arXiv subjects

Orion Reblitz-Richardson

Publications and source records attributed to Orion Reblitz-Richardson.

11 recordsLinked to original sources

arXiv2026

Calibrating Interpretability Instruments Before Trusting Their Verdicts

Causal claims about large language model (LLM) internals rest on measurements. Those might include a projection, a cosine, an ablation delta, or an interchange patch among others. These measurements fail in specific, diagnosable ways that return a plausible number instead of an error, so a broken instrument can easily read as a finding. A covariance-matched null can saturate until every direction looks typical, a per-head attribution can overshoot the true residual write threefold on reordered-normalization architectures, an interchange patch can go sign-chaotic because its outcome is pinned at a ceiling, or a read-from verdict can be an artifact of measuring past the layer where the model already decided. This note documents six such failures from a causal interpretability program on refusal and moral representation, spanning several papers and a four-model open-weight panel; each mode is established on one or two of the four. For each we give the tell that catches it and a protocol keyed to a detectable trigger (reordered normalization, massive activations, a low-dimensional decision channel), so we and readers can check whether a given setup is exposed. The discipline reduces to four moves: calibrate against a positive-control ladder, certify with an orthogonal cell, compute power before spending compute, and state every read-from verdict at a depth referenced to the model's commitment. The evidence is four architectures across three families within a single program; external replication across programs is future work.

cs.LG↗

arXiv2026

Refusal Reads Only a Slice of What the Model Knows: Harm-Keyed Routing and Its Exceptions Across Model Families

Alignment applied after pretraining is shallow in a measurable way: a single direction in a model's residual stream can be edited out, and the model stops refusing harmful requests. That fact says how easily refusal can be removed, not what the refusal decision was reading in the first place. We ask what it reads, and we separate that from what the model comprehends. Across four open-weight models spanning three families, moral comprehension is native to pretraining: a low-rank moral subspace crystallizes during pretraining, and alignment rotates it once without rebuilding it. The refusal gate, in contrast, is a fresh post-training construction with only a weak pretraining precursor, written into a narrow control-token channel where the refusal decision is orthogonal to the moral-judgment decision. The central result is causal and comes from one model, OLMo-3. A nested interchange rank sweep patches successively larger slices of the moral subspace between matched requests and reads how much of refusal's response transfers: as the basis widens, moral judgment keeps reading more of it, while refusal levels off at the level of a single harm direction, and about three-quarters of refusal's causal input lies outside the moral subspace altogether. Refusal reads the harm percept, not the moral content that judgment reads on the same patches. The picture is not uniform across families. Llama reads broad moral content; Qwen reads beyond the single harm cue but is unresolved at our sample size; GPT-OSS reads harm, and its refusals can be argued in either direction by its own reasoning trace. Where refusal reads only a low-rank slice and routes around the bulk of what the model knows, a rank-one edit removes it. Whether widening what refusal reads would also deepen the behavior is the open question this raises.

cs.LG↗

arXiv2026

How Language Models Organize and Structure Moral Knowledge

How do large language models (LLMs) organize moral knowledge? Models detect moral content broadly, but detection is a low bar. We ask whether they go further, distinguishing moral foundations from one another and organizing the relationships between them geometrically. We train six independent linear probes on open-weight language models, one per Moral Foundations Theory (MFT) category (care/harm, fair/cheat, lib/oppress, loy/betray, auth/subv, sanc/degrade), and examine how the resulting directions relate to each other in representation space. We find the directions neither collapse into a single moral detector nor isolate from one another. Rather, they span a near-maximal number of independent dimensions while sharing a positive common component. The shared component is the signature of integration, and it is moral-specific relative to a matched non-moral concept battery built identically (mean pairwise cosine 0.26 vs. 0.013). The geometry is consistent across architectures and scale and reaches its integration regime early in pre-training, well before probe accuracy saturates. The structure the model discovers shows no evidence of the individualizing/binding distinction predicted by Moral Foundations Theory (an underpowered test: only 20 candidate partitions exist) but rather reflects corpus statistics. Extending to moral dilemmas, each dilemma direction partially composes from its component foundations, at 2.7x a mismatched-pair baseline, while the majority of its variance encodes conflict-specific structure. The model represents moral tension itself, not a pre-resolved judgment.

cs.CL↗

arXiv2026

When Probing Accuracy Saturates, Fragility Resolves: A Complementary Metric for LLM Pre-Training Analysis

Standard linear probing declares a property "encoded" when a classifier on hidden states achieves high accuracy. The protocol works well on a snapshot but breaks across pre-training, with probe accuracy saturating within the first few thousand steps, leaving most of training invisible to the instrument. We introduce fragility, a complementary per-layer metric defined as the activation-noise level at which probe accuracy collapses. Fragility is sensitive to both the margin of separability and the redundancy of representation, both of which keep evolving long after accuracy plateaus. Applied to open-checkpoint language models, fragility recovers structure that accuracy alone cannot see. Moralized representations, our interest, emerge in stages, with high-accuracy detection of morally loaded words first and compositional encoding later. Because probe accuracy on its own tracks how lexically separable a dataset is, we establish the compositional encoding directly, by showing it transfers across construction types that share no contrast tokens. Where probing accuracy returns a flat answer, fragility returns a structured one. Raw critical noise also shows a layer-depth gradient, but we identify and correct a confound that, to our knowledge, has gone undocumented in noise-injection probing. That is, deeper layers carry larger activations, so a fixed absolute noise makes them look more robust than they are. Rescaling per layer removes most of the gradient, while within-layer comparisons stay valid and still separate identical-accuracy corpora by fragility. We recommend RMS-normalized critical noise for cross-layer claims and raw critical noise for within-layer comparisons.

cs.CL↗

arXiv2026

Output Dilution: Redundant but Fragile Representations in MoE Models

Mixture-of-Experts (MoE) models appear to encode moral content as robustly as dense models, yet prove far more fragile in their encoding. In OLMoE-1B-7B, linear probes recover moral valence from nearly every expert-layer combination, with mean peak-layer accuracy above 90%. But these representations collapse under levels of activation noise that a dense model of matched size easily tolerates, with a 4.2-fold difference in robustness. We trace this to output dilution. Because the MoE block averages across active experts before contributing to the residual stream, the feedforward signal reaching downstream layers is nearly two orders of magnitude smaller than in a dense MLP. Moral information, our interest, survives aggregation intact but at a scale trivially overwhelmed by perturbation. Routing itself remains stable under noise while the vulnerability originates entirely in the diluted aggregate. Checkpoint trajectories confirm this is architectural, not learned. Experts never specialize and accuracy saturates within the first few thousand steps. In sparse architectures, redundant encoding does not imply robust encoding.

cs.LG↗

arXiv2026

ExecuTorch -- A Unified PyTorch Solution to Run AI Models On-Device

Local execution of AI on edge devices is important for low latency and offline operation. However, deploying models on diverse hardware remains fragmented, often requiring model conversion or complete reimplementation outside the PyTorch ecosystem where the model was originally authored. We introduce ExecuTorch, a unified PyTorch-native deployment framework for edge AI. ExecuTorch enables seamless deployment of machine learning models across heterogeneous compute environments. It scales from embedded microcontrollers to complex system-on-chips (SoCs) with dedicated accelerators, powering devices ranging from wearables and smartphones to large compute clusters. ExecuTorch preserves PyTorch semantics while allowing customization, support for optimizations like quantization, and pluggable execution "backends". These features together enable fast experimentation, allowing researchers to validate deployment behavior entirely within PyTorch, bridging the gap between research and production.

cs.LG↗

arXiv2022

A Tour of Visualization Techniques for Computer Vision Datasets

We survey a number of data visualization techniques for analyzing Computer Vision (CV) datasets. These techniques help us understand properties and latent patterns in such data, by applying dataset-level analysis. We present various examples of how such analysis helps predict the potential impact of the dataset properties on CV models and informs appropriate mitigation of their shortcomings. Finally, we explore avenues for further visualization techniques of different modalities of CV datasets as well as ones that are tailored to support specific CV tasks and analysis needs.

cs.CV↗

arXiv2021

Investigating sanity checks for saliency maps with image and text classification

Saliency maps have shown to be both useful and misleading for explaining model predictions especially in the context of images. In this paper, we perform sanity checks for text modality and show that the conclusions made for image do not directly transfer to text. We also analyze the effects of the input multiplier in certain saliency maps using similarity scores, max-sensitivity and infidelity evaluation metrics. Our observations reveal that the input multiplier carries input's structural patterns in explanation maps, thus leading to similar results regardless of the choice of model parameters. We also show that the smoothness of a Neural Network (NN) function can affect the quality of saliency-based explanations. Our investigations reveal that replacing ReLUs with Softplus and MaxPool with smoother variants such as LogSumExp (LSE) can lead to explanations that are more reliable based on the infidelity evaluation metric.

cs.LG↗

arXiv2020

Investigating Saturation Effects in Integrated Gradients

Integrated Gradients has become a popular method for post-hoc model interpretability. De-spite its popularity, the composition and relative impact of different regions of the integral path are not well understood. We explore these effects and find that gradients in saturated regions of this path, where model output changes minimally, contribute disproportionately to the computed attribution. We propose a variant of IntegratedGradients which primarily captures gradients in unsaturated regions and evaluate this method on ImageNet classification networks. We find that this attribution technique shows higher model faithfulness and lower sensitivity to noise com-pared with standard Integrated Gradients. A note-book illustrating our computations and results is available at https://github.com/vivekmig/captum-1/tree/ExpandedIG.

cs.CV↗

arXiv2020

Mind the Pad -- CNNs can Develop Blind Spots

We show how feature maps in convolutional networks are susceptible to spatial bias. Due to a combination of architectural choices, the activation at certain locations is systematically elevated or weakened. The major source of this bias is the padding mechanism. Depending on several aspects of convolution arithmetic, this mechanism can apply the padding unevenly, leading to asymmetries in the learned weights. We demonstrate how such bias can be detrimental to certain tasks such as small object detection: the activation is suppressed if the stimulus lies in the impacted area, leading to blind spots and misdetection. We propose solutions to mitigate spatial bias and demonstrate how they can improve model accuracy.

cs.CV↗

arXiv2020

Captum: A unified and generic model interpretability library for PyTorch

In this paper we introduce a novel, unified, open-source model interpretability library for PyTorch [12]. The library contains generic implementations of a number of gradient and perturbation-based attribution algorithms, also known as feature, neuron and layer importance algorithms, as well as a set of evaluation metrics for these algorithms. It can be used for both classification and non-classification models including graph-structured models built on Neural Networks (NN). In this paper we give a high-level overview of supported attribution algorithms and show how to perform memory-efficient and scalable computations. We emphasize that the three main characteristics of the library are multimodality, extensibility and ease of use. Multimodality supports different modality of inputs such as image, text, audio or video. Extensibility allows adding new algorithms and features. The library is also designed for easy understanding and use. Besides, we also introduce an interactive visualization tool called Captum Insights that is built on top of Captum library and allows sample-based model debugging and visualization using feature importance metrics.

cs.LG↗