Deploying high-accuracy neural networks on resource-constrained edge devices remains challenging, as existing approaches treat training, compression, and hardware synthesis as separate stages, leaving a gap between software-trained models and efficient end-to-end deployment with limited support for interpretability. We propose Bern2Edge, an end-to-end framework that uses knowledge distillation to convert a pretrained teacher feed-forward network into hardware-efficient representations via Bernstein polynomial activations. This representation enables two deployment paths: (i) a high-fidelity LUT-based realization that preserves model fidelity under compression, and (ii) a symbolic rule-based representation derived from Bernstein activation geometry, enabling interpretable inference with explicit input-space constraints. The resulting BNNs achieve up to 2.12 percentage-point (pp) accuracy improvement over ReLU under identical compression constraints. At the system level, Bern2Edge achieves up to 99.8% latency reduction and 95.2% BRAM reduction relative to a W8A8 quantized teacher on an AMD Xilinx KV260 FPGA, while maintaining accuracy within 0.5 pp, and further deploys on a low-power Spartan-7 XC7S15 FPGA. The rule-based path reduces DSP usage by up to 89.0% at a cost of 1.5 pp in total accuracy.
Top-k Mixture-of-Experts (MoE) routing is discontinuous, so a deployment-motivated numerical disturbance -- simulated 4-bit KV-cache quantization read by a protected BF16 gate -- pushes tokens across decision boundaries and flips which experts fire. This paper proposes no new mitigation; it supplies a causal apparatus, empirical findings, and a detection-limit result. A four-run apparatus prices the route-mediated fraction (RMF) of quantization damage, a token-level attribution decomposes it by mechanism, and pre-registered probes carry the findings across three architectures. On OLMoE-1B-7B at 4-bit KV (pilot), about a third of the damage is routing-mediated: RMF ~ 0.31 (discovery 0.31 [0.20, 0.41]; process-replicated mean 0.313 +/- 0.020; pre-registered re-execution 0.231). The deployable router margin detects that a flip occurred (AUC 0.772) but cannot tell a harmful flip from a helpful one (at chance): among the tested local, inference-observable router statistics we find no predictor of a flip's loss sign above chance -- an empirical benefit-detection barrier bounding selective repair restricted to this feature family. The signed-flip tax and sign-inseparability carry cross-model; the clean-reference remedy's payout is architecture-modulated; a controlled same-checkpoint flag-swap re-scopes the gate's normalization convention to a damage-magnitude moderator, not a route-recoverability mechanism. A real int4 KV kernel yields a fraction compatible with the fake-quant dose curve but underpowered (95% CI [-0.111, 0.394] includes zero) -- ruling out gross disagreement, not an independent replication. Hypotheses, thresholds, and evaluations were pre-registered before measurement, with misses reported; a pre-registered held-out read replicates the partition and the near-cancelling tax out of sample, while the strict impossibility exclusion narrowly misses.
Pietro Tropeano, Maria Maistro, Tuukka Ruotsalo +1cs.LG
Pruning Large Language Models (LLMs) reduces memory and inference costs by removing parts of the network, producing smaller models that retain most of their accuracy. As attention layers are the most resource-intensive parts of LLMs, pruning them is a promising compression strategy. Prior work shows that up to 33% of attention layers can be pruned with minimal accuracy loss. Nevertheless, the impact of attention pruning on model interpretability, specifically faithfulness and confidence calibration, remains unstudied. To address this gap, we study how pruning attention layers affects explanation faithfulness and confidence calibration across five LLMs and eight datasets. While the pruned models often maintain high accuracy, we find that their faithfulness and calibration often degrade. Notably, faithfulness and calibration can fluctuate significantly, even when accuracy remains stable, highlighting a misalignment between model confidence, interpretability, and accuracy. Our findings suggest that layer pruning can affect LLMs' interpretability and reliability in ways not captured by accuracy and efficiency measures alone. We recommend including explainability and calibration metrics when evaluating pruned models.
Lucia Quirke, Louis Jaburi, David Johnston +6cs.LG
Data attribution is a promising field in interpretability that aims to explain model behavior through the influence of its training data, with applications including debugging undesirable model behavior and training dataset curation. However, significant engineering effort is required to perform it at scale, and many cutting edge techniques lack open-source tooling and support. Bergson is an open source library that aims to enable faster progress in the field by providing a host of techniques that scale to very large language models and pre-training datasets. The library natively supports on-disk gradient stores and multi-node distributed training, and provides quality of life tools for researchers. Finally, we introduce the first open-source implementations of three leading data attribution methods: MAGIC, SOURCE, and TrackStar. The library is available at https://github.com/EleutherAI/bergson .
Ian Seet, Jonas Bozenhard, Simon Ostermanncs.LG cs.AI
Recent advances in generative AI, especially powerful Large Language Models (LLMs), raise concerns over the interpretability, safety and sustainability of these large and opaque AI models. The power of such architectures is derived not only from the scalability of deep neural networks (DNNs), but also massively parallel hardware such as GPU clusters. The diffuse nature of DNNs gives them great function-approximation capability when provided with sufficient training data but imposes a cost in interpretability and computational efficiency. Observing that localised machine learning (ML) models tend to be more interpretable and computationally efficient than DNNs on small datasets, we reason by analogy that similar advantages may apply to specific localised hardware ML architectures. We argue that localised architectures with lower bandwidth but higher expressivity per node have the potential to be fundamentally more interpretable than DNNs running on GPU clusters while remaining competitive for smaller datasets. We then survey the suitability of various hardware ML paradigms for implementing such localised architectures and evaluate their per-node expressivity, energy efficiency and practical maturity of the technology required.
Quantization is a standard path to deploying large language models, and quantized models are typically judged acceptable when perplexity or downstream accuracy remains close to the full-precision original. But behavioral parity need not imply feature fidelity: the sparse-autoencoder (SAE) features used to interpret a full-precision model may change after weight rounding. We test this directly by using a frozen SAE as a fixed measurement basis, encoding full-precision and round-to-nearest (RTN) quantized activations on identical tokens, and measuring per-feature survival by Pearson correlation across bit-widths from INT8 to INT4 on Pythia-70M and Gemma-2-2B. Our central finding is that perplexity can miss feature damage: on Gemma-2-2B, INT7 improves perplexity while degrading 18.7% of active SAE features, and under sliding-window evaluation INT6 also improves perplexity while only 51.3% of active features survive. Feature survival is graded rather than cliff-like, with 62.4% of active Pythia features and 51.3% of active Gemma features surviving at INT6; most non-surviving features are blurred rather than fully damaged. Survival is also predictable from full-precision feature statistics alone, with cross-validated AUC 0.92--0.97 and peak activation as the strongest marginal predictor. Finally, RTN quantization and matched-perplexity magnitude pruning damage strongly overlapping feature sets, with Jaccard overlap 0.79--0.86 and damage-score Spearman correlation 0.98. These results show that behavioral metrics alone are insufficient evidence that full-precision interpretability findings transfer to quantized models, motivating feature-level audits of compression.
Transformer models are widely deployed in critical AI applications, yet faults in their attention mechanisms, projections, and other internal components often degrade behavior silently without raising runtime errors. Existing fault diagnosis techniques often target generic deep neural networks and cannot identify which transformer component is responsible for an observed symptom. In this article, we present DEFault++, a hierarchical learning-based diagnostic technique that operates at three level of abstraction: it detects whether a fault is present, classifies it into one of 12 transformer-specific fault categories (covering both attention-internal mechanisms and surrounding architectural components), and identifies the underlying root cause from up to 45 mechanisms. To facilitate both training and evaluation, we construct DEFault-bench, a benchmark of 3,739 labeled instances obtained through systematic mutation testing. These instances are created across seven transformer models and nine downstream tasks using DEForm, a transformer-specific mutation technique we developed for this purpose. DEFault++ measures runtime behavior at the level of individual transformer components. It organizes these measurements through a Fault Propagation Graph (FPG) derived from the transformer architecture. It then produces an interpretable diagnosis using prototype matching combined with supervised contrastive learning. On DEFault-bench, DEFault++ exceeds an AUROC of 0.96 for detection and a Macro-F1 of 0.85 for both categorization and root-cause diagnosis on encoder and decoder architectures. In a developer study with 21 practitioners, the accuracy of choosing correct repair actions increased from 57.1% without support to 83.3% when using DEFault++.