SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2309.08600 · 2023

Sparse Autoencoders Find Highly Interpretable Features in Language Models

Lee Sharkey, Eleutherai, Logan Riggs, Aidan Ewart, Hoagy Cunningham, Robert Huben

arXiv · PDF · Open in the Atlas

Code that ran

8 of 11 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 1 violated it and 4 have no contract verdict; 3 failures of our instrument; 3 no recorded run), measured 2026-11-22; repository links from the Papers with Code archive attributed under CC BY-SA, no affiliation; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2309.08600.json · badge: /api/ran/2309.08600.svg

Each repository's role is a label: a classifier's output recorded on the link, or, where no role is recorded, what the Papers with Code archive marked. The line under it names its basis. Neither says who wrote the code.

RepositoryRoleRan
hoagyc/sparse_coding canonical (classifier label)
classifier basis: census
4 ran · 1 no recorded run
jbloomaus/saelens community (classifier label)
classifier basis not recorded on the link
4 ran · 2 no recorded run
FunctionStatusWhere it lives
ablate_feature_intervention Ran
ran; our contract draft was wrong, not the code
hoagyc/sparse_coding/standard_metrics.py
pointer only (licence: NONE); the stored copy get_code serves, from HoagyC/sparse_coding, is pointer only (licence: NONE) · get_code("5551196e92fa7236")
How it ran

The body of this sample is withheld: GitHub detected no licence in this paper's repository, hoagyc/sparse_coding when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

collect_lite Ran
ran, violated the contract we drafted
hoagyc/sparse_coding/cluster_runs.py
pointer only (licence: NONE); the stored copy get_code serves, from HoagyC/sparse_coding, is pointer only (licence: NONE) · get_code("03016136842a1f30")
How it ran

The body of this sample is withheld: GitHub detected no licence in this paper's repository, hoagyc/sparse_coding when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

filter_learned_dicts Ran
ran; our contract draft was wrong, not the code
hoagyc/sparse_coding/big_sweep.py
pointer only (licence: NONE); the stored copy get_code serves, from HoagyC/sparse_coding, is pointer only (licence: NONE) · get_code("99158f9a6ffb0e31")
How it ran

The body of this sample is withheld: GitHub detected no licence in this paper's repository, hoagyc/sparse_coding when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

get_eval_everything_config Ran
ran on a synthesized input
jbloomaus/saelens/sae_lens/evals.py
code served (permissive licence) · get_code("82ff8a8d62edb55d")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned EvalConfig. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
from pathlib import Path

torch.manual_seed(0)

# The function get_eval_everything_config takes three optional parameters:
# - batch_size_prompts: int | None
# - n_eval_reconstruction_batches: int
# - n_eval_sparsity_variance_batches: int

ARGS = {
    "batch_size_prompts": 32,
    "n_eval_reconstruction_batches": 10,
    "n_eval_sparsity_variance_batches": 1,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type EvalConfig — a EvalConfig output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a EvalConfig output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_model_tensor_name Ran
ran; our contract draft was wrong, not the code
hoagyc/sparse_coding/standard_metrics.py
pointer only (licence: NONE); the stored copy get_code serves, from HoagyC/sparse_coding, is pointer only (licence: NONE) · get_code("200b3f8d8c27361c")
How it ran

The body of this sample is withheld: GitHub detected no licence in this paper's repository, hoagyc/sparse_coding when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

get_special_token_from_cfg Ran
ran on a synthesized input
jbloomaus/saelens/sae_lens/pretokenize_runner.py
code served (permissive licence) · get_code("568742d62c1af63c")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned int, ndim 0, dtype family int, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
from transformers import AutoTokenizer

# Create a minimal stand-in tokenizer class that has the required attributes
class MinimalTokenizer:
    def __init__(self):
        self.bos_token_id = 101
        self.eos_token_id = 102
        self.sep_token_id = 103

tokenizer = MinimalTokenizer()

ARGS = {
    "cfg_token": "bos",
    "tokenizer": tokenizer,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch, transformers — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type int, ndim 0, dtype family int, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
json_dict Ran
ran on a synthesized input
jbloomaus/saelens/sae_lens/config.py
code served (permissive licence) · get_code("04cd1bc157c2e856")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import json
from typing import Any

def json_dict(s: str) -> Any:
    res = json.loads(s)
    if res is not None and (not isinstance(res, dict)):
        raise ValueError(f'Expected a dictionary, got {type(res)}')
    return res

# Test case 1: valid JSON dictionary
valid_dict_str = json.dumps({"key": "value", "nested": {"inner": 42}})

# Test case 2: valid empty dictionary
empty_dict_str = json.dumps({})

# Test case 3: valid dictionary with various types
complex_dict_str = json.dumps({"string": "test", "number": 123, "list": [1, 2, 3], "bool": True, "null": None})

ARGS = {
    "s": valid_dict_str
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
special_token Ran
ran on a synthesized input
jbloomaus/saelens/sae_lens/config.py
code served (permissive licence) · get_code("aac5ca4eeebb7d88")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

ARGS = {
    's': 'bos'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
calc_expected_interference Not yet run
unverified (harvested, no recorded run)
hoagyc/sparse_coding/big_sweep.py
pointer only (licence: NONE); the stored copy get_code serves, from HoagyC/sparse_coding, is pointer only (licence: NONE) · get_code("4d866dd8ed7f5121")
How it ran

The body of this sample is withheld: GitHub detected no licence in this paper's repository, hoagyc/sparse_coding when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. No invocation recorded.

get_sae_class Not yet run
unverified (harvested, no recorded run)
jbloomaus/saelens/sae_lens/registry.py
code served (permissive licence) · get_code("52c2cc2bd7a7d6e7")
How it ran

no invocation recorded.

get_sae_training_class Not yet run
unverified (harvested, no recorded run)
jbloomaus/saelens/sae_lens/registry.py
code served (permissive licence) · get_code("782e1939a5b1737f")
How it ran

no invocation recorded.

Also present in other papers' code

Of 11 samples lifted from this paper's linked repositories, 11 carry a function identity and 0 do not. Of the 11 distinct functions with an identity, 5 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 5 later) and 6 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

FunctionAlso present inEarliest-datedOthers named
ablate_feature_intervention 1 other paper
1 later
2606.11722 (2026-06)
ICA Lens: Interpreting Language Models Without Training Another Dictionary
calc_expected_interference 1 other paper
1 later
2606.11722 (2026-06)
ICA Lens: Interpreting Language Models Without Training Another Dictionary
collect_lite 1 other paper
1 later
2606.11722 (2026-06)
ICA Lens: Interpreting Language Models Without Training Another Dictionary
filter_learned_dicts 1 other paper
1 later
2606.11722 (2026-06)
ICA Lens: Interpreting Language Models Without Training Another Dictionary
get_model_tensor_name 1 other paper
1 later
2606.11722 (2026-06)
ICA Lens: Interpreting Language Models Without Training Another Dictionary

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

One of the roadblocks to a better understanding of neural networks' internals is polysemanticity, where neurons appear to activate in multiple, semantically distinct contexts. Polysemanticity prevents us from identifying concise, humanunderstandable explanations for what neural networks are doing internally. One hypothesised cause of polysemanticity is superposition, where neural networks represent more features than they have neurons by assigning features to an overcomplete set of directions in activation space, rather than to individual neurons. Here, we attempt to identify those directions, using sparse autoencoders to reconstruct the internal activations of a language model. These autoencoders learn sets of sparsely activating features that are more interpretable and monosemantic than directions identified by alternative approaches, where interpretability is measured by automated methods. Moreover, we show that with our learned set of features, we can pinpoint the features that are causally responsible for counterfactual behaviour on the indirect object identification task (Wang et al., 2022) to a finer degree than previous decompositions. This work indicates that it is possible to resolve superposition in language models using a scalable, unsupervised method. Our method may serve as a foundation for future mechanistic interpretability work, which we hope will enable greater model transparency and steerability. * Equal contribution Code to replicate experiments can be found at https://github.com/HoagyC/sparse_coding

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2309.08600")
get_code_for_paper("2309.08600")
have("2309.08600")

The run record, dated, one paper per request, free:

curl https://syntology.ai/api/ran/2309.08600.json

A badge for a README (the split and the date, never a ratio):

[![Syntology run record](https://syntology.ai/api/ran/2309.08600.svg)](https://syntology.ai/paper/2309.08600)

Connect an agent — have() is free.