A new benchmark shows current vision-language models can name foods but fail at estimating portion and nutrient values and often give unsafe dietary advice for chronic-disease patients.
How Alignment and Jailbreak Work: Explain LLM Safety through Intermediate Hidden States
5 Pith papers cite this work. Polarity classification is still indexing.
5
Pith papers citing it
abstract
Large language models (LLMs) rely on safety alignment to avoid responding to malicious user inputs. Unfortunately, jailbreak can circumvent safety guardrails, resulting in LLMs generating harmful content and raising concerns about LLM safety. Due to language models with intensive parameters often regarded as black boxes, the mechanisms of alignment and jailbreak are challenging to elucidate. In this paper, we employ weak classifiers to explain LLM safety through the intermediate hidden states. We first confirm that LLMs learn ethical concepts during pre-training rather than alignment and can identify malicious and normal inputs in the early layers. Alignment actually associates the early concepts with emotion guesses in the middle layers and then refines them to the specific reject tokens for safe generations. Jailbreak disturbs the transformation of early unethical classification into negative emotions. We conduct experiments on models from 7B to 70B across various model families to prove our conclusion. Overall, our paper indicates the intrinsical mechanism of LLM safety and how jailbreaks circumvent safety guardrails, offering a new perspective on LLM safety and reducing concerns. Our code is available at https://github.com/ydyjya/LLM-IHS-Explanation.
citation-role summary
background 1
citation-polarity summary
years
2026 5roles
background 1polarities
background 1representative citing papers
SafeSpec integrates a latent safety head into speculative LLM decoding with rollback and reflective multi-sampling, cutting attack success rates 15% on Qwen3-32B while retaining 2.06x speedup on normal workloads.
Causal mediation analysis shows harmful LLM outputs arise in late layers from MLP failures and gating neurons, with early layers handling harm context detection and signal propagation.
Benign fine-tuning collapses safety geometry in guard models like Granite Guardian, dropping refusal to 0%, but Fisher-Weighted Safety Subspace Regularization restores it to 75% while improving robustness.
Final-token safety probes miss jailbreaks whose unsafe evidence appears earlier in prefill; a clean-trained PCA-HMM trajectory model recovers many of those misses without catastrophic false positives.
citing papers explorer
-
OmniFood-Bench: Evaluating VLMs for Nutrient Reasoning and Personalized Health Advice
A new benchmark shows current vision-language models can name foods but fail at estimating portion and nutrient values and often give unsafe dietary advice for chronic-disease patients.
-
SafeSpec: Fast and Safe LLM via Dynamic Reflective Sampling
SafeSpec integrates a latent safety head into speculative LLM decoding with rollback and reflective multi-sampling, cutting attack success rates 15% on Qwen3-32B while retaining 2.06x speedup on normal workloads.
-
Why Do Large Language Models Generate Harmful Content?
Causal mediation analysis shows harmful LLM outputs arise in late layers from MLP failures and gating neurons, with early layers handling harm context detection and signal propagation.
-
When Safety Geometry Collapses: Fine-Tuning Vulnerabilities in Agentic Guard Models
Benign fine-tuning collapses safety geometry in guard models like Granite Guardian, dropping refusal to 0%, but Fisher-Weighted Safety Subspace Regularization restores it to 75% while improving robustness.
-
Before the Last Token: Diagnosing Final-Token Safety Probe Failures
Final-token safety probes miss jailbreaks whose unsafe evidence appears earlier in prefill; a clean-trained PCA-HMM trajectory model recovers many of those misses without catastrophic false positives.