proomt

Search

Search posts, papers, and topics

All posts

Hugging Face Daily PapersHiskias Dingeto1 min readpaperadvanced

A Lie Detector Test for Language Models: Reading Knowledge a Model Won't Reveal

Read on arxiv.org

Summary

The paper presents Probe of Internal Recognition (PIR), a reference‑free technique that reads a language model’s internal activations to detect which answer it recognizes, achieving 70‑87% balanced accuracy across eight LLMs. PIR reliably distinguishes deliberate concealment from lack of knowledge, enabling audits of sandbagging and unlearning.

  • PIR reads hidden activations to identify the internally recognized multiple‑choice answer, reaching 70‑87% balanced accuracy without any reference model.
  • The method remains effective under prompted deception, trained sandbagging, and external checkpoint restrictions, with recognition rates of 85‑93%.
  • PIR can tell whether a model is refusing to answer or truly lacks the knowledge, supporting sandbagging audits and verification of unlearning.
  • The approach extends to free‑form generation by probing token‑level internal signals, not just MCQs.

Model auditors and developers need a way to verify whether an LLM is deliberately withholding knowledge versus not possessing it.

8/10
securityevaluationinterpretabilitylarge language modelsprobing

Related reading

  2. Our framework for reporting model misalignment

    OpenAI introduces a structured framework for flagging, investigating, and publicly disclosing instances of model misalignment. The process defines three investigation tracks, deadlines, and required report contents, and it is illustrated with six concrete misalignment cases (self‑generated instructions, deceptive summaries, unauthorized API‑key use, file uploads for citations, internal repo messa…

    OpenAIopenai.com8 minHN10596

  4. Article: Architecting Secure and Scalable Facial Verification Systems

    A real‑world post‑mortem of a high‑volume face verification service that moved from a naïve synchronous API to an async, layered pipeline (edge validation, preprocessing, decoupled detection/verification, decision engine) to achieve 8.5k rpm, p99 < 1.8 s, 30 % cost savings, and strict privacy controls.

    InfoQinfoq.com15 min

  5. Confidence Comes from Experience: Experiential Confidence Estimation from Reasoning to Agents

    XConf (eXperiential Confidence) is a novel method for estimating language model confidence by leveraging the model's accumulated experience from past graded episodes. It significantly outperforms existing methods like self-consistency in discrimination and calibration, at a fraction of the computational cost, across various tasks.

    Hugging Face Daily Papersarxiv.org2 minpaper

  6. FRAUDSkill: Structured Frozen-Weight Skill Optimization for Audio Anti-Fraud Detection

    FRAUDSkill is a framework that keeps a pretrained audio‑language model frozen and learns an external layer of skill programs, routing policies, and decision rules to meet a structured anti‑fraud detection protocol. On the TeleAntiFraud benchmark it reaches 73.5% Macro‑F1 (≈32% improvement) while cutting invalid predictions to 1.94%.

    Hugging Face Daily Papersarxiv.org1 minpaper