LatentMAS Repro (2511.20639)
Collection
Reproduction of LatentMAS (orid syG9I9ofd8): honest negative on accuracy + partial efficiency. • 1 item • Updated
Dataset Viewer
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code: StreamingRowsError
Exception: TypeError
Message: Couldn't cast array of type
struct<status: string, latent_steps: int64, n: int64, correct: int64, accuracy: double>
to
{'status': Value('string'), 'latent_steps': Value('int64'), 'error': Value('string'), 'trace': Value('string')}
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 149, in get_rows_or_raise
return get_rows(
dataset=dataset,
...<4 lines>...
column_names=column_names,
)
File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
return func(*args, **kwargs)
File "/src/services/worker/src/worker/utils.py", line 129, in get_rows
rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
File "/src/services/worker/src/worker/utils.py", line 489, in safe_iter
yield from ds.decode(False) if ds.features else ds
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2818, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2355, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2380, in _iter_arrow
for key, pa_table in self.ex_iterable._iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 536, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, in _iter_arrow
for key, pa_table in self.generate_tables_fn(**gen_kwags):
~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
self._cast_table(pa_table, json_field_paths=json_field_paths),
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
pa_table = table_cast(pa_table, self.info.features.arrow_schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2369, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2303, in cast_table_to_schema
cast_array_to_feature(
~~~~~~~~~~~~~~~~~~~~~^
table[name] if name in table_column_names else pa.array([None] * len(table), type=schema.field(name).type),
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
feature,
^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 1852, in wrapper
return pa.chunked_array([func(chunk, *args, **kwargs) for chunk in array.chunks])
~~~~^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2059, in cast_array_to_feature
_c(array.field(name) if name in array_fields else null_array, subfeature)
~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 1854, in wrapper
return func(array, *args, **kwargs)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2149, in cast_array_to_feature
raise TypeError(f"Couldn't cast array of type\n{_short_str(array.type)}\nto\n{_short_str(feature)}")
TypeError: Couldn't cast array of type
struct<status: string, latent_steps: int64, n: int64, correct: int64, accuracy: double>
to
{'status': Value('string'), 'latent_steps': Value('int64'), 'error': Value('string'), 'trace': Value('string')}Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Reproduction — LatentMAS (arXiv 2511.20639, OpenReview syG9I9ofd8)
Reproduction of the two scored claims of "Latent Collaboration in Multi-Agent Systems" (LatentMAS, ICML 2026 Spotlight) for the ICML-2026-agent-repro hackathon.
- Claim 1 (accuracy): LatentMAS outperforms single agents and text-based MAS, up to +14.6% across 9 benchmarks.
- Claim 2 (efficiency): 70.8%–83.7% fewer output tokens and 4–4.3× faster end-to-end inference, with lossless information exchange.
What this is
bench.py is a self-contained UV script that runs the official Gen-Verse/LatentMAS
implementation (HF backend — the backend the paper's README says reproduces the published
numbers) for all three methods on the same data subset with the same model, and
measures for each:
- accuracy
- end-to-end inference wall-clock (model already loaded)
- total output (generated text) tokens = Σ over agents of
tokenize(agent.output)
Then it computes output-token reduction %, speedup ×, and accuracy deltas of LatentMAS vs TextMAS / single-agent baseline.
Mechanism being measured
In LatentMAS the 3 reasoning agents (Planner/Critic/Refiner) communicate via latent thoughts written into a shared KV cache and emit no decoded text (0 output tokens); only the final Judger decodes text. TextMAS decodes full text at all 4 agents. That ~4× fewer decoded agents is the source of both the token reduction and the speedup.
Faithfulness notes (documented, HF-numerics-neutral)
Run against the official repo main. Three minimal, documented shims applied at runtime in
prepare_repo() / driver so the 2025-era code runs on a current UV env:
from vllm import SamplingParamsmade optional (HF backend needs no vLLM engine).- Legacy
gsm8kdataset id →openai/gsm8k(namespaced id now required). - Judger decode on top of the long prefilled latent KV cache uses a manual batched
autoregressive loop calling
model.forwarddirectly (same as the officialgenerate_latent_batch), becausetransformers.generate()recomputescache_positionasarange(0, seq_len)[past_len:]— empty whenpast_len >> prompt_len, which is exactly the LatentMAS regime. Sampling (temperature/top_p) and the latent cache are preserved.transformers==4.51.3pinned (contemporaneous with the paper's Qwen3 experiments; keeps the subscriptableDynamicCachethe official_past_lengthrelies on).
Run
# smoke (cheap GPU, tiny model + subset)
hf jobs uv run --flavor t4-small --timeout 25m -d --secrets HF_TOKEN \
-e MODEL=Qwen/Qwen3-0.6B -e TASK=gsm8k -e MAX_SAMPLES=4 \
-e MAX_NEW_TOKENS=256 -e LATENT_STEPS=4 -e GENERATE_BS=4 -e THINK=0 bench.py
# full (A100, real paper model)
hf jobs uv run --flavor a100-large --timeout 60m -d --secrets HF_TOKEN \
-e MODEL=Qwen/Qwen3-4B -e TASK=gsm8k -e MAX_SAMPLES=50 \
-e MAX_NEW_TOKENS=2048 -e LATENT_STEPS=8 -e GENERATE_BS=25 -e THINK=1 \
-e PUSH_DATASET=ai-sherpa/repro-latentmas-bundle -e RUN_TAG=qwen3_4b_gsm8k bench.py
Results JSON is pushed to dataset ai-sherpa/repro-latentmas-bundle under results/.
- Downloads last month
- 74