arXiv subjects

Brady Steele

Publications and source records attributed to Brady Steele.

5 recordsLinked to original sources

arXiv2026

Annotation Entropy Predicts Per-Example Learning Dynamics in LoRA Fine-Tuning

We find that LoRA fine-tuning exhibits un-learning on contested examples: items with high annotator disagreement show increasing loss during training, a qualitatively distinct pattern largely absent under full fine-tuning and consistent across all six models tested (four encoder, two decoder-only). This discovery emerges from correlating annotation entropy, computed from ChaosNLI's 100 labels per example, with per-example area under the loss curve (AULC) on SNLI and MNLI. The correlation is positive in all 25 conditions tested (Spearman $\rho = 0.06$-$0.43$), with decoder-only models showing stronger correlations than encoders at matched LoRA rank. The effect survives partial-correlation controls and replicates across seeds and datasets. A preliminary noise-injection experiment is consistent with these findings.

cs.LG↗

arXiv2026

Subspace Geometry Governs Catastrophic Forgetting in Low-Rank Adaptation

Low-Rank Adaptation (LoRA) has emerged as a parameter-efficient approach for adapting large pre-trained models, yet its behavior under continual learning remains poorly understood. We present a geometric theory characterizing catastrophic forgetting in LoRA through the lens of gradient subspace interactions. Our central finding is that forgetting is governed by a simple geometric law: $\mathcal{F} = \alpha(1 - \cos^2\theta_{\min}) + \beta$, where $\theta_{\min}$ is the minimum principal angle between task gradient subspaces. This formulation reveals an approximate rank-invariance property, at high subspace angles, forgetting becomes largely independent of the adapter rank (coefficient of variation $\approx 0.8\%$ in controlled synthetic settings; CV $\approx 10$-$19\%$ on real benchmarks, suggesting this is regime-dependent rather than absolute). We validate our theory on synthetic tasks ($r=0.994$ correlation), Split-CIFAR100 with ViT-LoRA, and sequential GLUE with RoBERTa-LoRA. Our analysis reconciles seemingly contradictory findings in the literature: we show that rank affects forgetting only when task subspaces are similar (low angle), while orthogonal methods like O-LoRA provide minimal benefit when natural orthogonality is already high. These insights provide principled guidance for continual learning with parameter-efficient fine-tuning.

cs.LG↗

arXiv2026

Why LoRA Resists Label Noise: A Theoretical Framework for Noise-Robust Parameter-Efficient Fine-Tuning

Parameter-efficient fine-tuning methods like Low-Rank Adaptation (LoRA) have become the dominant paradigm for adapting large pretrained models. We present a theoretical framework explaining an underexplored property: LoRA's inherent resistance to label noise. Our analysis reveals three key insights. First, we prove that rank-$r$ LoRA cannot memorize all possible label assignments once the sample size exceeds $O(r(d+k-r))$, limiting its capacity to fit arbitrary noise. Second, we derive an optimal rank balancing approximation bias and noise-induced variance, showing it decreases with noise rate. Third, we establish temporal separation: clean patterns are learned early while noise memorization occurs later. We propose RACT (Rank-Aware Curriculum Training), leveraging rank discrepancy for noise detection. Experiments validate our predictions, with RACT achieving 91.1% F1 for noise detection on AG News while maintaining 91.46% accuracy, competitive with baselines that lack noise detection capability.

cs.LG↗

arXiv2026

On the Limits of Learned Importance Scoring for KV Cache Compression

We investigate learned KV cache compression through Speculative Importance Prediction (SIP), a 1.7M parameter non-query-aware scorer that predicts token importance from KV representations alone. Despite architectural sophistication (multi-horizon lookahead, cross-attention), SIP does not outperform simple baselines, including random selection, across 5 seeds, 4 retention levels, and 3 tasks. Key findings: (1) position-based heuristics (keep first 4 + last N tokens) match or exceed learned approaches; (2) prefill attention provides equivalent signal to complex learned scorers; (3) marginal information in KV representations beyond position and prefill attention appears limited for importance prediction. We hypothesize that circular dependence between future queries and generation trajectories contributes to this difficulty.

cs.LG↗

arXiv2026

Scaling Trends for Multi-Hop Contextual Reasoning in Mid-Scale Language Models

We present a controlled study of multi-hop contextual reasoning in large language models, providing a clean demonstration of the task-method dissociation: rule-based pattern matching achieves 100% success on structured information retrieval but only 6.7% on tasks requiring cross-document reasoning, while LLM-based multi-agent systems show the inverse pattern, achieving up to 80% on reasoning tasks where rule-based methods fail. Using a synthetic evaluation framework with 120 trials across four models (LLaMA-3 8B, LLaMA-2 13B, Mixtral 8x7B, DeepSeek-V2 16B), we report three key findings: (1) Multi-agent amplification depends on base capability: statistically significant gains occur only for models with sufficient reasoning ability (p < 0.001 for LLaMA-3 8B, p = 0.014 for Mixtral), with improvements of up to 46.7 percentage points, while weaker models show no benefit, suggesting amplification rather than compensation; (2) Active parameters predict reasoning performance: Mixtral's performance aligns with its ~12B active parameters rather than 47B total, consistent with the hypothesis that inference-time compute drives reasoning capability in MoE architectures; (3) Architecture quality matters: LLaMA-3 8B outperforms LLaMA-2 13B despite fewer parameters, consistent with known training improvements. Our results provide controlled quantitative evidence for intuitions about multi-agent coordination and MoE scaling, while highlighting the dependence of multi-agent benefits on base model capability. We release our evaluation framework to support reproducible research on reasoning in mid-scale models.

cs.AI↗