arXiv subjects

Pratinav Seth

Publications and source records attributed to Pratinav Seth.

At least 19 recordsLinked to original sources

arXiv2026

What Do Compliance Detectors Read? An Audit of Activation Probes and Guard Models

Regulatory compliance monitoring in deployed language models is increasingly implemented as a legal and audit control, checking model outputs against written rules spanning data protection, healthcare, financial regulation, and platform policy. Such monitoring is meaningful only if a detector's verdict depends on the stated rule rather than on surface features of the scenario. We show this condition fails across the current class of compliance detectors, a failure we call rule blindness. Deleting, permuting, or substituting the governing rule leaves detection accuracy unchanged for every guard and activation probe we test, including a policy-conditioned guard that correctly cites the governing clause yet barely changes its verdict when that clause is swapped for its permissive counterpart. A purpose-built benchmark crossing two rules with two scenarios, so that neither alone predicts the label, confirms the failure under a design no prior benchmark rules out, and shows that step by step reasoning, not any fast detector we test, is what escapes it. Auditing at scale requires a retraining-free detector, so we introduce the Internal Compliance Score (ICS): a training-free activation readout calibrated from ten labelled pairs and scored by a single projection. We hold ICS to the same scrutiny as the guards it audits: a pre-registered criterion for beating trivial baselines is not met, and a bag-of-words model matches its pooled generalisation exactly. It remains useful because it is inexpensive, letting us audit four deployed guard models, an 8B zero-shot judge, and thirteen benchmarks, and it raises the mechanically verified pass rate when used to rank candidate responses, though an adaptive white-box attack removes this gain. We release the counterfactual protocol and crossed-rule benchmark so rule blindness can be tested in future probe and guard claims.

cs.AI↗

arXiv2026

CircuitKIT : Circuit Discovery, Evaluation, and Application Toolkit for Mechanistic Interpretability

Circuit analysis can support not only model explanation but also downstream interventions such as pruning, editing, steering, and selective fine-tuning. However, conducting such analyses currently requires stitching together separate implementations for discovery, evaluation, and intervention, as well as hand-authoring the contrastive prompts required by many discovery methods. This fragmentation makes methods difficult to compare and limits their application beyond canonical tasks. We introduce CircuitKIT, a source-available library that connects the circuit-analysis workflow through a typed, serializable representation. CircuitKIT provides a suite of discovery algorithms, declarative interfaces for mapping structured data into discovery tasks, complementary circuit diagnostics, and downstream application modules. Together, these components provide common infrastructure for conducting and comparing circuit analyses. The library, examples, notebooks, and documentation are released at https://github.com/Lexsi-Labs/CircuitKIT .

cs.LG↗

arXiv2026

Faithfulness to Refusal: A Causal Audit of Neuron Selectors

Attribution scores increasingly identify which neuron rows of a language model matter for applications such as pruning, interpretability, and editing for safety, yet whether they identify causally important rows is rarely tested directly. We address this with two paired audits built on one-shot neuron-row zeroing. We first audit selectors at the language-modeling level: attribution methods substantially outperform activation and magnitude-based baselines at identifying dispensable rows across five LLMs. We then adapt the same intervention into a behavior test by driving it with a contrastive harmful-versus-benign signal; the attributed rows are sufficient to install refusal on hate and crime while keeping benign over-refusal low and preserving language model fluency, and specific in that layer-matched random controls at the same depths fail. Highly rank-stable selectors can be among the least causally valid. Refusal moreover lives in a redundant subspace, where different attribution methods install it through largely disjoint row sets, so the recovered edit is one realization of a sufficient set rather than a unique mechanism. Together, these findings show that rank-stability proxies miss the kinds of selector failures a direct causal audit can surface.%

cs.CL↗

arXiv2026

CuratorKIT : Data Curation and Synthetic Data Generation for LLM Post-Training

Data curation is a critical part of post-training pipelines for large language models, yet existing tools often treat ingestion, deduplication, synthetic generation, and quality filtering as separate stages. This fragmentation makes it difficult to audit pipeline decisions or understand why individual samples are rejected. CuratorKIT is an open-source Python library that covers this full lifecycle in a single configurable pipeline. The framework is composed of six source format readers and automatic schema detection, a pre-generation data hygiene layer for credentials, PII, and toxic content, eight LLM-powered generation tasks, three complementary quality gates with provenance-exact hallucination verification, structured adaptive recovery, and five training-ready export formats compatible with TRL, Unsloth, and AlignTune. Every pipeline decision is recorded in an append-only per-sample provenance chain, and rejected samples carry structured failure reasons rather than being silently discarded. CuratorKIT supports 100+ LLM providers through LiteLLM, exposes both a Python API and a YAML-driven CLI, and is designed for practitioners who need reproducible, auditable data pipelines at scale .

cs.CL↗

arXiv2026

ALIGNBEAM : Inference-Time Alignment Transfer via Cross-Vocabulary Logit Mixing

Domain fine-tuning degrades the safety of large language models: fine-tuned specialists readily comply with harmful prompts framed in domain language. Existing inference-time defenses that mix logits from a safe anchor model require both models to share a vocabulary, which rules them out for the cross-family specialists where safety is most degraded. We present ALIGNBEAM, a training-free method that lifts this restriction by translating anchor logits into the target model's vocabulary token-by-token at each decoding step; a small LLM judge then selects the safest among K candidate continuations. No weights are changed, and the safety-utility trade-off can be tuned at deployment without retraining. Across both cross-vocabulary and same-vocabulary evaluation pairs, ALIGNBEAM substantially raises refusal on adversarial benchmarks while keeping task accuracy and inference overhead within practical bounds. The results show that safety alignment can be transferred between model families at inference time, without touching either model's weights.

cs.CL↗

arXiv2026

Provenance-Grounded Gating and Adaptive Recovery in Synthetic Post-Training Data Curation

Synthetic post-training pipelines commonly filter generated samples with reward models or holistic LLM judges, yet two practices remain rarely examined together: whether the filtering signal is grounded in the source evidence that induced each generation, and whether rejected samples can be systematically recovered rather than permanently discarded. We present a controlled study of both questions across gate configurations, recovery strategies, and generator scales, using adversarially injected corpora to provide ground-truth failure labels. We find that exact source provenance improves faithfulness gating for stronger judges, that hallucination and reward gates reject largely disjoint sample populations making both necessary, and that an adaptive recovery pipeline combining failure diagnosis with targeted regeneration achieves higher yield, recovery rate, and injection recall than naive resampling. Downstream fine-tuning quality is driven primarily by generator scale, with filtration and recovery conditions contributing meaningfully but secondarily.

cs.CL↗

arXiv2026

Data Presentation Over Architecture: Resampling Strategies for Credit Risk Prediction with Tabular Foundation Models

Credit default prediction is a tabular learning problem with severe class imbalance, heterogeneous features, and tight latency budgets. Tabular Foundation Models (TFMs) approach this problem through in-context learning, which makes their predictions sensitive to how the context window is built. We benchmark four classical models and five TFMs on the Home Credit and Lending Club datasets, varying the context-construction strategy (seven options) and the context size (1K to 50K). On both datasets, the choice of context strategy explains more variance in AUC-ROC than the choice of TFM family: balanced and hybrid sampling add 3 to 4 AUC points over uniform sampling, and the gap exceeds the spread between TFMs. With a balanced context of 5K to 10K examples, the strongest TFMs reach the AUC of classical baselines trained on the full data, while also recovering meaningful default-class recall that default-threshold GBDTs do not. We frame this as evidence that context construction, rather than architecture choice, is the primary deployment lever for TFMs in imbalanced credit-risk settings.

cs.LG↗

arXiv2026

Pocket Foundation Models: Distilling TFMs into CPU-Ready Gradient-Boosted Trees

A fraud scorer needs to answer in under 2 ms. The best tabular foundation models (TFMs) take 151-1,275 ms on GPU. We close this gap by distilling the TFM offline into an XGBoost or CatBoost student that runs natively on CPU. The central obstacle is specific to in-context learning (ICL) teachers: they leak labels when scoring their own training set, so the soft targets collapse to near-one-hot vectors with no inter-class structure left to distill. Stratified out-of-fold (OOF) teacher labeling prevents this. Across 153 classification datasets drawn from TALENT, OpenML-CC18, TabZilla, and TabArena, distilling TabICLv2 into XGBoost gives 0.882 macro-mean AUC (96.5% of teacher AUC) at 1.9 ms on CPU, a 38x to 860x speedup across teacher-student pairs with a statistically significant edge over a tuned CatBoost baseline (Wilcoxon p = 0.0008; 51% win rate). Four further findings: teacher rank transfers exactly to student rank; gains concentrate on low-dimensional data (< 21 features: +0.011 over CatBoost vs. >21 features: +0.001); multi-teacher averaging helps MLP students (+0.006, p = 0.003) but adds less than 0.001 for tree students; and on high-dimensional tasks where the teacher itself trails CatBoost, distillation makes things worse rather than better. The full pipeline is open-sourced as part of the TabTune library.

cs.LG↗

arXiv2026

Ensembling Tabular Foundation Models - A Diversity Ceiling And A Calibration Trap

Tabular foundation models (TFMs) now match or beat tuned gradient-boosted trees on a growing fraction of tabular tasks, but no single TFM wins on every dataset. Ensembling is the go to fix here, and it works less well than expected. Six modern TFMs form a near-redundant pool: their mean pairwise Q-statistic is $0.961$, close enough to $1$ that any convex combination is bounded above. We benchmark six ensemble strategies over six TFMs on 153 OpenML classification tasks. The best ensemble, two-level cascade stacking, buys $+0.18\%$ accuracy over the strongest single TFM at $253\times$ the compute. A Friedman and Nemenyi analysis places three ensembles and the best base TFM in a single equivalence group; three other ensembles are significantly \emph{worse} than the best base. Stacking with a logistic-regression meta-learner is the most striking case: competitive accuracy and ROC-AUC, the worst log-loss rank among the ensembles. The meta-learner improves accuracy by sharpening class boundaries, which destroys calibration. We recommend greedy selection as the practical default.

cs.LG↗

arXiv2026

Distilling Tabular Foundation Models for Structured Health Data

Tabular foundation models (TFMs) achieve strong performance on health datasets, but their inference cost and infrastructure requirements limit practical use. We study whether their predictive behavior can be transferred to lightweight tabular models through knowledge distillation. Since in-context TFMs condition on the training set at inference time, naive distillation can introduce context leakage; we address this with stratified out-of-fold teacher labeling. Across $19$ healthcare datasets, $6$ TFM teachers, $4$ student families, and several multi-teacher ensembles, we find that distilled students retain at least $90\%$ of teacher AUC, outperforming teachers in some cases, while running at least $26\times$ faster on CPU and preserving calibration and fairness critical for health applications. Moreover, multi-teacher averaging does not consistently improve over the best single teacher. Leakage-aware distillation is thus a viable route for bringing TFM-quality predictions into inference-constrained health settings.

cs.LG↗

arXiv2026

Forgetting That Sticks: Quantization-Permanent Unlearning via Circuit Attribution

Standard unlearning evaluations measure behavioral suppression in full precision, immediately after training, despite every deployed language model being quantized first. Recent work has shown that 4-bit post-training quantization can reverse machine unlearning; we show this is not a tuning artefact but a systematic dual failure: gradient-based methods that achieve meaningful forgetting lose it under compression, while methods that survive quantization barely change the model. Both failures trace to the same root cause: across all baselines, per-parameter updates lie 47-828x below the NF4 quantization bin width; updates diffused across billions of parameters cannot clear quantization bin boundaries, a consequence we formalize as a sparsity-permanence tradeoff. We present MANSU (Mechanistic-Aligned Null-Space Unlearning), which resolves both modes by combining causal circuit attribution to isolate the minimal forget-set subgraph, circuit-restricted null-space projection with a diagonal-Fisher retain bound, and a per-parameter magnitude floor guaranteeing quantization survival by construction. We additionally introduce Circuit Attribution Divergence (CAD), a mechanistic verification metric distinguishing structural erasure from behavioral suppression, a distinction existing metrics cannot make. Across multiple model families and hazard benchmarks, MANSU is the first method to jointly satisfy all four properties with margin on each (meaningful forgetting, retain preservation, non-positive PTQ gap, and structural erasure), while gradient-based baselines recover up to +0.05 accuracy under compression.

cs.LG↗

arXiv2026

Position: Behavioural Assurance Cannot Verify the Safety Claims Governance Now Demands

This position paper argues that behavioural assurance, even when carefully designed, is being asked to carry safety claims it cannot verify. AI governance frameworks enacted between 2019 and early 2026 require reviewable evidence of properties such as the absence of hidden objectives, resistance to loss-of-control precursors, and bounded catastrophic capability; current assurance methodologies (primarily behavioural evaluations and red-teaming) are epistemically limited to observable model outputs and cannot verify the latent representations or long-horizon agentic behaviours these frameworks presume to regulate. We formalize this structural mismatch as the audit gap, the divergence between required and achievable verification access, and introduce the concept of fragile assurance to describe cases where the evidential structure does not support the asserted safety claim. Through an analysis of a 21-instrument inventory, we identify an incentive gradient where geopolitical and industrial pressures systematically reward surface-level behavioral proxies over deep structural verification. Finally, we propose a technical pivot: bounding the weight of behavioral evidence in legal text and extending voluntary pre-deployment access with mechanistic-evidence classes, specifically linear probes, activation patching, and before/after-training comparisons.

cs.LG↗

arXiv2026

AlignTune: Modular Toolkit for Post-Training Alignment of Large Language Models

Post-training alignment is central to deploying large language models (LLMs), yet practical workflows remain split across backend-specific tools and ad-hoc glue code, making experiments hard to reproduce. We identify backend interference, reward fragmentation, and irreproducible pipelines as key obstacles in alignment research. We introduce AlignTune, a modular toolkit exposing a unified interface for supervised fine-tuning (SFT) and RLHF-style optimization with interchangeable TRL and Unsloth backends. AlignTune standardizes configuration, provides an extensible reward layer (rule-based and learned), and integrates evaluation over standard benchmarks and custom tasks. By isolating backend-specific logic behind a single factory boundary, AlignTune enables controlled comparisons and reproducible alignment experiments.

cs.CL↗

arXiv2026

$C$-$\Delta\Theta$: Circuit-Restricted Weight Arithmetic for Selective Refusal

Modern deployments require LLMs to enforce safety policies at scale, yet many controls rely on inference-time interventions that add recurring compute cost and serving complexity. Activation steering is widely used, but it requires runtime hooks and scales cost with the number of generations; conditional variants improve selectivity by gating when steering is applied but still retain an inference-time control path. We ask whether selective refusal can be moved entirely offline: can a mechanistic understanding of category-specific refusal be distilled into a circuit-restricted weight update that deploys as a standard checkpoint? We propose C-{\Delta}{\theta} Circuit Restricted Weight Arithmetic}, which (i) localizes refusal-causal computation as a sparse circuit using EAP-IG and (ii) computes a constrained weight update {\Delta}{\theta}C supported only on that circuit (typically <5% of parameters). Applying {\Delta}{\theta}C yields a drop-in edited checkpoint with no inference-time hooks, shifting cost from per request intervention to a one-time offline update. We evaluate category-targeted selectivity and capability retention on refusal and utility benchmarks.

cs.CL↗

arXiv2026

Exploring Fine-Tuning for Tabular Foundation Models

Tabular Foundation Models (TFMs) have recently shown strong in-context learning capabilities on structured data, achieving zero-shot performance comparable to traditional machine learning methods. We find that zero-shot TFMs already achieve strong performance, while the benefits of fine-tuning are highly model and data-dependent. Meta-learning and PEFT provide moderate gains under specific conditions, whereas full supervised fine-tuning (SFT) often reduces accuracy or calibration quality. This work presents the first comprehensive study of fine-tuning in TFMs across benchmarks including TALENT, OpenML-CC18, and TabZilla. We compare Zero-Shot, Meta-Learning, Supervised (SFT), and parameter-efficient (PEFT) approaches, analyzing how dataset factors such as imbalance, size, and dimensionality affect outcomes. Our findings cover performance, calibration, and fairness, offering practical guidelines on when fine-tuning is most beneficial and its limitations.

cs.LG↗

arXiv2025

Orion-Bix: Bi-Axial Attention for Tabular In-Context Learning

Tabular data drive most real-world machine learning applications, yet building general-purpose models for them remains difficult. Mixed numeric and categorical fields, weak feature structure, and limited labeled data make scaling and generalization challenging. To this end, we introduce Orion-Bix, a tabular foundation model that combines biaxial attention with meta-learned in-context reasoning for few-shot tabular learning. Its encoder alternates standard, grouped, hierarchical, and relational attention, fusing their outputs through multi-CLS summarization to capture both local and global dependencies efficiently. A label-aware ICL head adapts on the fly and scales to large label spaces via hierarchical decision routing. Meta-trained on synthetically generated, structurally diverse tables with causal priors, Orion-Bix learns transferable inductive biases across heterogeneous data. Delivered as a scikit-learn compatible foundation model, it outperforms gradient-boosting baselines and remains competitive with state-of-the-art tabular foundation models on public benchmarks, showing that biaxial attention with episodic meta-training enables robust, few-shot-ready tabular learning. The model is publicly available at https://github.com/Lexsi-Labs/Orion-BiX .

cs.LG↗

arXiv2025

TabTune: A Unified Library for Inference and Fine-Tuning Tabular Foundation Models

Tabular foundation models represent a growing paradigm in structured data learning, extending the benefits of large-scale pretraining to tabular domains. However, their adoption remains limited due to heterogeneous preprocessing pipelines, fragmented APIs, inconsistent fine-tuning procedures, and the absence of standardized evaluation for deployment-oriented metrics such as calibration and fairness. We present TabTune, a unified library that standardizes the complete workflow for tabular foundation models through a single interface. TabTune provides consistent access to seven state-of-the-art models supporting multiple adaptation strategies, including zero-shot inference, meta-learning, supervised fine-tuning (SFT), and parameter-efficient fine-tuning (PEFT). The framework automates model-aware preprocessing, manages architectural heterogeneity internally, and integrates evaluation modules for performance, calibration, and fairness. Designed for extensibility and reproducibility, TabTune enables consistent benchmarking of adaptation strategies of tabular foundation models.

cs.LG↗

arXiv2025

Orion-MSP: Multi-Scale Sparse Attention for Tabular In-Context Learning

Tabular data remain the predominant format for real-world applications. Yet, developing effective neural models for tabular data remains challenging due to heterogeneous feature types and complex interactions occurring at multiple scales. Recent advances in tabular in-context learning (ICL), such as TabPFN and TabICL, have achieved state-of-the-art performance comparable to gradient-boosted trees (GBTs) without task-specific fine-tuning. However, current architectures exhibit key limitations: (1) single-scale feature processing that overlooks hierarchical dependencies, (2) dense attention with quadratic scaling in table width, and (3) strictly sequential component processing that prevents iterative representation refinement and cross-component communication. To address these challenges, we introduce Orion-MSP, a tabular ICL architecture featuring three key innovations: (1) multi-scale processing to capture hierarchical feature interactions; (2) block-sparse attention combining windowed, global, and random patterns for scalable efficiency and long-range connectivity; and (3) a Perceiver-style memory enabling safe bidirectional information flow across components. Across diverse benchmarks, Orion-MSP matches or surpasses state-of-the-art performance while scaling effectively to high-dimensional tables, establishing a new standard for efficient tabular in-context learning. The model is publicly available at https://github.com/Lexsi-Labs/Orion-MSP .

cs.AI↗