Item titled POEM describes the TERM jailbreak defense
The supplied abstract is TERM, tested against seven jailbreaks on Llama, Qwen, and Gemma, including Qwen3 from 0.6B to 32B.
Stories per day · 30 days
Relatedattention1cwe1cweval1ensembles1gpt-51implicature1jailbreak1language-models1llm-defense1neural-thickets1over-refusal1pragmatics1reasoning-models1secure-code-generation1vision1
8 stories, newest first · one card per story
Stories tagged #inference-time
Neural Thickets gains traced to output diversity rather than weight experts
Perturb-Select-Vote on frozen models matches or beats full-weight perturbations on most reasoning tasks and helps vision under domain shift.
SecGent improves secure code generation without dropping required behavior
On CWEval, open-weight models gain 15.5% FS@1 over RAG, and Claude Sonnet 5.5 reaches 92.1%.
Extra thinking makes reasoning models cancel some and or implicatures
Models first entertain the enriched reading, then cancel it; early answers recover implicatures without losing literal accuracy.
Understanding and Mitigating Token-Pruning-Induced Vulnerabilities in VLMs
Token pruning can amplify jailbreak semantics in VLMs; Safety-Aware Pruning cuts attack success by up to 62%.
Beyond Memory: Harnessing Long-Horizon Agents with Explicit Belief States
PoS keeps explicit belief states so long-horizon LLM agents avoid Belief Trapping and beat memory baselines.
Before It Fades: Reinforcing Temporal Representations at Inference Time in VideoLLMs
Temporal Activation Injection restores fading mid-layer timing signals in VideoLLMs without training.
World in World: Explore the World with World Models
World in World is a training-free inference-time interface enabling camera-controlled rerendering, long-horizon revisiting, and motion transfer with frozen causal video world models.