SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2310.15213 · ICLR · 2024

Function Vectors in Large Language Models

David Bau, Eric Todd, Millicent Li, Sen Sharma, Aaron Mueller, Byron Wallace

arXiv · PDF · Open in the Atlas

Code that ran

6 of 7 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 2 have no contract verdict; 3 failures of our instrument; 1 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/2310.15213.json · badge: /api/ran/2310.15213.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
ericwtodd/function_vectors community (classifier label)
classifier basis not recorded on the link
6 ran · 1 no recorded run
FunctionStatusWhere it lives
add_function_vector Ran
ran on a synthesized input
ericwtodd/function_vectors/src/utils/intervention_utils.py
code served (permissive licence) · get_code("0095bc7827bfbc5e")
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 function. 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
import torch.nn as nn

torch.manual_seed(0)

# Create a small function vector
fv_vector = torch.randn(4)

# Device should be CPU
device = 'cpu'

# Edit layer index - which layer to apply the intervention to
edit_layer = 2

# Token index to add the function vector at
idx = -1

# Create a mock output that could come from a layer
# The function checks if output is a tuple and accesses output[0][:, idx]
# So we need a tuple where the first element is a tensor with shape (batch, seq_len, hidden_dim)
mock_output = (torch.randn(2, 5, 4),)  # batch=2, seq_len=5, hidden_dim=4

ARGS = {
    'edit_layer': edit_layer,
    'fv_vector': fv_vector,
    'device': device,
    'idx': idx,
}
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 function — a function output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a function 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
compute_best_token_rank Ran
ran; our contract draft was wrong, not the code
ericwtodd/function_vectors/src/utils/eval_utils.py
pointer only (licence: MIT); the stored copy get_code serves, from weichuanw/llm-based-mt-via-model-editing, is pointer only (licence: NONE) · get_code("a9853ae06ebc5529")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (MIT), but get_code serves the stored copy, first harvested from weichuanw/llm-based-mt-via-model-editing, and GitHub detected no licence in that repository 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.

compute_individual_token_rank Ran
ran; our contract draft was wrong, not the code
ericwtodd/function_vectors/src/utils/eval_utils.py
pointer only (licence: MIT); the stored copy get_code serves, from weichuanw/llm-based-mt-via-model-editing, is pointer only (licence: NONE) · get_code("06461c0c4a4708ce")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (MIT), but get_code serves the stored copy, first harvested from weichuanw/llm-based-mt-via-model-editing, and GitHub detected no licence in that repository 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.

compute_top_k_accuracy Ran
ran, violated the contract we drafted
ericwtodd/function_vectors/src/utils/eval_utils.py
pointer only (licence: MIT); the stored copy get_code serves, from weichuanw/llm-based-mt-via-model-editing, is pointer only (licence: NONE) · get_code("6351069e7766eb7a")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (MIT), but get_code serves the stored copy, first harvested from weichuanw/llm-based-mt-via-model-editing, and GitHub detected no licence in that repository 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_module Ran
ran; our contract draft was wrong, not the code
ericwtodd/function_vectors/src/utils/intervention_utils.py
code served (permissive licence) · get_code("8953f9ef90cf6afd")
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) and returned Linear. 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
import torch.nn as nn

torch.manual_seed(0)

# Create a simple model with named modules
model = nn.Sequential(
    nn.Linear(4, 8),
    nn.ReLU(),
    nn.Linear(8, 3)
)

# The function uses model.named_modules() to iterate through modules
# and compares the name 'n' with the 'name' parameter
# We need to provide a valid module name that exists in the model

# Get a valid module name from the model
valid_name = None
for n, m in model.named_modules():
    if n == "0":  # First Linear layer
        valid_name = n
        break

ARGS = {
    "model": model,
    "name": "0"
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Linear — a Linear output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Linear 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
optim_loop Ran
ran on a synthesized input
ericwtodd/function_vectors/src/vocab_reconstruction.py
code served (permissive licence) · get_code("a97b73b55d820f6a")
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 Parameter, ndim 2, dtype family float; no output shape is stated. The image it ran in was not recorded. The code mentions cuda, and the harness aliases cuda to the CPU for every sample; whether this run needed that is not recorded. If your copy returns the same type and ndim 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
import torch.nn as nn
from torch.optim import Adam

torch.manual_seed(0)

# Build a minimal decoder nn.Module
# The function calls decoder(v_n) and passes result to loss_fn
class MinimalDecoder(nn.Module):
    def __init__(self, input_dim, output_vocab):
        super().__init__()
        self.linear = nn.Linear(input_dim, output_vocab)
    
    def forward(self, x):
        return self.linear(x)

# Build minimal loss function
def minimal_loss_fn(pred, target):
    # loss_fn is called with decoder output and Z (target)
    return torch.nn.functional.mse_loss(pred, target)

# Create v_n: learnable parameter (batch_size=2, embedding_dim=4)
v_n = nn.Parameter(torch.randn(2, 4))

# Create target: logits for vocabulary (batch_size=2, vocab_size=100000)
# Using a larger vocab than restrict_vocab to trigger the topk logic
target = torch.randn(2, 100000)

# Create decoder
decoder = MinimalDecoder(input_dim=4, output_vocab=100000)

# Create loss function
loss_fn = minimal_loss_fn

# Create optimizer
optimizer = Adam([v_n], lr=0.01)

# Number of steps
n_steps = 5

# Verbose flag
verbose = False

# Restrict vocab size
restrict_vocab = 50400

ARGS = {
    'v_n': v_n,
    'target': target,
    'decoder': decoder,
    'loss_fn': loss_fn,
    'optimizer': optimizer,
    'n_steps': n_steps,
    'verbose': verbose,
    'restrict_vocab': restrict_vocab
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Parameter, ndim 2, dtype family float — no output shape is stated: observe_signature.py recorded Tensor ndim 2, which disagrees with the fixture run
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
load_gpt_model_and_tokenizer Not yet run
unverified (harvested, no recorded run)
ericwtodd/function_vectors/src/utils/model_utils.py
code served (permissive licence) · get_code("3eba226ef375d48e")
How it ran

no invocation recorded.

Also present in other papers' code

Of 7 samples lifted from this paper's linked repositories, 7 carry a function identity and 0 do not. Of the 7 distinct functions with an identity, 4 are also present in other papers whose code we hold (1 with at least one paper dated earlier, 0 the same month, 4 later) and 3 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
get_module 5 other papers
1 dated earlier, 4 later · 1 only as an assembly brick
2309.15098 (2023-09)
Attention Satisfies: A Constraint-Satisfaction Lens on Factual Errors of Language Models
2405.01008 (2024-05)
On Mechanistic Knowledge Localization in Text-to-Image Generative Models
2405.15349 (2024-05)
Everything is Editable: Extend Knowledge Editing to Unstructured Data in Large Language Models
and 2 more not named here
compute_best_token_rank 1 other paper
1 later
2410.07054 (2024-10)
Mitigating the Language Mismatch and Repetition Issues in LLM-based Machine Translation via Model Editing
compute_individual_token_rank 1 other paper
1 later
2410.07054 (2024-10)
Mitigating the Language Mismatch and Repetition Issues in LLM-based Machine Translation via Model Editing
compute_top_k_accuracy 1 other paper
1 later
2410.07054 (2024-10)
Mitigating the Language Mismatch and Repetition Issues in LLM-based Machine Translation via Model Editing

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

We report the presence of a simple neural mechanism that represents an inputoutput function as a vector within autoregressive transformer language models (LMs). Using causal mediation analysis on a diverse range of in-context-learning (ICL) tasks, we find that a small number attention heads transport a compact representation of the demonstrated task, which we call a function vector (FV). FVs are robust to changes in context, i.e., they trigger execution of the task on inputs such as zero-shot and natural text settings that do not resemble the ICL contexts from which they are collected. We test FVs across a range of tasks, models, and layers and find strong causal effects across settings in middle layers. We investigate the internal structure of FVs and find while that they often contain information that encodes the output space of the function, this information alone is not sufficient to reconstruct an FV. Finally, we test semantic vector composition in FVs, and find that to some extent they can be summed to create vectors that trigger new complex tasks. Our findings show that compact, causal internal vector representations of function abstractions can be explicitly extracted from LLMs.

For agents

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

get_harvested_code_for_paper("2310.15213")
get_code_for_paper("2310.15213")
have("2310.15213")

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

curl https://syntology.ai/api/ran/2310.15213.json

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

[![Syntology run record](https://syntology.ai/api/ran/2310.15213.svg)](https://syntology.ai/paper/2310.15213)

Connect an agent — have() is free.