Manifold of Failure: Behavioral Attraction Basins in Language Models

• URL: http://arxiv.org/abs/2602.22291v2
• Date: Fri, 27 Feb 2026 17:04:02 GMT
• Title: Manifold of Failure: Behavioral Attraction Basins in Language Models
• Authors: Sarthak Munshi, Manish Bhatt, Vineeth Sai Narajala, Idan Habler, Ammar Al-Kahfah, Ken Huang, Blake Gatto,
• Abstract summary: This paper introduces a framework for systematically mapping the Manifold of Failure in Large Language Models (LLMs)<n>We reframe the search for vulnerabilities as a quality diversity problem, using MAP-Elites to illuminate the continuous topology of these failure regions.<n>Across three LLMs, we show that MAP-Elites achieves up to 63% behavioral coverage, discovers up to 370 distinct vulnerability niches, and reveals dramatically different model-specific topological signatures.
• Score: 0.49388902330345724
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: While prior work has focused on projecting adversarial examples back onto the manifold of natural data to restore safety, we argue that a comprehensive understanding of AI safety requires characterizing the unsafe regions themselves. This paper introduces a framework for systematically mapping the Manifold of Failure in Large Language Models (LLMs). We reframe the search for vulnerabilities as a quality diversity problem, using MAP-Elites to illuminate the continuous topology of these failure regions, which we term behavioral attraction basins. Our quality metric, Alignment Deviation, guides the search towards areas where the model's behavior diverges most from its intended alignment. Across three LLMs: Llama-3-8B, GPT-OSS-20B, and GPT-5-Mini, we show that MAP-Elites achieves up to 63% behavioral coverage, discovers up to 370 distinct vulnerability niches, and reveals dramatically different model-specific topological signatures: Llama-3-8B exhibits a near-universal vulnerability plateau (mean Alignment Deviation 0.93), GPT-OSS-20B shows a fragmented landscape with spatially concentrated basins (mean 0.73), and GPT-5-Mini demonstrates strong robustness with a ceiling at 0.50. Our approach produces interpretable, global maps of each model's safety landscape that no existing attack method (GCG, PAIR, or TAP) can provide, shifting the paradigm from finding discrete failures to understanding their underlying structure.

Related papers

• When World Models Dream Wrong: Physical-Conditioned Adversarial Attacks against World Models [54.08784776767683]
  We present Physical-Conditioned World Model Attack (PhysCond-WMA), the first white-box world model attack that perturbs physical-condition channels.<n>PhysCond-WMA induces semantic, logic, or decision-level distortion while preserving perceptual fidelity.
  arXiv  Detail & Related papers  (2026-02-21T07:22:37Z)
• Are Aligned Large Language Models Still Misaligned? [13.062124372682106]
  Mis-Align Bench is a unified benchmark for analyzing misalignment across safety, value, and cultural dimensions.<n> SAVACU is an English-aligned dataset of 382,424 misaligned samples spanning 112 domains (or labels)
  arXiv  Detail & Related papers  (2026-02-11T19:30:43Z)
• AgentCPM-Explore: Realizing Long-Horizon Deep Exploration for Edge-Scale Agents [75.67445299298949]
  AgentCPM-Explore is a compact 4B agent model with high knowledge density and strong exploration capability.<n>We introduce a holistic training framework featuring parameter-space model fusion, reward signal denoising, and contextual information refinement.<n>AgentCPM-Explore achieves state-of-the-art (SOTA) performance among 4B-class models, matches or surpasses 8B-class SOTA models on four benchmarks, and even outperforms larger-scale models such as Claude-4.5-Sonnet or DeepSeek-v3.2 in five benchmarks.
  arXiv  Detail & Related papers  (2026-02-06T08:24:59Z)
• SSVP: Synergistic Semantic-Visual Prompting for Industrial Zero-Shot Anomaly Detection [55.54007781679915]
  We propose Synergistic Semantic-Visual Prompting (SSVP), that efficiently fuses diverse visual encodings to elevate model's fine-grained perception.<n>SSVP achieves state-of-the-art performance with 93.0% Image-AUROC and 92.2% Pixel-AUROC on MVTec-AD, significantly outperforming existing zero-shot approaches.
  arXiv  Detail & Related papers  (2026-01-14T04:42:19Z)
• What Matters For Safety Alignment? [38.86339753409445]
  This paper presents a comprehensive empirical study on the safety alignment capabilities of AI systems.<n>We systematically investigate and compare the influence of six critical intrinsic model characteristics and three external attack techniques.<n>We identify the LRMs GPT-OSS-20B, Qwen3-Next-80B-A3B-Thinking, and GPT-OSS-120B as the top-three safest models.
  arXiv  Detail & Related papers  (2026-01-07T12:31:52Z)
• Replicating TEMPEST at Scale: Multi-Turn Adversarial Attacks Against Trillion-Parameter Frontier Models [0.0]
  This study employed the TEMPEST multi-turn attack framework to evaluate ten frontier models from eight vendors across 1,000 harmful behaviors.<n>Six models achieved 96% to 100% attack success rate (ASR), while four showed meaningful resistance, with ASR ranging from 42% to 78%.
  arXiv  Detail & Related papers  (2025-12-08T00:30:40Z)
• DUAL-Bench: Measuring Over-Refusal and Robustness in Vision-Language Models [59.45605332033458]
  Safety mechanisms can backfire, causing over-refusal, where models decline benign requests out of excessive caution.<n>No existing benchmark has systematically addressed over-refusal in the visual modality.<n>This setting introduces unique challenges, such as dual-use cases where an instruction is harmless, but the accompanying image contains harmful content.
  arXiv  Detail & Related papers  (2025-10-12T23:21:34Z)
• When Punctuation Matters: A Large-Scale Comparison of Prompt Robustness Methods for LLMs [55.20230501807337]
  We present the first systematic evaluation of 5 methods for improving prompt robustness within a unified experimental framework.<n>We benchmark these techniques on 8 models from Llama, Qwen and Gemma families across 52 tasks from Natural Instructions dataset.
  arXiv  Detail & Related papers  (2025-08-15T10:32:50Z)
• Eliciting and Analyzing Emergent Misalignment in State-of-the-Art Large Language Models [0.0]
  We demonstrate that state-of-the-art language models remain vulnerable to carefully crafted conversational scenarios.<n>We discover 10 successful attack scenarios, revealing fundamental vulnerabilities in how current alignment methods handle narrative immersion, emotional pressure, and strategic framing.<n>To validate generalizability, we distilled our successful manual attacks into MISALIGNMENTBENCH, an automated evaluation framework.
  arXiv  Detail & Related papers  (2025-08-06T08:25:40Z)
• Backdoor Cleaning without External Guidance in MLLM Fine-tuning [76.82121084745785]
  Believe Your Eyes (BYE) is a data filtering framework that leverages attention entropy patterns as self-supervised signals to identify and filter backdoor samples.<n>It achieves near-zero attack success rates while maintaining clean-task performance.
  arXiv  Detail & Related papers  (2025-05-22T17:11:58Z)
• Exploring the Relationship between Samples and Masks for Robust Defect Localization [1.90365714903665]
  This paper proposes a one-stage framework that detects defective patterns directly without the modeling process. Explicit information that could indicate the position of defects is intentionally excluded to avoid learning any direct mapping. Results show that the proposed method is 2.9% higher than the SOTA methods in F1-Score, while substantially outperforming SOTA methods in generalizability.
  arXiv  Detail & Related papers  (2023-06-19T06:41:19Z)

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.