Residual Stream Analysis of Overfitting And Structural Disruptions

• URL: http://arxiv.org/abs/2603.13318v1
• Date: Wed, 04 Mar 2026 11:13:58 GMT
• Title: Residual Stream Analysis of Overfitting And Structural Disruptions
• Abstract summary: We show that safety data exhibits substantially lower token entropy and 2-gram diversity compared to general instruction data.<n>To uncover the root cause, we introduce FlowLens, a stable PCA-based tool for residual-stream geometry analysis.<n>We propose Variance Concentration Loss, an auxiliary regularizer that penalizes excessive variance concentration in mid-layer residuals.
• Score: 26.444437808784482
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Ensuring that large language models (LLMs) remain both helpful and harmless poses a significant challenge: fine-tuning on repetitive safety datasets, where unsafe prompts are paired with standard refusal templates, often leads to false refusals, in which benign queries are declined. We first quantify this effect, showing that safety data exhibits substantially lower token entropy and 2-gram diversity (0.048) compared to general instruction data. To uncover the root cause, we introduce FlowLens, a stable PCA-based tool for residual-stream geometry analysis, and reveal that higher proportions of safety examples concentrate variance along a few components, reducing representational smoothness and driving false refusals (false refusal rate rises from 63 percent to 84 percent as safety data increases from 0 percent to 40 percent). Guided by these insights, we propose Variance Concentration Loss (VCL), an auxiliary regularizer that penalizes excessive variance concentration in mid-layer residuals. Empirical results demonstrate that VCL reduces false refusals by over 35 percentage points while maintaining or improving performance on general benchmarks such as MMLU and GSM8K.

Related papers

• From Detection to Refusal: Safer LLMs via Circuit-Guided Weight Scaling [23.223636694641502]
  Large Language Models (LLMs) are vulnerable to generating unsafe content under adversarial prompting.<n>We study safety from a mechanistic interpretability perspective and characterize a multi-stage *safety circuit*<n>We show that suppressing upstream Harmful Detection Heads disrupts downstream refusal behavior and that safety neurons mediate this interaction.
  arXiv  Detail & Related papers  (2026-08-30T08:01:42Z)
• MASH-Bench: Diagnosing Cross-Source Failure in Mass-Shooting Risk Classification [2.1067460775364286]
  Public mass-shooting databases differ substantially in coverage, feature availability, and reporting practices.<n>We introduce MASH-Bench, a harmonized benchmark of 6,968 incidents from four U.S. databases.<n>We evaluate cross-source risk classification using leave-one-dataset-out evaluation.
  arXiv  Detail & Related papers  (2026-08-23T15:28:10Z)
• Alignment Collapse Under KV Cache Quantization: Diagnosis and Mitigation [6.129872931808218]
  Key-value ( KV) cache quantization is widely used to reduce Large Language Model (LLM) inference memory.<n>In this study, we explore alignment preservation under KV cache quantization.<n>We find that low-bit quantization can silently destroy safety alignment.
  arXiv  Detail & Related papers  (2026-06-01T02:02:20Z)
• AgentDrift: Unsafe Recommendation Drift Under Tool Corruption Hidden by Ranking Metrics in LLM Agents [2.995458991057093]
  We introduce a paired-trajectory protocol that replays real financial dialogues under clean and contaminated tool-output conditions.<n>We observe the evaluation-blindness pattern: recommendation quality is largely preserved under contamination.<n>Even narrative-only corruption (biased headlines, no numerical manipulation) induces significant drift while completely evading consistency monitors.
  arXiv  Detail & Related papers  (2026-03-13T01:54:00Z)
• OOD-MMSafe: Advancing MLLM Safety from Harmful Intent to Hidden Consequences [64.01706941950489]
  Current safety paradigms primarily target malicious intent or situational violations.<n>We propose shifting the safety frontier toward consequence-driven safety, a paradigm essential for the robust deployment of autonomous and embodied agents.<n>We develop the Consequence-Aware Safety Policy Optimization (CASPO) framework, which integrates the model's intrinsic reasoning as a dynamic reference for token-level self-distillation rewards.
  arXiv  Detail & Related papers  (2026-03-10T14:16:43Z)
• SCAFFOLD-CEGIS: Preventing Latent Security Degradation in LLM-Driven Iterative Code Refinement [12.69450437027072]
  This paper reveals the iterative refinement paradox: specification drift during multi-objective optimization causes security to degrade gradually over successive iterations.<n>We propose the SCAFFOLD-CEGIS framework that transforms security constraints from implicit prompts into explicit verifiable constraints.
  arXiv  Detail & Related papers  (2026-03-09T15:54:18Z)
• Sharp Convergence Rates for Masked Diffusion Models [53.117058231393834]
  We develop a total-variation based analysis for the Euler method that overcomes limitations.<n>Our results relax assumptions on score estimation, improve parameter dependencies, and establish convergence guarantees.<n>Overall, our analysis introduces a direct TV-based error decomposition along the CTMC trajectory and a decoupling-based path-wise analysis for FHS.
  arXiv  Detail & Related papers  (2026-02-26T00:47:51Z)
• Selective Risk Certification for LLM Outputs via Information-Lift Statistics: PAC-Bayes, Robustness, and Skeleton Design [6.908972852063454]
  We develop information-lift certificates that compare model probabilities to a skeleton baseline, accumulating evidence into PAC-Bayes valid under heavy-tailed bounds.<n>Across eight datasets, our method achieves 77.2% coverage at 2% risk, outperforming recent 2023-2024 baselines by 8.6-15.1 percentage points.
  arXiv  Detail & Related papers  (2025-09-16T00:05:54Z)
• Rethinking Safety in LLM Fine-tuning: An Optimization Perspective [56.31306558218838]
  We show that poor optimization choices, rather than inherent trade-offs, often cause safety problems, measured as harmful responses to adversarial prompts.<n>We propose a simple exponential moving average (EMA) momentum technique in parameter space that preserves safety performance.<n>Our experiments on the Llama families across multiple datasets demonstrate that safety problems can largely be avoided without specialized interventions.
  arXiv  Detail & Related papers  (2025-08-17T23:46:36Z)
• Benchmarking Reasoning Robustness in Large Language Models [76.79744000300363]
  We find significant performance degradation on novel or incomplete data.<n>These findings highlight the reliance on recall over rigorous logical inference.<n>This paper introduces a novel benchmark, termed as Math-RoB, that exploits hallucinations triggered by missing information to expose reasoning gaps.
  arXiv  Detail & Related papers  (2025-03-06T15:36:06Z)
• Mind the Confidence Gap: Overconfidence, Calibration, and Distractor Effects in Large Language Models [0.6091702876917281]
  Large Language Models (LLMs) show remarkable proficiency in natural language tasks.<n>Overconfidence-misalignment between predicted confidence and true correctness poses significant risks in critical decision-making applications.<n>We present a comprehensive analysis on calibration in LLMs across nine LLMs and three factual Question-Answering datasets.
  arXiv  Detail & Related papers  (2025-02-16T07:46:09Z)
• Exploring Response Uncertainty in MLLMs: An Empirical Evaluation under Misleading Scenarios [49.53589774730807]
  Multimodal large language models (MLLMs) have recently achieved state-of-the-art performance on tasks ranging from visual question answering to video understanding.<n>We reveal a response uncertainty phenomenon: twelve state-of-the-art open-source MLLMs overturn a previously correct answer in 65% of cases after receiving a single deceptive cue.
  arXiv  Detail & Related papers  (2024-11-05T01:11:28Z)
• Geometry-Aware Instrumental Variable Regression [56.16884466478886]
  We propose a transport-based IV estimator that takes into account the geometry of the data manifold through data-derivative information. We provide a simple plug-and-play implementation of our method that performs on par with related estimators in standard settings.
  arXiv  Detail & Related papers  (2024-05-19T17:49:33Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.