REVIEW 25 cited by
LLMs Encode Harmfulness and Refusal Separately
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
LLMs Encode Harmfulness and Refusal Separately
read the original abstract
LLMs are trained to refuse harmful instructions, but do they truly understand harmfulness beyond just refusing? Prior work has shown that LLMs' refusal behaviors can be mediated by a one-dimensional subspace, i.e., a refusal direction. In this work, we identify a new dimension to analyze safety mechanisms in LLMs, i.e., harmfulness, which is encoded internally as a separate concept from refusal. There exists a harmfulness direction that is distinct from the refusal direction. As causal evidence, steering along the harmfulness direction can lead LLMs to interpret harmless instructions as harmful, but steering along the refusal direction tends to elicit refusal responses directly without reversing the model's judgment on harmfulness. Furthermore, using our identified harmfulness concept, we find that certain jailbreak methods work by reducing the refusal signals without reversing the model's internal belief of harmfulness. We also find that adversarially finetuning models to accept harmful instructions has minimal impact on the model's internal belief of harmfulness. These insights lead to a practical safety application: The model's latent harmfulness representation can serve as an intrinsic safeguard (Latent Guard) for detecting unsafe inputs and reducing over-refusals that is robust to finetuning attacks. For instance, our Latent Guard achieves performance comparable to or better than Llama Guard 3 8B, a dedicated finetuned safeguard model, across different jailbreak methods. Our findings suggest that LLMs' internal understanding of harmfulness is more robust than their refusal decision to diverse input instructions, offering a new perspective to study AI safety.
Forward citations
Cited by 25 Pith papers
-
OpenSafeIntent: Evaluating Intent-Calibrated Safe Completion Across Dual-Use Prompt Sets
OpenSafeIntent benchmark shows models fail to calibrate safety across intent shifts in matched dual-use prompts, indicating current evaluations are insufficient.
-
The Piggyback Hypothesis of Generalization: Explaining and Mitigating Emergent Misalignment
Shared chat-template tokens piggyback narrow finetuning behaviors onto out-of-domain queries; regularizing their KV states (TReFT) reduces emergent misalignment and other off-topic generalization.
-
Low-Resource Safety Failures Are Action Failures, Not Representation Failures
Low-resource safety failures are action failures because the harmfulness representation transfers but the decision calibration does not; this is fixed by recalibrating a high-resource gate with 1-4 target-language examples.
-
Model-Adaptive Tool Necessity Reveals the Knowing-Doing Gap in LLM Tool Use
Model-adaptive tool necessity shows 26-54% mismatch with actual tool calls across LLMs, driven by nearly orthogonal hidden-state signals for cognition versus action.
-
How Alignment Routes: Localizing, Scaling, and Controlling Policy Circuits in Language Models
Alignment policy in language models is implemented as an early-commitment routing circuit of detection gates and amplifier heads that can be localized, scaled, and directly controlled without removing the underlying c...
-
Minionese: Comprehensive Benchmark and Mechanistic Study of Multilingual LLM Safety
Multilingual jailbreaks succeed mainly by routing harm through misaligned residual subspaces that under-project onto an intact, effectively one-dimensional cross-lingual refusal direction.
-
How Alignment Routes: Localizing, Scaling, and Controlling Policy Circuits in Language Models
Alignment-trained LMs route refusal through an intermediate-layer attention gate that triggers amplifier heads; modulating the gate controls policy from hard refusal to compliance, and encodings that evade the gate by...
-
Agent Security Needs Redefinition through a Holistic Framework
Agent security should be redefined around four contextual authorization properties instead of the content of the action performed.
-
RoCo-ACE: Rollout-Conditioned Online Distillation for Retention-Aware Knowledge Injection
A rollout-conditioned contrastive distillation loss plus sparse anchored cross-entropy improves injected-knowledge accuracy in MLLMs while keeping retention close to the base model.
-
When Behavioral Safety Evaluation Fails: A Representation-Level Perspective
Behavioral safety metrics for LLMs are insufficient because models can maintain safe outputs while remaining vulnerable to latent-space interventions, as shown via dissociated models and the new Latent Vulnerability Score.
-
Defending Jailbreak Attacks on Large Language Models via Manifold Trajectory Kinetics
MTK detects jailbreaks by monitoring the evolution of prompt neighborhood structures on the data manifold through LLM layers, reporting 95% TPR at 5% FPR on benign and 2% on pseudo-malicious prompts plus 85% TPR under...
-
The Piggyback Hypothesis of Generalization: Explaining and Mitigating Emergent Misalignment
The Piggyback Hypothesis attributes emergent misalignment to chat-template tokens piggybacking finetuned behavior; Token-Regularized Finetuning (TReFT) mitigates it by regularizing prefix token representations.
-
Model-Adaptive Tool Necessity Reveals the Knowing-Doing Gap in LLM Tool Use
LLMs show a knowing-doing gap in tool use: they often recognize when tools are needed via internal states but fail to translate that into actual tool calls, with mismatches of 26-54% on arithmetic and factual tasks.
-
Probing Persona-Dependent Preferences in Language Models
Linear probes on residual-stream activations identify a shared preference vector in LLMs that tracks choices across prompts and causally steers decisions even for anti-correlated personas.
-
Probing Persona-Dependent Preferences in Language Models
Linear probes on residual-stream activations extract a preference vector that tracks and steers pairwise task choices across personas in Gemma-3-27B and Qwen-3.5-122B, including anti-correlated evil personas.
-
Before the Last Token: Diagnosing Final-Token Safety Probe Failures
Final-token probes miss distributed unsafe evidence in jailbreaks, but a PCA-HMM model on prefill trajectories recovers many misses without naive pooling's false positives.
-
Harmful Intent as a Geometrically Recoverable Feature of LLM Residual Streams
Harmful intent is linearly separable in LLM residual streams across 12 models and multiple architectures, reaching mean AUROC 0.982 while showing protocol-dependent directions and strong generalization to held-out har...
-
Harmful Intent as a Geometrically Recoverable Feature of LLM Residual Streams
Harmful intent is geometrically recoverable as a linear direction or angular deviation in LLM residual streams, with high AUROC across 12 models, stable under alignment variants including abliterated ones, and transfe...
-
LLM Safety From Within: Detecting Harmful Content with Internal Representations
SIREN identifies safety neurons via linear probing on internal LLM layers and combines them with adaptive weighting to detect harm, outperforming prior guard models with 250x fewer parameters.
-
Different Paths to Harmful Compliance: Behavioral Side Effects and Mechanistic Divergence Across LLM Jailbreaks
Different LLM jailbreak techniques achieve similar harmful compliance but lead to distinct behavioral side effects and mechanistic changes.
-
Blind Refusal: Language Models Refuse to Help Users Evade Unjust, Absurd, and Illegitimate Rules
Language models refuse 75.4% of requests to evade defeated rules and do so even after recognizing reasons that undermine the rule's legitimacy.
-
Can Aha Moments Be Fake? Towards Quantifying Decorative and True Thinking in Chain-of-Thought
Chain-of-thought steps in LLMs vary in causal influence; many are decorative, TTS identifies them, and a latent steering direction can switch whether a model 'thinks' through a step.
-
Can Aha Moments Be Fake? Towards Quantifying Decorative and True Thinking in Chain-of-Thought
LLMs interleave true causal reasoning steps with decorative ones in CoT, with only ~2.3% of steps having high causal impact on AIME for Qwen-2.5, and a steering direction can force internal use of specific steps.
-
Geometric Configurations of Perturbed Jailbreak Prompts
In six small open-weight LLMs, jailbreak prompts are linearly separable in last-token embeddings by surface form, but not by refusal/compliance behavior.
-
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.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.