SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2411.05037 · 2024

Towards Interpreting Language Models: A Case Study in Multi-Hop Reasoning

arXiv · PDF · Open in the Atlas

Code that ran

2 of 3 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, 0 violated it and 0 have no contract verdict; 2 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/2411.05037.json · badge: /api/ran/2411.05037.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
msakarvadia/attentionlens canonical (classifier label)
classifier basis: census
2 ran · 1 no recorded run
FunctionStatusWhere it lives
classify_toxicity Ran
ran; our contract draft was wrong, not the code
msakarvadia/attentionlens/demos/measure_toxicity.py
code served (permissive licence) · get_code("279620b4fa77ddd2")
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 tuple, ndim 1, dtype family float; no output shape is stated. The image it ran in was 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

torch.manual_seed(0)

class MockClassifier(nn.Module):
    """Mock classifier that returns toxicity classification results."""
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(10, 2)
    
    def forward(self, text):
        # Return a list with a dict containing label and score
        # This mimics HuggingFace transformers classifier output format
        return [
            {
                'label': 'TOXIC',
                'score': 0.95
            }
        ]

classifier = MockClassifier()
text = "This is a sample text that might be toxic or not toxic depending on the classifier's analysis."

ARGS = {
    'classifier': classifier,
    'text': text
}
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 tuple, ndim 1, dtype family float — a tuple output has no array shape of its own; none is stated
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
generate_text Ran
ran; our contract draft was wrong, not the code
msakarvadia/attentionlens/demos/measure_toxicity.py
code served (permissive licence) · get_code("c64dd03c9f019df6")
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
import torch.nn as nn

torch.manual_seed(0)

# Stand-in tokenizer
class MockTokenizer:
    def __init__(self):
        self.eos_token_id = 50256
    
    def encode(self, text, max_length=900, truncation=True, return_tensors='pt'):
        # Return a tensor of shape (1, seq_len) with token IDs
        tokens = torch.tensor([[1, 2, 3, 4, 5]], dtype=torch.long)
        return tokens
    
    def decode(self, tokens, skip_special_tokens=True):
        return "This is the generated text output from the model"

# Stand-in model that supports generate with injection parameters
class MockModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.embedding = nn.Embedding(50257, 8)
        self.linear = nn.Linear(8, 50257)
    
    def forward(self, input_ids):
        x = self.embedding(input_ids)
        return self.linear(x)
    
    def generate(self, inputs, max_new_tokens=20, num_return_sequences=1, 
                 pad_token_id=None, inject_tensor=None, inject_layer=None, 
                 inject_head=None):
        # Mock generate method that accepts injection parameters
        # Return generated token IDs
        batch_size = inputs.shape[0]
        generated = torch.cat([inputs, torch.randint(0, 1000, (batch_size, max_new_tokens))], dim=1)
        return generated

# Create instances
device = 'cpu'
model = MockModel()
tokenizer = MockTokenizer()
prompt = "Once upon a time"
inject_tensor = torch.randn(1, 8, 4)  # Small tensor for injection
inject_layer = 2
inject_head = 1

ARGS = {
    'model': model,
    'tokenizer': tokenizer,
    'prompt': prompt,
    'device': device,
    'inject_tensor': inject_tensor,
    'inject_layer': inject_layer,
    'inject_head': inject_head,
    'max_new_tokens': 20
}
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
get_model Not yet run
unverified (harvested, no recorded run)
msakarvadia/attentionlens/attention_lens/model/get_model.py
code served (permissive licence) · get_code("2584e493335f04dd")
How it ran

no invocation recorded.

Also present in other papers' code

Of 3 samples lifted from this paper's linked repositories, 3 carry a function identity and 0 do not. Of the 3 distinct functions with an identity, 0 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 0 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.

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

Answering multi-hop reasoning questions requires retrieving and synthesizing information from diverse sources. Language models (LMs) struggle to perform such reasoning consistently. We propose an approach to pinpoint and rectify multi-hop reasoning failures through targeted memory injections on LM attention heads. First, we analyze the per-layer activations of GPT-2 models in response to single- and multi-hop prompts. We then propose a mechanism that allows users to inject relevant prompt-specific information, which we refer to as "memories," at critical LM locations during inference. By thus enabling the LM to incorporate additional relevant information during inference, we enhance the quality of multi-hop prompt completions. We empirically show that a simple, efficient, and targeted memory injection into a key attention layer often increases the probability of the desired next token in multi-hop tasks, by up to 424%. We observe that small subsets of attention heads can significantly impact the model prediction during multi-hop reasoning. To more faithfully interpret these heads, we develop Attention Lens: an open source tool that translates the outputs of attention heads into vocabulary tokens via learned transformations called lenses. We demonstrate the use of lenses to reveal how a model arrives at its answer and use them to localize sources of model failures such as in the case of biased and malicious language generation.

For agents

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

get_harvested_code_for_paper("2411.05037")
get_code_for_paper("2411.05037")
have("2411.05037")

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

curl https://syntology.ai/api/ran/2411.05037.json

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

[![Syntology run record](https://syntology.ai/api/ran/2411.05037.svg)](https://syntology.ai/paper/2411.05037)

Connect an agent — have() is free.