When large language model confidence signals fail

Following yesterday's agent-evaluation coverage, this revision tests 265,000 samples and finds that agreement predicts correctness only weakly and depends on the test regime. Self-consistency can help allocate compute, but it isn't a standalone confidence score.

Do LLMs Know What They Know? (v3)

arXiv

Version 3 separates factual accuracy from how well confidence tracks knowledge. Its Signal Detection Theory analysis treats calibration and metacognitive sensitivity as different measurements.

Prompt Framing Distorts Count-Based Evaluation (v2)

arXiv

Count-F1 can rise when a prompt supplies the expected error count even if error localization barely improves. This revision argues for pairing count scores with span- or edit-aware metrics.

Revealing Hidden Model Behaviors with Task-Specific Self-Reports (v2)

arXiv

The Stabilized Adapter for self-Report makes a fine-tuned model describe hidden behavior in plain language. The reported study tests seven implanted behaviors and reduces hallucinated reports relative to its cited baseline.

Adversarial Pragmatics for AI Safety Evaluation (v3)

arXiv

This diagnostic framework measures task success and policy compliance separately from risk, refusal, attribution, and confidence. Ambiguous safety behavior no longer disappears inside one pass/fail label.

Linguistic Firewall (v3)

arXiv

ANTAP routes work through active capability tests instead of agent descriptions or static representations. This removes description manipulation from the routing path and measures what an agent can demonstrate.

ATOD: Annealed Turn-Aware On-Policy Distillation (v4)

arXiv

ATOD starts multi-turn agent training with dense teacher guidance, then increases reinforcement-learning weight as training progresses. It also concentrates the distillation signal on turns with more disagreement or uncertainty.

Functional Cache Grafting (v2)

arXiv

FCGraft retrieves validated code skeletons with their key-value caches and stitches them into an embodied-agent policy. It patches only task-specific regions to target generation latency and control-code robustness.

Ouroboros-Spatial (v2)

arXiv

Ouroboros-Spatial uses a frozen proposer to generate spatial questions and executable ground truth, while a solver's confidence steers the next training round. The loop aims to spend fewer examples on tasks that are already trivial or still uninformative.

VendorBench-100 (v2)

arXiv

VendorBench-100 compares commercial APIs with vision-language models and open-source deepfake detectors under one protocol. Version 2 corrects a scoring-normalization error and narrows the paper's calibration claim.

G2VD: Generalizable AI-Generated Video Detection (v2)

arXiv

G2VD uses counterfactual intervention and causal disentanglement to reduce reliance on generator-specific shortcuts. The target is video detection that holds up on generators outside the training set.

RWGBench (v4)

arXiv

RWGBench evaluates related-work generation as a citation decision rather than a text-similarity task. It checks which sources were selected, whether they fit the surrounding text, how the material is organized, and where each citation sits in the scholarly record.