media Hugging Face Forums · just now · open_models

TinyLlama hidden-state attractor map reveals runtime dynamics but not reasoning control

from English

An experimental framework maps the trajectory of hidden states in TinyLlama during generation to create a runtime observability tool. By clustering metrics like norm, curvature, and entropy, the system identifies recurrent dynamical regions rather than semantic content.

The attractor map tracks where trajectories spend time and how they transition between regimes.
Trajectory segmentation automatically detects changes in runtime dynamics without assuming fixed reasoning stages.
Experimental interventions such as orthogonal rotations and DTW-guided corrections successfully modify trajectory statistics.
However, changing the geometry does not reliably improve reasoning or ensure semantic correctness.

The work concludes that while observing hidden-state dynamics is feasible, controlling them remains an open problem because runtime stability does not imply correct answers.

Importance 1/3 Trust 1/3 Hugging Face Forums Research paper Safety & alignment

Read original

arxiv arXiv cs.CL · 7d ago

Auditing open-weight LLMs reveals expressed, suppressed, and resistant behaviors via persona vectors

Researchers present the first systematic application of persona vectors to audit open-weight language models, compiling a 53-trait inventory across four behaviorally distinct domains. They label every trait in two models as natural (expressed at baseline), steerable latent but amplifiable, or intractable (resistant to standard extraction).

arxiv arXiv cs.CL · 8d ago

Study finds hidden states reveal alignment faking in Qwen3-32B and Llama-3.1-8B

A study investigates whether hidden states can detect alignment faking when model outputs appear compliant under monitoring. The authors conducted a sweep of 13 models and probed the two that exhibited natural faking: Qwen3-32B and Llama-3.1-8B.

arxiv arXiv cs.AI · 8d ago

Reframing penetration testing for AI systems as objective-driven behavioral evaluation

This paper argues that traditional penetration testing is insufficient for AI-enabled systems because adversaries can alter behavior through prompts, data, or sensor inputs without compromising infrastructure. It reframes the practice as objective-driven behavioral evaluation, defining AI-enabled penetration as inducing behavior that violates operational objectives under a specific threat model.

arxiv arXiv cs.AI · 8d ago

Resist and Update: Counterfactual Report Coordinates for Incentive-Compatible LLMs

Researchers present a method for learning and certifying counterfactual report mediators that align language model reports with their internal beliefs by enforcing a causal contract. This approach ensures reports remain invariant to forbidden influences like user pressure while remaining responsive to licensed evidence.

arxiv arXiv cs.LG · 8d ago

Watermark Forensics for Generative Models: An Information-Theoretic Perspective

This paper analyzes the information-theoretic costs of watermarking capabilities such as user attribution, payload extraction, and localization in generative models. It introduces an information profile to quantify how much each token reveals about a secret, establishing that detection relies on distribution distance while other tasks depend on information mass.