nik/Masters_Thesis
1
0 Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Add literature review protocol, searches and first snowball round

Browse Source 
Six seeds snowballed (Semantic Scholar) plus logged keyword searches
(OpenAlex, arXiv): 2,127 unique records, 480 title/abstract screened.
AI first-pass suggestions are in their own column; decisions are pending.

Claude-Session: https://claude.ai/code/session_01S8KN7NuUCm8nXLmy6eDuVv
This commit is contained in:
Cormac Shannon committed 2026-10-28 18:00:06 +00:00
1 parent 4369bcefa8
commit 95b5857189
28 files changed
+5547
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

02-literature/search/results/snowball-fwd-stability2024.csv
+17
View File
@@ -0,0 +1,17 @@
s2_id,arxiv,doi,year,title,first_author,n_authors,venue,citations,abstract
9653176d4dafd012f0b04775889b80cc7b8c6b91,2608.02091,,2026,"One QK Channel, Many Sources: Tracing Low-Precision Attention Collapse",Shu-Xiao Xie,6,,0,"A bfloat16 transformer can train normally, then collapse abruptly. Prior work links collapse to structured attention errors and shows QK normalization disrupts their compounding. Distinct low-precision errors trigger the same collapse, leaving unclear whether each needs a fix at its source or one shared route can be blocked instead. We isolated the fault behind a reproduced GPT-2-class collapse to the streaming-softmax accumulator, where an fp32 streaming core repairs it, and turned it into an assay for moving a controlled error across sources. Using it, we found that errors placed outside attention still drove the same QK spectral runaway, and that correcting only QK kept training stable while the fault stayed active. This is a source-channel dissociation: fault source is not failure channel. It held across tested architectures and scales, and reproduced on a second GPU architecture. As a causal probe, projecting each update off the current QK weights'leading three singular directions held the query projection's largest singular value to 11.1, whereas removing equal energy elsewhere left it at 237: the QK channel causally drives the early runaway. What lets the injected error in is temporal sign-coherence, its per-head sign persisting across steps, not aggregate deviation; once inside, the runaway shows as attention-logit saturation. QK-Guard, a dormant controller, tests this by switching on parameter-free QK normalization at the first monitored threshold crossing. On the runs designated for this test, the QK-local action prevented the failure of each matched or same-configuration unguarded run; on plain GPT-2, all 12 final train and validation losses were within 0.03 nat of same-configuration always-on QK-norm, and both methods ran 60k steps without collapse. Intervention at the QK locus therefore suffices in place of a fix at each source."
cacb1f26dd4ca32fb6663fb7ca0489a05b598a6e,2606.08105,,2026,A Unifying View of Attention Sinks: From Mechanisms to Architectural Interventions,Lukas Fesser,5,,1,"When attention concentrates on a single token, a sink, what is the model actually computing? Attention sinks are ubiquitous in softmax transformers, yet this shared visual signature can hide fundamentally different algorithms. We show that visually similar sink patterns can reflect two distinct mechanisms: (i) adaptive nop, where a head suppresses its update by routing to a null token, and (ii) broadcast, where a sink aggregates and redistributes global information. Each mechanism leaves distinct traces (nop-sinks exhibit negligible value norms; broadcast sinks induce low-rank outputs), which we formalize on synthetic tasks and use to derive practical diagnostics. Applied to pretrained vision transformers, these diagnostics reveal that both mechanisms exist at scale: sinks transition from CLS in early layers to patches in deeper layers and concentrate in specialized heads. Causal interventions further connect these signatures to near-null suppression and shared residual contributions. We then use architectural interventions to show how these computations can be reorganized: gating eliminates detected nop-like sinks but increases broadcast-like sinks, registers relocate rather than remove sink computation, and our position-free global pathway provides an explicit route for shared communication that reduces the broadcast-like sinks induced by gating. On dense probes, combining gating with the global pathway gives the strongest results among the tested variants, despite retaining some broadcast-like sinks. Overall, we find that the same attention pattern can reflect two very different computations, and that effective intervention depends not only on identifying the computation, but also on providing architectural alternatives through which the model can reorganize it."
dd91f0cdb8cc1da2b7c211fcaac523d14834eea1,2605.12492,10.48550/arXiv.2605.12492,2026,Pion: A Spectrum-Preserving Optimizer via Orthogonal Equivalence Transformation,Kexuan Shi,6,arXiv.org,6,"We introduce Pion, a spectrum-preserving optimizer for large language model (LLM) training based on orthogonal equivalence transformation. Unlike additive optimizers such as Adam and Muon, Pion updates each weight matrix through left and right orthogonal transformations, preserving its singular values throughout training. This yields an optimization mechanism that modulates the geometry of weight matrices while keeping their spectral norm fixed. We derive the Pion update rule, systematically examine its design choices, and analyze its convergence behavior along with several key properties. Empirical results show that Pion offers a stable and competitive alternative to standard optimizers for both LLM pretraining and finetuning."
3b07bb0e7ff9068eae67d7a4867c4a72518d5df0,2601.22095,10.48550/arXiv.2601.22095,2026,GeoNorm: Unify Pre-Norm and Post-Norm with Geodesic Optimization,Chuan-Yang Zheng,14,arXiv.org,1,"The placement of normalization layers, specifically Pre-Norm and Post-Norm, remains an open question in Transformer architecture design. In this work, we rethink these approaches through the lens of manifold optimization, interpreting the outputs of the Feed-Forward Network (FFN) and attention layers as update directions in optimization. Building on this perspective, we introduce GeoNorm, a novel method that replaces standard normalization with geodesic updates on the manifold. Furthermore, analogous to learning rate schedules, we propose a layer-wise update decay for the FFN and attention components. Comprehensive experiments demonstrate that GeoNorm consistently outperforms existing normalization methods in Transformer models. Crucially, GeoNorm can be seamlessly integrated into standard Transformer architectures, achieving performance improvements with negligible additional computational cost."
e476fff74e14b5da75a5dc911a180ff830e24efb,2601.18588,10.48550/arXiv.2601.18588,2026,Stability as a Liability:Systematic Breakdown of Linguistic Structure in LLMs,X. Meng,10,arXiv.org,0,"Training stability is typically regarded as a prerequisite for reliable optimization in large language models. In this work, we analyze how stabilizing training dynamics affects the induced generation distribution. We show that under standard maximum likelihood training, stable parameter trajectories lead stationary solutions to approximately minimize the forward KL divergence to the empirical distribution, while implicitly reducing generative entropy. As a consequence, the learned model can concentrate probability mass on a limited subset of empirical modes, exhibiting systematic degeneration despite smooth loss convergence. We empirically validate this effect using a controlled feedback-based training framework that stabilizes internal generation statistics, observing consistent low-entropy outputs and repetitive behavior across architectures and random seeds. It indicates that optimization stability and generative expressivity are not inherently aligned, and that stability alone is an insufficient indicator of generative quality."
52058c41b3b6a1e48018794f206c8f59f27a24a4,2601.04575,10.48550/arXiv.2601.04575,2026,Scaling Behavior Cloning Improves Causal Reasoning: An Open Model for Real-Time Video Game Playing,Yuguang Yue,6,arXiv.org,11,"Behavior cloning has seen a resurgence as scaling model and data sizes demonstrate strong performance. In this work, we introduce an open recipe for training a video game playing foundation model designed for inference in realtime on a consumer GPU. We release all data (8300+ hours of high quality human gameplay), training and inference code, and pretrained checkpoints under an open license. Empirically, we show that our best model achieves performance competitive with human players across a variety of 3D games. We use this recipe to investigate the scaling laws of behavior cloning, with a focus on causal reasoning. In a controlled toy setting, we first demonstrate that increasing training data and network depth leads to the model learning a more causal policy. We then validate these findings at scale, analyzing models up to 1.2 billion parameters. We observe that the causal improvements seen in the toy domain hold true as model size and training steps increase."
652b1768e9627342a1d696d05cf54c08a32afdde,2511.21377,10.48550/arXiv.2511.21377,2025,Controlling changes to attention logits,Ben Anson,2,arXiv.org,3,"Stability of neural network weights is critical when training transformer models. The query and key weights are particularly problematic, as they tend to grow large without any intervention. Applying normalization to queries and keys, known as `QK norm', fixes stability issues in practice, but is not always applicable. For example, QK norm is not compatible with Multi Latent Attention (MLA) because QK norm requires full materialization of queries and keys during inference, which is not done in MLA. In this paper we suggest that controlling the changes to logits is important for stability. We show that these changes are controllable by assigning parameter-dependent learning rates to the query and key weights. We find that our cheap intervention allows us to increase the base learning rate of the network, outperform other methods in the MLA setting, and achieve performance competitive with QK norm when using Multi-head Attention."
1bdb73a64b839b0a708a2a5b0970e177c8494630,2510.04848,10.48550/arXiv.2510.04848,2025,Instability in Downstream Task Performance During LLM Pretraining,Yuto Nishida,3,Conference on Empirical Methods in Natural Language Processing,5,"When training large language models (LLMs), it is common practice to track downstream task performance throughout the training process and select the checkpoint with the highest validation score. However, downstream metrics often exhibit substantial fluctuations, making it difficult to identify the checkpoint that truly represents the best-performing model. In this study, we empirically analyze the stability of downstream task performance in an LLM trained on diverse web-scale corpora. We find that task scores frequently fluctuate throughout training, both at the aggregate and example levels. To address this instability, we investigate two post-hoc checkpoint integration methods: checkpoint averaging and ensemble, motivated by the hypothesis that aggregating neighboring checkpoints can reduce performance volatility. We demonstrate both empirically and theoretically that these methods improve downstream performance stability without requiring any changes to the training procedure."
3d70db88eb672e1869c7793631eae9f3c65a1b0f,2510.04202,10.48550/arXiv.2510.04202,2025,Spectral Alignment as Predictor of Loss Explosion in Neural Network Training,Hai-Quan Qiu,5,arXiv.org,2,"Loss explosions in training deep neural networks can nullify multi-million dollar training runs. Conventional monitoring metrics like weight and gradient norms are often lagging and ambiguous predictors, as their values vary dramatically across different models and even between layers of the same model, making it difficult to establish a unified standard for detecting impending failure. We introduce Spectral Alignment (SA), a novel, theoretically-grounded metric that monitors the distributional alignment between layer inputs and the principal singular vectors of weight matrices. We show that a collapse in the sign diversity of this alignment is a powerful early predictor of representational collapse and training divergence. Empirical results on language models demonstrate that monitoring the SA distribution provides a significantly earlier and clearer warning of loss explosions than traditional scalar metrics. SA's low computational overhead makes it a practical tool for safeguarding model training."
ba5c84133fe893bcbcc398e9df6ec63e98a4a675,,10.1117/12.3050511,2025,LeakGPT: detecting water pipe leaks using vision language models,Li-Xin Tu,8,SPIE Smart Structures and Materials + Nondestructive Evaluation,3,"Aging water infrastructures pose significant risks to water systems, and detecting water pipe leaks is essential because leaks can lead to service shutdowns and disrupt daily life and industrial operations. Among various inspection techniques, acoustic methods have gained attention due to their ability to detect small leaks that other approaches might miss. Typically, audio signals are converted into spectrograms to enhance the leak-related features. Traditionally, human experts manually analyze these spectrograms which is a both time-consuming and labor-intensive process. More recent approaches employ neural networks to analyze spectrograms automatically. However, these methods often produce results that are hard to interpret, making it challenging for human experts to trust their conclusions. In contrast, vision language models (VLMs) have shown strong capabilities in interpreting images, while their potential for analyzing spectrograms remains unexplored. Herein, we fine-tuned a VLM, which we refer to as LeakGPT, using a customized dataset collected from field water pipeline inspection. Since current VLMs lack specialized knowledge in this domain, we address this gap by leveraging a large language model (LLM) to generate question-answer pairs from manually labeled spectrograms, forming a domain-specific dataset. This dataset is then used to fine-tune three foundational VLMs (Llama3.2, Qwen2, and Pixtral) to establish benchmark performance. Experimental results indicate that LeakGPT, fine-tuned on the Pixtral foundation model, can effectively interpret spectrograms and detect water pipe leaks, indicating its potential as a useful tool in water pipe leak inspection."
0ccc4f409c3630c4a3d68170a12fc72e15906559,2503.04598,10.48550/arXiv.2503.04598,2025,HybridNorm: Towards Stable and Efficient Transformer Training via Hybrid Normalization,Zhijian Zhuo,8,Neural Information Processing Systems,25,"Transformers have become the de facto architecture for a wide range of machine learning tasks, particularly in large language models (LLMs). Despite their remarkable performance, many challenges remain in training deep transformer networks, especially regarding the position of the layer normalization. While Pre-Norm structures facilitate more stable training owing to their stronger identity path, they often lead to suboptimal performance compared to Post-Norm. In this paper, we propose $\textbf{HybridNorm}$, a simple yet effective hybrid normalization strategy that integrates the advantages of both Pre-Norm and Post-Norm. Specifically, HybridNorm employs QKV normalization within the attention mechanism and Post-Norm in the feed-forward network (FFN) of each transformer block. We provide both theoretical insights and empirical evidence to demonstrate that HybridNorm improves the gradient flow and the model robustness. Extensive experiments on large-scale transformer models, including both dense and sparse variants, show that HybridNorm consistently outperforms both Pre-Norm and Post-Norm approaches across multiple benchmarks. These findings highlight the potential of HybridNorm as a more stable and effective technique for improving the training and performance of deep transformer models. Code is available at https://github.com/BryceZhuo/HybridNorm."
e6b867fd81eb90c67ee64d9a768d3c5c867c9ab4,2501.18795,10.48550/arXiv.2501.18795,2025,Rope to Nope and Back Again: A New Hybrid Attention Strategy,Bowen Yang,7,Neural Information Processing Systems,52,"Long-context large language models (LLMs) have achieved remarkable advancements, driven by techniques like Rotary Position Embedding (RoPE) (Su et al., 2023) and its extensions (Chen et al., 2023; Liu et al., 2024c; Peng et al., 2023). By adjusting RoPE parameters and incorporating training data with extended contexts, we can train performant models with considerably longer input sequences. However, existing RoPE-based methods exhibit performance limitations when applied to extended context lengths. This paper presents a comprehensive analysis of various attention mechanisms, including RoPE, No Positional Embedding (NoPE), and Query-Key Normalization (QK-Norm), identifying their strengths and shortcomings in long-context modeling. Our investigation identifies distinctive attention patterns in these methods and highlights their impact on long-context performance, providing valuable insights for architectural design. Building on these findings, we propose a novel architecture featuring a hybrid attention mechanism that integrates global and local attention spans. This design not only surpasses conventional RoPE-based transformer models with full attention in both long and short context tasks but also delivers substantial efficiency gains during training and inference."
b8e9ac58409f7dc6096fad17d8444b7f6bf88879,2411.00918,10.48550/arXiv.2411.00918,2024,LIBMoE: A Library for comprehensive benchmarking Mixture of Experts in Large Language Models,Nam V. Nguyen,5,arXiv.org,5,"Mixture of experts (MoE) architectures have become a cornerstone for scaling up and are a key component in most large language models such as GPT-OSS, DeepSeek-V3, Llama-4, and Gemini-2.5. However, systematic research on MoE remains severely constrained by the prohibitive computational costs of training and evaluation, restricting large-scale studies accessible to most researchers. We introduce LibMoE, a unified framework for reproducible, efficient, and extensible MoE research that supports both pretraining and sparse-upcycling regimes. Beyond unified implementations, the framework provides transparent analytical tools for probing routing and expert dynamics. Leveraging this foundation, we conduct a comprehensive analysis along three dimensions: (i) routing dynamics, covering expert selection patterns, routing stability and optimality, and how routing entropy reveals task specialization and expert diversity; (ii) the effect of lightweight initialization on load balancing, demonstrating how subtle changes in router initialization shape early expert utilization; and (iii) training regime differences, revealing how sparse upcycling and full pretraining exhibit distinct routing patterns and stability profiles. By lowering the barrier to entry and standardizing evaluation, along with our comprehensive analysis, LibMoE broadens access to MoE research and establishes a reliable benchmark to guide future innovations. GitHub: \href{https://github.com/Fsoft-AIC/LibMoE}{https://github.com/Fsoft-AIC/LibMoE}."
80d86aea613a91272a27c81c2d6a43763da0bece,2410.13060,,2024,AERO: Entropy-Guided Framework for Private LLM Inference,Nandan Kumar Jha,2,,6,"Privacy-preserving computation enables language model inference directly on encrypted data yet suffers from prohibitive latency and communication overheads, primarily due to nonlinear functions. Removing nonlinearities, however, can trigger one of two failure modes restricting the potential for nonlinearity removal: entropy collapse in deeper layers, which destabilizes training, and entropic overload in early layers, causing under-utilization of attention heads. To address these challenges, we introduce AERO, an entropy-guided framework to strategically eliminates costly nonlinear operations from transformer architectures, which employs an adaptive recalibration through a head-wise entropy regularizer with learnable per-head strengths, enabling each head to adjust its entropy level while penalizing extreme entropies and fostering functional diversity through a tolerance margin. Experiments show AERO can save 3.4$\times$ communication and 1.4$\times$ latency, without any performance penalty."
8cf5cef12b2f86bc037b989e54057d878d0e53cf,,10.48550/arXiv.2606.08105,2026,"A Unifying View of Attention Sinks: Two Algorithms, Two Solutions",Lukas Fesser,5,arXiv.org,2,
88fc6f475d9ea9cae96972e0afb326ff6b4f826a,,,,Variance Sensitivity Induces Attention Entropy Collapse in Transformers,J. Hong,2,,0,
1 s2_id arxiv doi year title first_author n_authors venue citations abstract
2 9653176d4dafd012f0b04775889b80cc7b8c6b91 2608.02091 2026 One QK Channel, Many Sources: Tracing Low-Precision Attention Collapse Shu-Xiao Xie 6 0 A bfloat16 transformer can train normally, then collapse abruptly. Prior work links collapse to structured attention errors and shows QK normalization disrupts their compounding. Distinct low-precision errors trigger the same collapse, leaving unclear whether each needs a fix at its source or one shared route can be blocked instead. We isolated the fault behind a reproduced GPT-2-class collapse to the streaming-softmax accumulator, where an fp32 streaming core repairs it, and turned it into an assay for moving a controlled error across sources. Using it, we found that errors placed outside attention still drove the same QK spectral runaway, and that correcting only QK kept training stable while the fault stayed active. This is a source-channel dissociation: fault source is not failure channel. It held across tested architectures and scales, and reproduced on a second GPU architecture. As a causal probe, projecting each update off the current QK weights'leading three singular directions held the query projection's largest singular value to 11.1, whereas removing equal energy elsewhere left it at 237: the QK channel causally drives the early runaway. What lets the injected error in is temporal sign-coherence, its per-head sign persisting across steps, not aggregate deviation; once inside, the runaway shows as attention-logit saturation. QK-Guard, a dormant controller, tests this by switching on parameter-free QK normalization at the first monitored threshold crossing. On the runs designated for this test, the QK-local action prevented the failure of each matched or same-configuration unguarded run; on plain GPT-2, all 12 final train and validation losses were within 0.03 nat of same-configuration always-on QK-norm, and both methods ran 60k steps without collapse. Intervention at the QK locus therefore suffices in place of a fix at each source.
3 cacb1f26dd4ca32fb6663fb7ca0489a05b598a6e 2606.08105 2026 A Unifying View of Attention Sinks: From Mechanisms to Architectural Interventions Lukas Fesser 5 1 When attention concentrates on a single token, a sink, what is the model actually computing? Attention sinks are ubiquitous in softmax transformers, yet this shared visual signature can hide fundamentally different algorithms. We show that visually similar sink patterns can reflect two distinct mechanisms: (i) adaptive nop, where a head suppresses its update by routing to a null token, and (ii) broadcast, where a sink aggregates and redistributes global information. Each mechanism leaves distinct traces (nop-sinks exhibit negligible value norms; broadcast sinks induce low-rank outputs), which we formalize on synthetic tasks and use to derive practical diagnostics. Applied to pretrained vision transformers, these diagnostics reveal that both mechanisms exist at scale: sinks transition from CLS in early layers to patches in deeper layers and concentrate in specialized heads. Causal interventions further connect these signatures to near-null suppression and shared residual contributions. We then use architectural interventions to show how these computations can be reorganized: gating eliminates detected nop-like sinks but increases broadcast-like sinks, registers relocate rather than remove sink computation, and our position-free global pathway provides an explicit route for shared communication that reduces the broadcast-like sinks induced by gating. On dense probes, combining gating with the global pathway gives the strongest results among the tested variants, despite retaining some broadcast-like sinks. Overall, we find that the same attention pattern can reflect two very different computations, and that effective intervention depends not only on identifying the computation, but also on providing architectural alternatives through which the model can reorganize it.
4 dd91f0cdb8cc1da2b7c211fcaac523d14834eea1 2605.12492 10.48550/arXiv.2605.12492 2026 Pion: A Spectrum-Preserving Optimizer via Orthogonal Equivalence Transformation Kexuan Shi 6 arXiv.org 6 We introduce Pion, a spectrum-preserving optimizer for large language model (LLM) training based on orthogonal equivalence transformation. Unlike additive optimizers such as Adam and Muon, Pion updates each weight matrix through left and right orthogonal transformations, preserving its singular values throughout training. This yields an optimization mechanism that modulates the geometry of weight matrices while keeping their spectral norm fixed. We derive the Pion update rule, systematically examine its design choices, and analyze its convergence behavior along with several key properties. Empirical results show that Pion offers a stable and competitive alternative to standard optimizers for both LLM pretraining and finetuning.
5 3b07bb0e7ff9068eae67d7a4867c4a72518d5df0 2601.22095 10.48550/arXiv.2601.22095 2026 GeoNorm: Unify Pre-Norm and Post-Norm with Geodesic Optimization Chuan-Yang Zheng 14 arXiv.org 1 The placement of normalization layers, specifically Pre-Norm and Post-Norm, remains an open question in Transformer architecture design. In this work, we rethink these approaches through the lens of manifold optimization, interpreting the outputs of the Feed-Forward Network (FFN) and attention layers as update directions in optimization. Building on this perspective, we introduce GeoNorm, a novel method that replaces standard normalization with geodesic updates on the manifold. Furthermore, analogous to learning rate schedules, we propose a layer-wise update decay for the FFN and attention components. Comprehensive experiments demonstrate that GeoNorm consistently outperforms existing normalization methods in Transformer models. Crucially, GeoNorm can be seamlessly integrated into standard Transformer architectures, achieving performance improvements with negligible additional computational cost.
6 e476fff74e14b5da75a5dc911a180ff830e24efb 2601.18588 10.48550/arXiv.2601.18588 2026 Stability as a Liability:Systematic Breakdown of Linguistic Structure in LLMs X. Meng 10 arXiv.org 0 Training stability is typically regarded as a prerequisite for reliable optimization in large language models. In this work, we analyze how stabilizing training dynamics affects the induced generation distribution. We show that under standard maximum likelihood training, stable parameter trajectories lead stationary solutions to approximately minimize the forward KL divergence to the empirical distribution, while implicitly reducing generative entropy. As a consequence, the learned model can concentrate probability mass on a limited subset of empirical modes, exhibiting systematic degeneration despite smooth loss convergence. We empirically validate this effect using a controlled feedback-based training framework that stabilizes internal generation statistics, observing consistent low-entropy outputs and repetitive behavior across architectures and random seeds. It indicates that optimization stability and generative expressivity are not inherently aligned, and that stability alone is an insufficient indicator of generative quality.
7 52058c41b3b6a1e48018794f206c8f59f27a24a4 2601.04575 10.48550/arXiv.2601.04575 2026 Scaling Behavior Cloning Improves Causal Reasoning: An Open Model for Real-Time Video Game Playing Yuguang Yue 6 arXiv.org 11 Behavior cloning has seen a resurgence as scaling model and data sizes demonstrate strong performance. In this work, we introduce an open recipe for training a video game playing foundation model designed for inference in realtime on a consumer GPU. We release all data (8300+ hours of high quality human gameplay), training and inference code, and pretrained checkpoints under an open license. Empirically, we show that our best model achieves performance competitive with human players across a variety of 3D games. We use this recipe to investigate the scaling laws of behavior cloning, with a focus on causal reasoning. In a controlled toy setting, we first demonstrate that increasing training data and network depth leads to the model learning a more causal policy. We then validate these findings at scale, analyzing models up to 1.2 billion parameters. We observe that the causal improvements seen in the toy domain hold true as model size and training steps increase.
8 652b1768e9627342a1d696d05cf54c08a32afdde 2511.21377 10.48550/arXiv.2511.21377 2025 Controlling changes to attention logits Ben Anson 2 arXiv.org 3 Stability of neural network weights is critical when training transformer models. The query and key weights are particularly problematic, as they tend to grow large without any intervention. Applying normalization to queries and keys, known as `QK norm', fixes stability issues in practice, but is not always applicable. For example, QK norm is not compatible with Multi Latent Attention (MLA) because QK norm requires full materialization of queries and keys during inference, which is not done in MLA. In this paper we suggest that controlling the changes to logits is important for stability. We show that these changes are controllable by assigning parameter-dependent learning rates to the query and key weights. We find that our cheap intervention allows us to increase the base learning rate of the network, outperform other methods in the MLA setting, and achieve performance competitive with QK norm when using Multi-head Attention.
9 1bdb73a64b839b0a708a2a5b0970e177c8494630 2510.04848 10.48550/arXiv.2510.04848 2025 Instability in Downstream Task Performance During LLM Pretraining Yuto Nishida 3 Conference on Empirical Methods in Natural Language Processing 5 When training large language models (LLMs), it is common practice to track downstream task performance throughout the training process and select the checkpoint with the highest validation score. However, downstream metrics often exhibit substantial fluctuations, making it difficult to identify the checkpoint that truly represents the best-performing model. In this study, we empirically analyze the stability of downstream task performance in an LLM trained on diverse web-scale corpora. We find that task scores frequently fluctuate throughout training, both at the aggregate and example levels. To address this instability, we investigate two post-hoc checkpoint integration methods: checkpoint averaging and ensemble, motivated by the hypothesis that aggregating neighboring checkpoints can reduce performance volatility. We demonstrate both empirically and theoretically that these methods improve downstream performance stability without requiring any changes to the training procedure.
10 3d70db88eb672e1869c7793631eae9f3c65a1b0f 2510.04202 10.48550/arXiv.2510.04202 2025 Spectral Alignment as Predictor of Loss Explosion in Neural Network Training Hai-Quan Qiu 5 arXiv.org 2 Loss explosions in training deep neural networks can nullify multi-million dollar training runs. Conventional monitoring metrics like weight and gradient norms are often lagging and ambiguous predictors, as their values vary dramatically across different models and even between layers of the same model, making it difficult to establish a unified standard for detecting impending failure. We introduce Spectral Alignment (SA), a novel, theoretically-grounded metric that monitors the distributional alignment between layer inputs and the principal singular vectors of weight matrices. We show that a collapse in the sign diversity of this alignment is a powerful early predictor of representational collapse and training divergence. Empirical results on language models demonstrate that monitoring the SA distribution provides a significantly earlier and clearer warning of loss explosions than traditional scalar metrics. SA's low computational overhead makes it a practical tool for safeguarding model training.
11 ba5c84133fe893bcbcc398e9df6ec63e98a4a675 10.1117/12.3050511 2025 LeakGPT: detecting water pipe leaks using vision language models Li-Xin Tu 8 SPIE Smart Structures and Materials + Nondestructive Evaluation 3 Aging water infrastructures pose significant risks to water systems, and detecting water pipe leaks is essential because leaks can lead to service shutdowns and disrupt daily life and industrial operations. Among various inspection techniques, acoustic methods have gained attention due to their ability to detect small leaks that other approaches might miss. Typically, audio signals are converted into spectrograms to enhance the leak-related features. Traditionally, human experts manually analyze these spectrograms which is a both time-consuming and labor-intensive process. More recent approaches employ neural networks to analyze spectrograms automatically. However, these methods often produce results that are hard to interpret, making it challenging for human experts to trust their conclusions. In contrast, vision language models (VLMs) have shown strong capabilities in interpreting images, while their potential for analyzing spectrograms remains unexplored. Herein, we fine-tuned a VLM, which we refer to as LeakGPT, using a customized dataset collected from field water pipeline inspection. Since current VLMs lack specialized knowledge in this domain, we address this gap by leveraging a large language model (LLM) to generate question-answer pairs from manually labeled spectrograms, forming a domain-specific dataset. This dataset is then used to fine-tune three foundational VLMs (Llama3.2, Qwen2, and Pixtral) to establish benchmark performance. Experimental results indicate that LeakGPT, fine-tuned on the Pixtral foundation model, can effectively interpret spectrograms and detect water pipe leaks, indicating its potential as a useful tool in water pipe leak inspection.
12 0ccc4f409c3630c4a3d68170a12fc72e15906559 2503.04598 10.48550/arXiv.2503.04598 2025 HybridNorm: Towards Stable and Efficient Transformer Training via Hybrid Normalization Zhijian Zhuo 8 Neural Information Processing Systems 25 Transformers have become the de facto architecture for a wide range of machine learning tasks, particularly in large language models (LLMs). Despite their remarkable performance, many challenges remain in training deep transformer networks, especially regarding the position of the layer normalization. While Pre-Norm structures facilitate more stable training owing to their stronger identity path, they often lead to suboptimal performance compared to Post-Norm. In this paper, we propose $\textbf{HybridNorm}$, a simple yet effective hybrid normalization strategy that integrates the advantages of both Pre-Norm and Post-Norm. Specifically, HybridNorm employs QKV normalization within the attention mechanism and Post-Norm in the feed-forward network (FFN) of each transformer block. We provide both theoretical insights and empirical evidence to demonstrate that HybridNorm improves the gradient flow and the model robustness. Extensive experiments on large-scale transformer models, including both dense and sparse variants, show that HybridNorm consistently outperforms both Pre-Norm and Post-Norm approaches across multiple benchmarks. These findings highlight the potential of HybridNorm as a more stable and effective technique for improving the training and performance of deep transformer models. Code is available at https://github.com/BryceZhuo/HybridNorm.
13 e6b867fd81eb90c67ee64d9a768d3c5c867c9ab4 2501.18795 10.48550/arXiv.2501.18795 2025 Rope to Nope and Back Again: A New Hybrid Attention Strategy Bowen Yang 7 Neural Information Processing Systems 52 Long-context large language models (LLMs) have achieved remarkable advancements, driven by techniques like Rotary Position Embedding (RoPE) (Su et al., 2023) and its extensions (Chen et al., 2023; Liu et al., 2024c; Peng et al., 2023). By adjusting RoPE parameters and incorporating training data with extended contexts, we can train performant models with considerably longer input sequences. However, existing RoPE-based methods exhibit performance limitations when applied to extended context lengths. This paper presents a comprehensive analysis of various attention mechanisms, including RoPE, No Positional Embedding (NoPE), and Query-Key Normalization (QK-Norm), identifying their strengths and shortcomings in long-context modeling. Our investigation identifies distinctive attention patterns in these methods and highlights their impact on long-context performance, providing valuable insights for architectural design. Building on these findings, we propose a novel architecture featuring a hybrid attention mechanism that integrates global and local attention spans. This design not only surpasses conventional RoPE-based transformer models with full attention in both long and short context tasks but also delivers substantial efficiency gains during training and inference.
14 b8e9ac58409f7dc6096fad17d8444b7f6bf88879 2411.00918 10.48550/arXiv.2411.00918 2024 LIBMoE: A Library for comprehensive benchmarking Mixture of Experts in Large Language Models Nam V. Nguyen 5 arXiv.org 5 Mixture of experts (MoE) architectures have become a cornerstone for scaling up and are a key component in most large language models such as GPT-OSS, DeepSeek-V3, Llama-4, and Gemini-2.5. However, systematic research on MoE remains severely constrained by the prohibitive computational costs of training and evaluation, restricting large-scale studies accessible to most researchers. We introduce LibMoE, a unified framework for reproducible, efficient, and extensible MoE research that supports both pretraining and sparse-upcycling regimes. Beyond unified implementations, the framework provides transparent analytical tools for probing routing and expert dynamics. Leveraging this foundation, we conduct a comprehensive analysis along three dimensions: (i) routing dynamics, covering expert selection patterns, routing stability and optimality, and how routing entropy reveals task specialization and expert diversity; (ii) the effect of lightweight initialization on load balancing, demonstrating how subtle changes in router initialization shape early expert utilization; and (iii) training regime differences, revealing how sparse upcycling and full pretraining exhibit distinct routing patterns and stability profiles. By lowering the barrier to entry and standardizing evaluation, along with our comprehensive analysis, LibMoE broadens access to MoE research and establishes a reliable benchmark to guide future innovations. GitHub: \href{https://github.com/Fsoft-AIC/LibMoE}{https://github.com/Fsoft-AIC/LibMoE}.
15 80d86aea613a91272a27c81c2d6a43763da0bece 2410.13060 2024 AERO: Entropy-Guided Framework for Private LLM Inference Nandan Kumar Jha 2 6 Privacy-preserving computation enables language model inference directly on encrypted data yet suffers from prohibitive latency and communication overheads, primarily due to nonlinear functions. Removing nonlinearities, however, can trigger one of two failure modes restricting the potential for nonlinearity removal: entropy collapse in deeper layers, which destabilizes training, and entropic overload in early layers, causing under-utilization of attention heads. To address these challenges, we introduce AERO, an entropy-guided framework to strategically eliminates costly nonlinear operations from transformer architectures, which employs an adaptive recalibration through a head-wise entropy regularizer with learnable per-head strengths, enabling each head to adjust its entropy level while penalizing extreme entropies and fostering functional diversity through a tolerance margin. Experiments show AERO can save 3.4$\times$ communication and 1.4$\times$ latency, without any performance penalty.
16 8cf5cef12b2f86bc037b989e54057d878d0e53cf 10.48550/arXiv.2606.08105 2026 A Unifying View of Attention Sinks: Two Algorithms, Two Solutions Lukas Fesser 5 arXiv.org 2
17 88fc6f475d9ea9cae96972e0afb326ff6b4f826a Variance Sensitivity Induces Attention Entropy Collapse in Transformers J. Hong 2 0