arXiv subjects

Dmitrii Troitskii

Publications and source records attributed to Dmitrii Troitskii.

5 recordsLinked to original sources

arXiv2026

Shortcuts in the Tail: Debiasing via Post-Hoc Spectral Compression of Fine-Tuning Updates

Fine-tuning often introduces spurious correlations alongside task knowledge, causing systematic failures on underrepresented groups. Existing mitigations require retraining, group labels, or curated counterfactual data. We show a simple post-hoc intervention reduces shortcut reliance without any of these: truncating the tail of the SVD of $ΔW = W_\mathrm{ft} - W_\mathrm{base}$ reduces the spurious-group gap while preserving task accuracy. Across three instruction-tuned models ($0.5$B--$7$B) and four classification benchmarks, top-$k$ truncation reduces the gap on every cell at $<2$ pp accuracy loss, by up to $5\times$ on CivilComments. We propose this works because the shortcut response sits in the tail of the singular ordering of $ΔW$, a claim about how truncation behaves rather than about the raw singular values, which are broadly distributed and look the same across all four datasets. A controlled boundary case in which fine-tuning has only a shortcut to learn shows the predicted FT-to-base collapse, and bottom-/random-$k$ and matched-rank LoRA controls rule out generic low-rank approximation and rank-constrained training as the explanation. We read this as preliminary evidence that the singular basis of $ΔW$ is a useful coordinate system for studying what fine-tuning has learned.

cs.LG↗

arXiv2026

Improving Evaluation Realism with Inference-Time Compute and Deployment Scaffolds

A core obstacle to alignment evaluation is evaluation awareness: capable models can tell when they are being tested rather than deployed, weakening the conclusions a safety evaluation can support. We present two techniques that make simulated alignment evaluations harder to distinguish from real deployments. Our first technique, critique refinement, spends additional inference-time compute on each simulator action: the simulator generates multiple candidate actions, refines them using feedback from an instance of the target model on how to make them more realistic, and continues the evaluation with the most deployment-like candidate. Our second technique, DISH (Deployment-Imitating SWE-Agent Harness), wraps the target in an agent harness, reducing the gap between simulated and real deployment environments in coding settings. We test the techniques on multiple target models and find that they compose: applying both yields larger realism gains than either alone. Our results show that automated approaches can improve the realism of alignment evaluations, and that these improvements use additional compute more effectively than making the audits longer.

cs.AI↗

arXiv2025

Internal states before wait modulate reasoning patterns

Prior work has shown that a significant driver of performance in reasoning models is their ability to reason and self-correct. A distinctive marker in these reasoning traces is the token wait, which often signals reasoning behavior such as backtracking. Despite being such a complex behavior, little is understood of exactly why models do or do not decide to reason in this particular manner, which limits our understanding of what makes a reasoning model so effective. In this work, we address the question whether model's latents preceding wait tokens contain relevant information for modulating the subsequent reasoning process. We train crosscoders at multiple layers of DeepSeek-R1-Distill-Llama-8B and its base version, and introduce a latent attribution technique in the crosscoder setting. We locate a small set of features relevant for promoting/suppressing wait tokens' probabilities. Finally, through a targeted series of experiments analyzing max activating examples and causal interventions, we show that many of our identified features indeed are relevant for the reasoning process and give rise to different types of reasoning patterns such as restarting from the beginning, recalling prior knowledge, expressing uncertainty, and double-checking.

cs.AI↗

arXiv2025

NNsight and NDIF: Democratizing Access to Open-Weight Foundation Model Internals

We introduce NNsight and NDIF, technologies that work in tandem to enable scientific study of the representations and computations learned by very large neural networks. NNsight is an open-source system that extends PyTorch to introduce deferred remote execution. The National Deep Inference Fabric (NDIF) is a scalable inference service that executes NNsight requests, allowing users to share GPU resources and pretrained models. These technologies are enabled by the Intervention Graph, an architecture developed to decouple experimental design from model runtime. Together, this framework provides transparent and efficient access to the internals of deep neural networks such as very large language models (LLMs) without imposing the cost or complexity of hosting customized models individually. We conduct a quantitative survey of the machine learning literature that reveals a growing gap in the study of the internals of large-scale AI. We demonstrate the design and use of our framework to address this gap by enabling a range of research methods on huge models. Finally, we conduct benchmarks to compare performance with previous approaches. Code, documentation, and tutorials are available at https://nnsight.net/.

cs.LG↗

arXiv2023

Finding Neurons in a Haystack: Case Studies with Sparse Probing

Despite rapid adoption and deployment of large language models (LLMs), the internal computations of these models remain opaque and poorly understood. In this work, we seek to understand how high-level human-interpretable features are represented within the internal neuron activations of LLMs. We train $k$-sparse linear classifiers (probes) on these internal activations to predict the presence of features in the input; by varying the value of $k$ we study the sparsity of learned representations and how this varies with model scale. With $k=1$, we localize individual neurons which are highly relevant for a particular feature, and perform a number of case studies to illustrate general properties of LLMs. In particular, we show that early layers make use of sparse combinations of neurons to represent many features in superposition, that middle layers have seemingly dedicated neurons to represent higher-level contextual features, and that increasing scale causes representational sparsity to increase on average, but there are multiple types of scaling dynamics. In all, we probe for over 100 unique features comprising 10 different categories in 7 different models spanning 70 million to 6.9 billion parameters.

cs.LG↗