Safety Mid-Training: Internalizing LLM Safety as a Foundational Capability
24d
- Publicado
- Coletado
Safety alignment for large language models (LLMs) is primarily performed during post-training, while recent work has also begun incorporating safety data into pre-training. Yet safety can degrade when later training no longer includes safety data, and gains in safety often come at the cost of general capability. This tension suggests that safety and general reasoning should not be treated as separate training goals. We instead view safety as an LLM capability that can be learned through a reasoning structure shared with general tasks. Both require the model to understand intent or task goals, apply relevant knowledge, and reason about potential consequences. We introduce safety mid-training, a stage between pre-training and post-training that jointly develops safety and general reasoning through this shared structure. Given safety and utility prompts, the base model generates structured documents and reasoning traces that are then used for mid-training. Safety mid-training reduces average attack success rate (ASR) by 72% to 81% relative to Qwen3-1.7B, 4B, and 8B base models, while also improving general capability. The benefits persist through subsequent post-training. After general-only SFT, mid-trained models largely retain their safety, achieving lower ASR than the baselines without sacrificing general capability. They also retain safety under general-only GRPO without safety rewards, while subsequent RL reaches a better safety-utility frontier when initialized from the mid-trained models. Together, these results suggest that learning safety and general reasoning within a shared mid-training stage can reduce the safety-utility trade-off and make safety more robust to subsequent training.
Autores
Zhengyue Zhao, Liwei Jiang, Yejin Choi, Chaowei Xiao1
Categorias
O overfeed.news indexa e aponta. Publicamos um trecho curto — o artigo completo fica em arXiv AI Papers.
Mais de arXiv AI Papers
Entre para seguir esta fonte
Estimating Image Creativity
What makes an image creative? Although this may seem like a highly subjective decision, research in psychology suggests that creativity may be based on a set of fundamental creativity dimensions such as novelty, surprise, and arousal. A deeper understanding of image creativity can benefit both research and practical applications. Exploring how human viewers use creativity dimensions to evaluate image creativity may allow us to predict a target image's human-rated creativity automatically. We present a novel framework to estimate the perceived creativity of an image by: (1) predicting the strengths of an image's novelty, surprise, and arousal creativity dimensions, and (2) combining the creativity dimension strengths to estimate an image creativity score. Controlled human-subject experiments were conducted on Amazon Mechanical Turk to collect data to train supervised machine learning models to estimate an image's strengths on the three creativity dimensions. A final machine learning model takes the predicted dimension strengths as input to produce an estimated image creativity score. Results showed that our estimated creativity scores are statistically equivalent to human ratings of creativity for the images we tested. The novel contributions of our research are: (1) identification and validation of three psychological creativity dimensions used to estimate image creativity; (2) supervised machine learning models to automatically estimate novelty, surprise, and arousal strengths for an input image; (3) a supervised machine learning model to predict a creativity score from novelty, surprise, and arousal strengths for an input image, and (4) experimental methods to confirm the accuracy, advantages, and limitations of our approach.
AROMA: Adaptive Representation-Oriented Multimodal Actuation for Vision-Language Model Control
Mechanistic localization does not by itself yield usable control: causal sensitivity may occupy little activation energy, an accurate local direction may fail at finite scale, and a reachable numeral target may be wrong. We introduce AROMA (Adaptive Representation-Oriented Multimodal Actuation), a frozen-VLM control framework separating where to intervene, which geometry and gain to use, how to reach a target, which target to pursue, and when to act. In Llama-3.2-11B-Vision-Instruct, four dimensions of a causally validated cross-attention head explain 89.82% of cardinality sensitivity but only 9.52% of median activation energy; magnitude-normalized causal subspace actuation converts this mismatch into finite control, improving exact-count accuracy on untouched procedural data by +8.00 pp. Iterative Minimal-Effort Actuation (MEA) reaches 81.25% of oracle targets versus 15.18% for a frozen direction under matched step budgets. With target and gate frozen before evaluation, targeted MEA improves the previous frozen policy from 34.444% to 38.730% on a new natural Stage-2 cohort (N=630; paired p=4.63e-7) and yields 65.741% versus 65.153% over all 7,992 examples while editing only 63 inputs. Prospective routing, replication across heads and models, and comparisons with attribution-steering baselines further characterize each stage, separating localization, actuation, reachability, target quality, and selective deployment as distinct control problems.
Moving Beyond the DRW Model: Learning Quasar Variability with Generative Models
The optical continuum variability of quasars is often described by a damped random walk (DRW) model, with characteristic damping timescales τ_{DRW} of hundreds of days and amplitudes σ_{DRW} of {O}(0.1) magnitudes. This model has been widely used as a null hypothesis to assess the significance of periodic signals in quasar light curves in the search for supermassive black hole binaries, the significance of extreme quasar variability like tidal disruption events (TDEs) and changing-look (CL) quasars, or to model time lags in reverberation mapping. However, the DRW model is known to be incomplete, and an incorrect null hypothesis can lead to biased significance estimates. We investigate the impact of such null-hypothesis mis-specification on the detection of periodic signals in synthetic quasar light curves and introduce the generative model Nullasar, capable of learning quasar variability directly from time-series data. We demonstrate that Nullasar, trained with flow matching, can learn delicate statistical properties of stochastic light-curve variability without explicitly assuming any noise model. We further demonstrate that Nullasar can learn correlated distributions of noise parameters and generate new light-curve realizations from a range of deviations from the Gaussian DRW noise model, approximating parametric descriptions of stochastic processes. When used as a null hypothesis for detecting periodic signals in synthetic light curves, Nullasar performs comparably to fitting the correct DRW model to each light curve, and outperforms parametric fitting when the variability follows a more flexible AR(2) model. Finally, we identify practical limitations, including sensitivity to contamination of the training set by non-null signals, and show that projecting the conditioning vector of the model onto its leading principal components largely mitigates it.
Mask-Pattern Position Augmentation: Training-Free Long-Context Extrapolation for Diffusion LLMs
Diffusion large language models (dLLMs) generate by iteratively unmasking tokens with bidirectional attention, but rotary position embeddings (RoPE) still confine them to a fixed pretraining window. Existing training-free extrapolation, notably LongLLaDA, copies NTK-based RoPE scaling from autoregressive (AR) models and applies one static, length-only transform at every position and denoising step. That transplant ignores a signal unique to dLLMs: the evolving mask pattern of the partially decoded sequence. We propose Mask-Pattern Position Augmentation (MPPA). MPPA linearly interpolates indices into the pretrained window, then adds a per-position, per-step offset from cumulative mask density. The offset restores rotary phase resolution that interpolation collapses, and it shrinks as tokens are committed, so late steps see near-clean interpolated geometry. On LLaDA-8B-Instruct at twice its 4096-token window, MPPA reduces perplexity from 19.51 under standard decoding to 3.12 at 8192 tokens, below the in-window figure, with 100% passkey retrieval and unchanged needle-in-a-haystack accuracy. These preliminary 2results motivate mask-conditioned position engineering as a diffusion-native path to long-context generation.