turn each
gold	string (JSON)	the answers the graders score against
meta	string (JSON)	generator settings and instance statistics

The budget is counted in o200k_base tokens for every agent. An agent may enforce a stricter budget of its own; that is part of the agent, not of the benchmark.

Each operation is shown to the agent as === {header} ({remaining} op(s) remaining after this) === followed by its body.

Usage

With datasets

import json
from datasets import load_dataset

ds = load_dataset("context-language-model/ContextBench", "kv_store", split="test")   # or "all"
row = ds[0]
ops = row["chunks"]            # list of {"index", "header", "body"}
gold = json.loads(row["gold"])

To evaluate an agent from Python, install the contextbench package, write an instance to disk, and pass your agent to contextbench.run:

from pathlib import Path
from contextbench import run

inst = Path("instances") / row["task_id"]
inst.mkdir(parents=True, exist_ok=True)
(inst / "chunks.jsonl").write_text(
    "".join(json.dumps(c) + "\n" for c in row["chunks"]))
(inst / "gold.json").write_text(row["gold"])
(inst / "meta.json").write_text(row["meta"])

result = run(my_agent, inst, budget=row["budget"], reserve=row["reserve"],
             out_dir=Path("runs") / row["task_id"])

With Harbor

The harbor/ directory holds one Harbor task per instance, and registry.json lists the datasets contextbench (every task) and contextbench-<family> (for example contextbench-kv-store). The same tasks are on GitHub, so Harbor runs them in one line:

harbor jobs start --repo RulinShao/ContextBench -d contextbench \
    -a contextbench.harbor.agents:Terminus2Stream -m <model>

Or download this repository and run Harbor from inside it:

hf download context-language-model/ContextBench --repo-type dataset --local-dir contextbench-hf
cd contextbench-hf
harbor jobs start --registry-path registry.json -d contextbench \
    -a contextbench.harbor.agents:Terminus2Stream -m <model>

Use -d contextbench-sudoku-sketchpad for one family, -i <task_id> for one task, -p harbor/ to run the downloaded directory without the registry, or -a oracle to check a task with its reference solution. Agents that deliver operations into a running conversation ship with the contextbench package (contextbench.harbor.agents).

Citation

ContextBench was introduced in the Context Language Models paper. If you use it, please cite:

@article{shao2026context,
  title   = {Context Language Models},
  author  = {Shao, Rulin and Shen, Shannon Zejiang and Yin, Junjie Oscar and Li, Yuetai and
             Wang, Minheng and Ivison, Hamish and Poovendran, Radha and Lambert, Nathan and
             Xiao, Teng and Lewis, Mike and Yih, Wen-tau and Zettlemoyer, Luke and Koh, Pang Wei},
  journal = {arXiv preprint arXiv:2609.37725},
  year    = {2026}
}

License

Apache-2.0.