Surgical Repair of Collapsed Attention Heads in ALiBi Transformers

• URL: http://arxiv.org/abs/2603.09616v1
• Date: Tue, 10 Mar 2026 12:57:49 GMT
• Title: Surgical Repair of Collapsed Attention Heads in ALiBi Transformers
• Abstract summary: We identify a systematic attention collapse pathology in the BLOOM family of transformer language models.<n>ALiBi positional encoding causes 31-44% of attention heads to attend almost entirely to the beginning-of-sequence token.<n>We introduce surgical reinitialization: targeted Q/K/V reinitialization with zeroed output projections.
• Score: 0.0
• License: http://creativecommons.org/licenses/by/4.0/
• 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 to 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.

Related papers

• On the Design of Qwen3.8-Next Architecture: Evaluation, Efficiency, and Training Stability [57.03393882933515]
  We describe the architecture and ablations of Qwen3.8-Flash-Next, a sparse mixture-of-experts model with 125B parameters, 6B activated per token, and 51B parameters of n-gram embedding tables held off the accelerator.<n>On fourteen pre-training benchmarks the model leads the 397B-A17B predecessor on eight and trails it on the rest by at most 2.6 points, at 1/3 the activated parameters, 1/3 the training tokens, and roughly 1/9 the training FLOPs.
  arXiv  Detail & Related papers  (2026-08-31T06:35:07Z)
• Output Vector Editing for Memorization Mitigation in Large Language Models [68.30351930772788]
  Large language models memorize and reproduce sequences from their training data, creating privacy, copyright, and security risks.<n>Existing neuron-level mitigation methods equate editing with zeroing out neuron activations, but the activation only controls whether a neuron engages.<n>We propose output vector editing, a constrained-optimization edit that locates a small set of neurons responsible for a memorized continuation.
  arXiv  Detail & Related papers  (2026-06-17T07:29:18Z)
• How Much Capacity Does EEG Denoising Need? Ultra-Compact Networks reveal Benchmark Saturation and Metric-Utility Gap [0.0]
  Deep learning EEG denoising architectures have scaled from tens of thousands to tens of millions of parameters, yet no prior study has isolated model capacity as the experimental variable.<n>We address both gaps by fixing architecture, loss, data split, and training recipe while sweeping only channel width from 1.05K to 40.26K parameters in a minimal depthwiseseparable convolutional U-Net.
  arXiv  Detail & Related papers  (2026-06-07T12:17:25Z)
• Human-Inspired Memory Architecture for LLM Agents [0.9507070656654629]
  We present a biologically-grounded memory architecture comprising six cognitive mechanisms.<n>Each mechanism addresses a specific failure mode of naive memory accumulation.<n>At S-tier scale (50 sessions), dedup-based consolidation yields a +13.3 pp improvement in preference recall.
  arXiv  Detail & Related papers  (2026-05-08T22:52:37Z)
• Perturbation Probing: A Two-Pass-per-Prompt Diagnostic for FFN Behavioral Circuits in Aligned LLMs [9.127363793428119]
  Perturbation probing generates task-specific causal hypotheses for FFN neurons in large language models.<n>Across eight behavioral circuits, 13 models, and four architecture families, we identify two circuit structures that organize LLM behavior.
  arXiv  Detail & Related papers  (2026-04-30T04:13:33Z)
• Fragile Reasoning: A Mechanistic Analysis of LLM Sensitivity to Meaning-Preserving Perturbations [2.7946918847372277]
  Large language models demonstrate strong performance on mathematical reasoning benchmarks, yet remain surprisingly fragile to meaning-preserving surface perturbations.<n>We systematically evaluate three open-weight LLMs, Mistral-7B, Llama-3-8B, and Qwen2.5-7B, on 677 GSM8K problems paired with semantically equivalent variants.<n>All three models exhibit substantial answer-flip rates (28.8%-45.1%), with number paraphrasing consistently more disruptive than name swaps.
  arXiv  Detail & Related papers  (2026-04-02T05:30:20Z)
• SCALE:Scalable Conditional Atlas-Level Endpoint transport for virtual cell perturbation prediction [46.645872209565574]
  We present a specialized large-scale foundation model for virtual cell perturbation prediction.<n>First, we build a BioNeMo-based training and inference framework that substantially improves data throughput.<n>Second, we formulate perturbation prediction as conditional transport and implement it with a set-aware flow architecture.<n>Third, we evaluate the model on Tahoe-100M using a rigorous cell-level protocol centered on biologically meaningful metrics.
  arXiv  Detail & Related papers  (2026-03-18T05:51:08Z)
• A Quantization-Aware Training Based Lightweight Method for Neural Distinguishers [4.751864584128287]
  In 2019, Gohr pioneered the application of deep neural networks to differential cryptanalysis.<n>This study proposes a lightweight neural distinguisher based on quantization-aware training.
  arXiv  Detail & Related papers  (2026-03-06T00:52:49Z)
• SurGo-R1: Benchmarking and Modeling Contextual Reasoning for Operative Zone in Surgical Video [14.263054454077272]
  We introduce ResGo, a benchmark of laparoscopic frames annotated with Go Zone bounding boxes.<n>We then present SurGo-R1, a model optimized via RLHF with a multi-turn phase-then-go architecture.<n>On unseen procedures, SurGo-R1 achieves 76.6% phase accuracy, 32.7 mIoU, and 54.8% hardcore accuracy.
  arXiv  Detail & Related papers  (2026-02-25T09:11:45Z)
• ShortcutBreaker: Low-Rank Noisy Bottleneck with Global Perturbation Attention for Multi-Class Unsupervised Anomaly Detection [59.89803740308262]
  ShortcutBreaker is a novel unified feature-reconstruction framework for MUAD tasks.<n>It features two key innovations to address the issue of shortcuts.<n>The proposed method achieves a remarkable image-level AUROC of 99.8%, 98.9%, 90.6%, and 87.8% on four datasets.
  arXiv  Detail & Related papers  (2025-10-21T06:51:30Z)
• Hemorica: A Comprehensive CT Scan Dataset for Automated Brain Hemorrhage Classification, Segmentation, and Detection [0.749500254646884]
  Hemorica is a publicly available collection of 372 head CT examinations acquired between 2012 and 2024.<n>Each scan has been exhaustively annotated for five ICH subtypes-epidural (EPH), subdural (SDH), subarachnoid (SAH), intraparenchymal (IPH)<n>Hemorica offers a unified, fine-grained benchmark that supports multi-task and curriculum learning.
  arXiv  Detail & Related papers  (2025-09-26T23:09:41Z)
• Organ-Agents: Virtual Human Physiology Simulator via LLMs [66.40796430669158]
  Organ-Agents is a multi-agent framework that simulates human physiology via LLM-driven agents.<n>We curated data from 7,134 sepsis patients and 7,895 controls, generating high-resolution trajectories across 9 systems and 125 variables.<n>Organ-Agents achieved high simulation accuracy on 4,509 held-out patients, with per-system MSEs 0.16 and robustness across SOFA-based severity strata.
  arXiv  Detail & Related papers  (2025-08-20T01:58:45Z)
• NeRF-based CBCT Reconstruction needs Normalization and Initialization [53.58395475423445]
  NeRF-based methods suffer from a local-global training mismatch between their two key components: the hash encoder and the neural network.<n>We introduce a Normalized Hash, which enhances feature consistency and mitigates the mismatch.<n>The neural network exhibits improved stability during early training, enabling faster convergence and enhanced reconstruction performance.
  arXiv  Detail & Related papers  (2025-06-24T16:01:45Z)
• Quantization Meets Reasoning: Exploring and Mitigating Degradation of Low-Bit LLMs in Mathematical Reasoning [39.56908863102256]
  Low-bit post-training quantization impairs mathematical reasoning up to 69.81% in harder settings.<n>We address two deployment-critical questions with process-level precision.<n>In our settings, as few as 332 curated examples and 3--5 minutes of compute on a single GPU recover 4-bit weight math reasoning toward the full-precision baseline.
  arXiv  Detail & Related papers  (2025-05-16T12:11:40Z)
• Training-free Ultra Small Model for Universal Sparse Reconstruction in Compressed Sensing [39.36305648162564]
  This paper proposes ultra-small artificial neural models called coefficients learning (CL)<n>CL enables training-free and rapid sparse reconstruction while inheriting the generality and interpretability of traditional iterative methods.<n>Compared to representative iterative methods, CLOMP improves efficiency by 100 to 1000 folds for large-scale data.
  arXiv  Detail & Related papers  (2025-01-20T16:50:59Z)
• Multilabel 12-Lead Electrocardiogram Classification Using Gradient Boosting Tree Ensemble [64.29529357862955]
  We build an algorithm using gradient boosted tree ensembles fitted on morphology and signal processing features to classify ECG diagnosis. For each lead, we derive features from heart rate variability, PQRST template shape, and the full signal waveform. We join the features of all 12 leads to fit an ensemble of gradient boosting decision trees to predict probabilities of ECG instances belonging to each class.
  arXiv  Detail & Related papers  (2020-10-21T18:11:36Z)
• Generalized Focal Loss: Learning Qualified and Distributed Bounding Boxes for Dense Object Detection [85.53263670166304]
  One-stage detector basically formulates object detection as dense classification and localization. Recent trend for one-stage detectors is to introduce an individual prediction branch to estimate the quality of localization. This paper delves into the representations of the above three fundamental elements: quality estimation, classification and localization.
  arXiv  Detail & Related papers  (2020-06-08T07:24:33Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.