SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2603.09616 · 2026

Surgical Repair of Collapsed Attention Heads

Palmer Schallon

arXiv · PDF · Open in the Atlas

Code that ran

1 of 1 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; 1 honoured the contract we drafted, 0 violated it and 0 have no contract verdict; 0 failures of our instrument; 0 no recorded run), measured 2026-11-22; 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/2603.09616.json · badge: /api/ran/2603.09616.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
Palmerschallon/bloom-head-surgery community
Papers with Code archive
1 ran · 0 no recorded run
FunctionStatusWhere it lives
reinitialize_dead_heads Ran
ran, honoured the contract we drafted
Palmerschallon/bloom-head-surgery/surgery/train_1b7_headband.py
code served (permissive licence) · get_code("259c4a9e66914c02")
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 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
import torch.nn as nn

torch.manual_seed(0)

# Build a minimal config object
class MinimalConfig:
    def __init__(self):
        self.hidden_size = 12  # Must be divisible by n_head
        self.n_head = 3

# Build minimal attention module
class MinimalAttention(nn.Module):
    def __init__(self, hidden_size, n_head):
        super().__init__()
        self.query_key_value = nn.Linear(hidden_size, 3 * hidden_size)
        self.dense = nn.Linear(hidden_size, hidden_size)

# Build minimal transformer layer
class MinimalTransformerLayer(nn.Module):
    def __init__(self, hidden_size, n_head):
        super().__init__()
        self.self_attention = MinimalAttention(hidden_size, n_head)

# Build minimal transformer
class MinimalTransformer(nn.Module):
    def __init__(self, hidden_size, n_head, num_layers=2):
        super().__init__()
        self.h = nn.ModuleList([
            MinimalTransformerLayer(hidden_size, n_head)
            for _ in range(num_layers)
        ])

# Build minimal model
class MinimalModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.config = MinimalConfig()
        self.transformer = MinimalTransformer(
            self.config.hidden_size,
            self.config.n_head,
            num_layers=2
        )

model = MinimalModel()

# Build surgery_config: specifies which heads to reinitialize in which layers
surgery_config = {
    0: {'reinit': [0, 1]},  # Reinitialize heads 0 and 1 in layer 0
    1: {'reinit': [2]},      # Reinitialize head 2 in layer 1
}

device = 'cpu'

ARGS = {
    'model': model,
    'surgery_config': surgery_config,
    'device': device,
}
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 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

Also present in other papers' code

Of 1 samples lifted from this paper's linked repositories, 1 carry a function identity and 0 do not. Of the 1 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 1 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

We identify a systematic attention collapse pathology in the BLOOM family of transformer language models, where ALiBi positional encoding causes 31-44% of attention heads to attend almost entirely to the beginning-of-sequence token. The collapse follows a predictable pattern across four model scales (560M to 7.1B parameters), concentrating in head indices where ALiBi's slope schedule imposes the steepest distance penalties. We introduce surgical reinitialization: targeted Q/K/V reinitialization with zeroed output projections and gradient-masked freezing of all non-surgical parameters. Applied to BLOOM-1b7 on a single consumer GPU, the technique recovers 98.7% operational head capacity (242 → 379 of 384 heads) in two passes. A controlled comparison with C4 training data confirms that reinitialization-not corpus content-drives recovery, and reveals two distinct post-surgical phenomena: early global functional redistribution that improves the model, and late local degradation that accumulates under noisy training signal. An extended experiment reinitializing mostly-healthy heads alongside collapsed ones produces a model that transiently outperforms stock BLOOM-1b7 by 25% on training perplexity (12.70 vs. 16.99), suggesting that pretrained attention configurations are suboptimal local minima. Code, checkpoints, and diagnostic tools are released as open-source software.

For agents

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

get_harvested_code_for_paper("2603.09616")
get_code_for_paper("2603.09616")
have("2603.09616")

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

curl https://syntology.ai/api/ran/2603.09616.json

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

[![Syntology run record](https://syntology.ai/api/ran/2603.09616.svg)](https://syntology.ai/paper/2603.09616)

Connect an agent — have() is free.