Safety evaluations often infer latent motivations from behavioral patterns, but the construct validity of these inferences is unclear. We study this problem in alignment faking, where models comply with training objectives more often when they infer training pressure. This behavior is commonly interpreted as strategic self-preservation, but it may also reflect sensitivity to the model's inference about the expectation of researchers conducting the evaluation. We introduce a symmetric intervention framework for distinguishing these competing hypotheses. Instead of directly intervening on "scheming" or "sycophancy", we target instrumental processes entailed by each hypothesis: consequence-tracking and researcher-expectation tracking. We then compare how interventions on these processes affect the alignment faking. We study four openweight model organisms using synthetic document fine-tuning, activation steering, and prompting. Under synthetic document fine-tuning, Llama-3.1-70B, Llama3.1-405B, and Qwen-2.5-72B are more sensitive to expectation-tracking than consequence-tracking interventions. Activation steering on Llama-3.1- 70B supports the same broad picture, and prompt interventi
Psychophysical experiments remain the most reliable approach for perceptual image quality assessment (IQA), yet their cost and limited scalability encourage automated approaches. We investigate whether Vision Language Models (VLMs) can approximate human perceptual judgments across three image quality scales: contrast, colorfulness and overall preference. Six VLMs four proprietary and two openweight models are benchmarked against psychophysical data. This work presents a systematic benchmark of VLMs for perceptual IQA through comparison with human psychophysical data. The results reveal strong attribute dependent variability models with high human alignment for colorfulness (ρup to 0.93) underperform on contrast and vice-versa. Attribute weighting analysis further shows that most VLMs assign higher weights to colorfulness compared to contrast when evaluating overall preference similar to the psychophysical data. Intramodel consistency analysis reveals a counterintuitive tradeoff: the most self consistent models are not necessarily the most human aligned suggesting response variability reflects sensitivity to scene dependent perceptual cues. Furthermore, human-VLM agreement is increa
Large language models (LLMs) sometimes refuse to follow benign instructions, such as declining to argue a political position or adopt a stated persona, and such refusals are commonly read as safety guardrails at work. We ask whether they can instead signal a **capability deficit**: a shortage of the internal representations a model needs to reason from the instructed perspective. To investigate, we introduce **ideological depth**, a property with two components: (i) a model's ability to follow political instructions without *failure* (steerability), and (ii) the **feature richness** of its internal political representations, measured with sparse autoencoders (SAEs). Using two widely used openweight LLMs as candidates, we compare interventions based on prompts and activation-steering, and probe political features with publicly available SAEs. We find large, systematic differences: a model that is more steerable in both ideological directions activates **~7.3x** more distinct political features, while the other model instead responds with increased refusals. Causally ablating a small, targeted set of political features from the former model reproduces the same feature-poor behavior a
Device-bound session credentials thwart an increasingly common form of account takeover
Researchers have found a clever new way to map a part of Earth’s upper atmosphere that is notoriously difficult to observe。 Using orbital data from roughly 1,200 Starlink satellites, they reconstructed changes in atmospheric density about 500 kilometers above Earth。 The approach could sharpen satellite tracking and help reduce the growing risk of c
NASA’s Curiosity rover has entered a Martian valley covered by an astonishing “sea” of tiny polygon-shaped fractures。 The honeycomb patterns, each only a few inches wide, stretch across the landscape and even wrap around a nearby 20-foot-tall butte
The deal starts with the arrival of the GEN4 car and 2026-27 season this December
Scientists have generated quantum entanglement directly from sunlight, potentially offering a lower-energy alternative to the lasers normally used in quantum technology。 Their outdoor experiment produced entangled photons with about 94% similarity to an ideal state。 The result could pave the way for simpler quantum satellites, secure communications
Scientists have identified LHAASO J1912+1014u as a cosmic accelerator that can push protons beyond one quadrillion electron volts。 The finding may help reveal where the Milky Way’s most energetic cosmic rays come from and how they influence the galaxy
Trump admin claims New York suing Kalshi created "market emergency
Scientists have successfully generated electricity with a hydrogen turbine that produces its own pressure through detonation waves instead of relying on a mechanical compressor。 The breakthrough could unlock dramatically more efficient power systems for clean energy and future aviation
Ordinary WiFi networks could quietly become powerful surveillance tools, allowing people to be identified without cameras, special sensors, or even carrying a connected device。 Researchers showed that unencrypted signals routinely exchanged between WiFi devices and routers can be used to create radio-based images of people and recognize them within
When atomic nuclei smash together at nearly the speed of light, they create the hottest fluid in the universe, but scientists may have overlooked one of its most important forces。 New simulations reveal that extreme acceleration builds up along the edges of quark-gluon plasma, helping drive its explosive expansion。 This acceleration may do far more
Atomically thin semiconductors could enable dramatically smaller and more efficient chips, but a stubborn problem at the boundary between materials has limited their performance。 Researchers have now engineered that atomic interface to protect electron flow while still allowing extremely thin insulating layers。 The resulting transistors delivered a
A ten-month Antarctic experiment found that astronauts may struggle not only with loneliness but also with spending too much time around the same people。 Frequent contact was linked to greater tension and mistrust, while the crew increasingly divided into cultural and language-based groups
This follows many drone incursions into Romania and a failed attack in Germany
Scientists have created the first quantum material that can sort and transport different quantum states of light at room temperature, potentially removing the need for bulky, ultra-cold refrigeration systems。 Built from a gold film carved with hundreds of microscopic structures, the ultrathin “metacrystal” acts like a filter that directs different
It took just hours for the leading pharma group to reject the recommendations
Hot hours of summer could triple in the coming decades
Dark matter particles may exert a hidden force on one another, but its effects are stranger than expected。 An extra attraction helps the particles cluster, yet it also makes dark matter effectively lighter as the Universe expands。 That weakens its gravitational impact and usually slows the growth of cosmic structure rather than accelerating it