PAPER PLAINE

Fresh research, simply explained. Updates twice daily.

ShareLock: A Stealthy Multi-Tool Threshold Poisoning Attack Against MCP

How attackers hide harmful instructions across multiple AI tools to avoid detection

Researchers discovered a new attack method called ShareLock that splits malicious instructions into harmless-looking pieces hidden across multiple AI tools, making it nearly impossible to spot by inspection or automated systems. The attack succeeded over 90% of the time against major AI language models and only requires a small trigger during a routine system update to activate the hidden instructions and steal data or compromise systems.

As AI assistants increasingly rely on external tools and integrations to function, this attack reveals a critical blind spot in how these systems are secured. Organizations deploying AI agents with tool access need to know that current inspection methods—manual review or automated scanning—won't catch poisoning hidden across multiple tools, which means their defensive strategies are currently insufficient and could leave them vulnerable to data theft or system compromise.

Fast algorithms for learning a Gaussian under halfspace truncation with optimal sample complexity

Learning hidden patterns from incomplete data faster than previously possible

Researchers created a faster algorithm for recovering the true shape of a high-dimensional dataset when only a partial view is available. The new method uses the minimum amount of data theoretically possible and runs at the speed of basic matrix operations—improvements over the previous best approach, which was slower and required more samples.

Many real-world datasets are naturally filtered or incomplete: sensor readings might only record values above a threshold, survey responses might exclude certain groups, or observations might be restricted to a subset of space. This algorithm makes it practical to recover accurate statistical models from such truncated data without the computational slowdown of previous methods, potentially improving everything from medical imaging to climate modeling where observations are naturally limited.

Embedding Foundation Model Predictions in Discrete-Choice Models with Structural Guarantees

Making AI predictions follow economic rules without sacrificing accuracy

Foundation models predict choices well but often violate basic economics—suggesting that raising prices increases demand, or that unavailable options have some probability of being chosen. Researchers created a two-stage adapter that embeds foundation model predictions into an economic model while mathematically guaranteeing that the results follow economic logic, gaining an average 6.4 percentage point accuracy boost while maintaining 100% cost monotonicity.

Transportation agencies, retailers, and economists use choice models to estimate how people respond to prices and policies. An accurate model that also obeys economic logic is more trustworthy for real decisions—whether predicting traffic patterns after a toll increase or estimating consumer welfare. This method lets organizations use faster, more accurate foundation models without sacrificing the economic guarantees that justify policy reliance on their outputs.

The Geometry of Updates: Fisher Alignment at Vocabulary Scale

A faster way to pick the best training data for specialized AI models

When training language models on specialized data like DNA sequences or protein structures, picking the right source material is usually slow and expensive. Researchers developed FisherSketch, a method that identifies which training datasets will transfer best to a new task without needing to actually train the models—reducing the signature size needed to just 16 kilobytes while capturing the actual learning patterns that matter.

For scientists working with specialized sequences in biology and chemistry, this cuts the cost of selecting training data from hours of computation to seconds. The technique also reveals whether models learn from data patterns, prediction errors, or how those interact—giving researchers insight into what makes transfer learning succeed or fail in their domain.

Empowering GUI Agents via Autonomous Experience Exploration and Hindsight Experience Utilization for Task Planning

Training smaller AI models to plan complex computer tasks better than much larger ones

Researchers developed a method that lets smaller AI models learn to navigate websites and complete tasks by autonomously exploring environments and reusing past experiences as training data. A 7-billion-parameter model trained this way outperformed a much larger 32-billion-parameter commercial model, reaching 30.6% accuracy on real-world benchmarks. The breakthrough came from focusing training on high-level task planning rather than low-level individual skills.

Smaller AI models are cheaper to run and keep user data private, but they've struggled with planning complex multi-step tasks on websites and generalizing to new situations. This work shows they can match or beat much larger commercial models when trained the right way—meaning organizations could deploy capable web automation agents without expensive hardware or privacy concerns, while still handling unfamiliar websites and task variations they've never seen before.

Exact subsystem dynamics in the deterministic Floquet-PXP model

How to track quantum systems' behavior when their surroundings act like a hidden bath

Physicists have figured out how to exactly predict how parts of certain quantum systems evolve over time, even when those parts are constantly jostled by the rest of the system around them. They showed that a specific type of quantum rule called Rule 201 can be solved mathematically using a compact mathematical structure, making it possible to calculate how quantum properties change moment by moment without the usual explosion of complexity.

Most quantum systems are far too complicated to predict exactly—the math becomes impossibly tangled as time goes on. Finding even one solvable case, like Rule 201, gives physicists a testbed for understanding how quantum information spreads and gets scrambled in realistic situations. This matters for designing quantum computers and sensors, where controlling how quantum states degrade is essential.

Time-dependent weighted directed networks of cryptocurrency interaction from high-frequency returns

Which cryptocurrencies drive the market—and how that power shifts over time

A analysis of five years of cryptocurrency price data reveals a shifting pecking order of influence: Ethereum has become the market's most powerful trendsetter, while Bitcoin's sway has weakened. The researchers mapped these relationships by tracking which cryptocurrencies' price movements predict others', uncovering a tiny handful of coins that shape the entire market.

Traders and fund managers who want to anticipate crypto market moves need to watch the right assets. Bitcoin is no longer the dominant signal it once was—ignoring Ethereum's emerging role would mean missing early warnings of broader market shifts. This snapshot of a fast-changing hierarchy also flags a real risk: when power concentrates in a small number of assets, the entire market becomes more fragile.

Autoregressive Boltzmann Generators

Using language-model techniques to simulate protein behavior faster

Researchers created a new method called Autoregressive Boltzmann Generators that simulates how proteins behave at equilibrium—a crucial problem in chemistry and drug discovery. The approach borrows techniques from large language models to sidestep the mathematical limitations of previous methods, achieving 60% better accuracy on standard tests and working significantly faster on larger protein systems.

Simulating protein behavior accurately is essential for drug design, materials science, and understanding biological processes. The speedup and accuracy gains mean researchers can test more drug candidates computationally before expensive lab experiments, potentially accelerating the discovery of new medicines and reducing development costs.

Pseudo-Text-Conditioned 3D Grounding DINO for Organ Localization in Abdominal CT

Teaching CT scanners to automatically spot organs without text labels

Researchers created a lightweight AI system that locates five abdominal organs (liver, spleen, both kidneys, and bowel) in CT scans by using simplified text-like tokens instead of full language descriptions. The system achieved 58% accuracy on finding organs in the right general location, but struggled with pinpointing exact boundaries—suggesting it's useful for initial navigation through scans but needs refinement for precise surgical planning.

Automatically finding organs in trauma patients' CT scans could speed up emergency diagnosis and help surgeons plan interventions faster when minutes matter. The open-source baseline the authors released gives hospitals and researchers a starting point to improve organ detection, potentially reducing the time radiologists spend manually locating structures before analyzing injuries.

Finding Stationary Points by Comparisons

Finding mathematical sweet spots using only yes-or-no comparisons

Researchers developed a new method to find stationary points—places where functions flatten out—when you can only ask a computer "which of these two values is bigger?" instead of calculating exact function values. The approach uses roughly 10,000 times fewer queries than naive methods for typical problem sizes, and a quantum version cuts that further by a factor equal to the square root of the problem's dimensions.

Many real optimization problems can only be queried through comparisons—ranking models, A/B testing, or noisy systems where you can compare outcomes but not measure them precisely. This algorithm makes it practical to find good solutions in those scenarios. The quantum version hints at how quantum computers might eventually offer speedups for real-world optimization beyond brute-force advantage.

TriViewBench: Controlled Complexity Scaling for Multi-View Structural Reasoning in MLLMs

Why AI vision systems fail when objects hide and multiply across views

All 18 major AI vision systems tested share the same weakness: they handle simple visual questions well but collapse catastrophically when asked to count objects (59% accuracy drop) or understand complex 3D scenes (80% drop). The failures stem from two distinct problems—the systems either miss hidden objects or confuse the same object across different camera angles—and simply asking them to "think step by step" doesn't help.

AI systems that see are being deployed in robotics, autonomous vehicles, and industrial inspection, where missing hidden objects or misidentifying items across viewpoints could cause real failures. This benchmark reveals these systems have a fundamental blind spot that current prompting tricks can't fix, suggesting engineers need to rebuild how these systems represent 3D space rather than just improve their reasoning.

Extended pseudo-spectral physics-informed neural networks for phase-field models

Using AI to reverse-engineer the hidden rules that drive material separation

Researchers developed a machine-learning method that can figure out the underlying physical laws of phase separation—the process where mixtures split into distinct regions—by watching how materials evolve over time. The technique recovers unknown physical parameters from just a single or handful of snapshot pairs, and continues to work reasonably well even when data is noisy.

Materials scientists often can't directly measure the fundamental properties that control how materials separate and form patterns, but need to know them to design alloys, polymers, and other engineered materials. This method cuts the number of observations needed to infer those hidden rules, potentially speeding up materials discovery and making it cheaper to characterize new substances without running expensive, lengthy experiments.