Science – Page 14

Unmasking Deepfakes: A Hidden Code for Image Authenticity

01.03.2026 by qfx

$A method embeds content-related information into images via quantized multi-scale tokens-created with a VQ-VAE and constrained by watermark capacity [latex] |h| \leq |m| [/latex]-enabling recovery of deepfakes even after malicious manipulations like object removal or inpainting, achieved through decoding the watermarked image to extract hidden tokens and generate a deepfake localization map [latex] M_{loc} [/latex].$

Researchers have developed a novel watermarking technique that embeds a multi-scale ‘fingerprint’ within images, enabling both deepfake detection and faithful recovery of original content.

Seeing Beyond the Surface: Closing the Gap in Image Reflection Removal

01.03.2026 by qfx

The study demonstrates that the proposed GFRRN method achieves competitive performance against state-of-the-art techniques when applied to image samples from both the SIR2 and Real20 datasets, suggesting its robustness across diverse visual conditions despite the inevitable challenges of real-world application.

A new network architecture tackles the challenges of removing reflections from single images, offering improved clarity and realism.

Lost in Translation: When Language Style Vanishes in Images

01.03.2026 by qfx

Despite explicit descriptive cues-including color, texture, and viewpoint-large language models consistently fail to synthesize images accurately reflecting the original visual details, manifesting in broadly similar outputs across models and a systematic inability to capture nuanced attributes like precise color or consistent perspective.

New research reveals that stylistic nuances present in text generated by language models are often lost when those models are used to create images, exposing a critical limitation in cross-modal AI.

Reading Minds with EEG: A New Era of Visual Reconstruction

01.03.2026 by qfx

Researchers have developed a new framework that accurately reconstructs visual information directly from brain activity recorded via electroencephalography (EEG).

Beyond the Limit: Reclaiming Data from Signal Saturation

28.02.2026 by qfx

$A reconstruction network, when burdened by inherent bias, falters in accurately representing signals - particularly those with low amplitude - but achieves natural homogeneity and improved performance when that bias is removed, as demonstrated by its ability to faithfully reconstruct both a signal [latex]\bm{x}[/latex] and a scaled version [latex]\frac{\bm{x}}{10}[/latex].$

A new self-supervised learning technique restores information lost when signals exceed measurement limits, opening doors for improved audio and image processing.

Beyond Autoregressive Bias: Why Parallel Generation Remains a Challenge for Diffusion Models

28.02.2026 by qfx

Conventional decoding methods, despite employing confidence-based strategies, ultimately revert to a strict left-to-right, autoregressive generation pattern-a limitation overcome by a novel approach that simultaneously explores multiple reasoning trajectories, effectively bypassing the single-stream bottleneck and enabling a more expansive decoding process →.

A new analysis reveals that diffusion language models struggle with truly parallel decoding due to inherent sequential dependencies learned from typical training data.

Smarter Product Search: Boosting E-commerce with Knowledge-Powered AI

28.02.2026 by qfx

Retrieval-Augmented Generation systems architect a pathway for knowledge integration, combining the strengths of pre-trained language models with information retrieved from external sources to mitigate the inherent decay of static datasets and foster ongoing relevance.

New research demonstrates how combining efficient data retrieval with intelligent reranking can dramatically improve the accuracy of product recommendations and information access in online shopping.

Automated Benchmark Creation Boosts Hardware Verification

28.02.2026 by qfx

EvolveGen establishes a workflow for iterative refinement, systematically evolving generative models through a cycle of proposal, evaluation, and adaptation-a process designed to push the boundaries of what’s currently achievable.

A new reinforcement learning framework automatically generates complex hardware designs to rigorously test formal verification tools.

Reasoning Without Limits: A New Approach to Adaptive Thinking

28.02.2026 by qfx

A two-stage training pipeline initializes a unified policy through hybrid fine-tuning on paired thinking and non-thinking data, then stabilizes optimization-even with significant variations in sequence length-using gradient regulation and correctness-preserving advantage shaping within a reinforcement learning framework to effectively determine when to engage in deliberate thought processes.

Researchers have developed a framework to enhance the reasoning capabilities of large models, allowing them to tackle complex problems with improved efficiency and accuracy.

Can AI Spot Deepfakes? FactGuard’s New Approach to Video Truth

28.02.2026 by qfx

FactGuard distinguishes itself from existing video misinformation detection methods through both enhanced explainability and a demonstrated improvement in overall performance metrics.

A new agentic framework uses artificial intelligence to actively investigate and verify the authenticity of video content, moving beyond passive detection.