Eye on AI Weekly Research Watch

Getting multiple AI agents to work together effectively in a shared physical environment is harder than it sounds — agents frequently act on outdated assumptions about their partners or issue redundant, mistimed communications. LLawCo addresses this by having agents reflect on past failures to extract high-level "laws of cooperation," such as knowing when to speak and when to wait, then fine-tunin...

Predicting how hard an exam question will be for human test-takers — without running expensive human trials — would transform educational assessment. This paper proposes using the reasoning traces of large language models as a proxy for human cognitive effort. Rather than treating these traces as raw text, Epi2Diff structures them into meaningful "cognitive episodes" — functional states like plann...

Remittances — money sent home by migrant workers — are a lifeline for many developing economies, yet surprisingly hard to forecast reliably. This study applies rigorous time-series and machine learning methods to 32 years of Sri Lankan migration and remittance data, finding that external factors like exchange rates and global oil prices drive inflows far more than domestic indicators. A multivaria...

Building robots that can understand and interact with the physical world requires massive amounts of 3D training data — but capturing that data with multi-camera rigs is expensive and impractical at scale. HAT-4D proposes using ordinary monocular video as a data source, reconstructing the 3D geometry and temporal dynamics of multiple interacting objects with the help of vision-language models and...

As AI systems increasingly act as proxies for human stakeholders in shared learning environments, a thorny question arises: how do you fairly reward each participant's contribution when different contributors have different values — and when some contributions might violate those values? This paper proposes a framework that filters gradient updates by each principal's value profile before computin...

Flow matching is a powerful framework for generating images and other data by learning to map noise to structure, but it suffers from a training-inference mismatch: models are trained on clean trajectories but must operate on drifted ones at test time. DEFAR turns this problem on its head, treating the drift itself as a useful signal. It uses the bias to learn corrective directions and to reinforc...

When AI agents autonomously write and merge code at scale, the usual way of evaluating them — task by task, in isolation — misses something important: the cumulative friction and technical debt that builds up in shared codebases over time. Studying over 930,000 agent-authored pull requests, this paper finds that about half of "integration friction" is a property of the repository ecosystem, not an...

Why do bigger neural networks tend to perform better — and by exactly how much? Scaling laws attempt to answer this, but most existing theory relies on simplified assumptions about infinite width or unlimited data. This work studies how generalization error changes as both model width and dataset size vary simultaneously in a tractable two-layer network, revealing a phase diagram with distinct reg...

Detecting defects in manufactured goods — a crack in a circuit board, a tear in fabric — requires models that can generalize across wildly different visual conditions. TopoTTA brings an unusual tool to this problem: persistent homology, a mathematical framework that captures the shape and connectivity of structures across scales. Rather than relying on simple pixel-confidence thresholds, it uses t...

As AI agents gain the ability to use tools, access memory, and coordinate with other agents, they become vulnerable to entirely new classes of attacks — malicious instructions injected through tool outputs, poisoned memory, or compromised peer agents. ANIS proposes a defense architecture modeled on the biological immune system, embedded directly inside the agent's reasoning process rather than bol...

Cellular networks in cities are under constant, unpredictable stress — traffic jams, concerts, and commutes all reshape how and where data flows. Predicting this demand accurately is essential for carriers to allocate bandwidth intelligently. PEHT introduces a transformer-based model that separates core network traffic signals from external urban mobility and congestion data, then fuses them effic...

The volume of scientific papers being published is growing faster than human reviewers can keep pace with — a crisis accelerated by AI-assisted research generation. This paper proposes a taxonomy of AI-human collaboration levels in peer review, then introduces PAT (Paper Assistant Tool), an agentic system that reads full manuscripts and produces structured evaluations, including checks of mathemat...

Designing computer chips is extraordinarily complex, requiring expertise across logic, timing, and physical layout — making it a compelling frontier for AI automation. HORIZON treats hardware design the same way modern AI treats software: as an evolving codebase that an agent can iteratively improve. By wrapping design tasks in a structured "project pack" with executable evaluators, the agent refi...

In competitive games — from poker to cybersecurity — there isn't always a single optimal strategy, but rather a whole family of equally valid equilibria. Which one an AI solver picks can quietly determine how it behaves against opponents who don't play perfectly. This work reveals that the choice of algorithm, not random chance, systematically drives which equilibrium gets selected. Regularized me...

Robotic hands capable of dexterous manipulation have made impressive strides, but teaching a single hand to perform multiple tasks simultaneously — without one task undoing another — remains a hard unsolved problem. Imagine a robot hand that already knows how to hold an object securely; asking it to also open a latch might cause it to loosen its grip. DexCompose addresses this by assigning explici...

Building a high-quality speech synthesis system typically requires training multiple specialized models independently, then orchestrating them at inference time — an expensive and memory-intensive process. This paper explores a more compact path: starting with a speech classifier already trained to recognize acoustic properties, and attaching a lightweight generative subnetwork that reuses its int...

IVF success rates are influenced by countless variables, but the physical conditions inside laboratory incubators — temperature stability, humidity adherence, recovery speed after disturbances — have historically been modeled crudely if at all. This paper demonstrates that richly engineered temporal features from environmental sensors, combined with a hierarchical Bayesian model that pools informa...

As AI agents gain access to tools with real-world consequences, attackers have begun automating their jailbreak campaigns — using language models to generate, evaluate, and refine prompts at scale. Standard defenses that simply refuse suspicious inputs inadvertently help attackers by providing clear feedback signals. This paper proposes a counterintuitive alternative: rather than blocking detected...

Language model agents that maintain long, multi-turn conversations place enormous pressure on GPU memory, primarily because the key-value cache — a stored record of prior context — grows with every exchange. At scale, this becomes a bottleneck that throttles how many users a system can serve simultaneously. UltraQuant attacks this problem with aggressive 4-bit compression of the KV cache, achievin...

As AI systems increasingly coordinate in networks — fleets of trading agents, swarms of robotic systems, distributed planning architectures — questions about collective behavior become urgent. When should agents synchronize tightly, and when should they maintain independence? This paper develops a formal framework borrowing concepts from physics and economics, modeling collective outcomes in terms...

Multi-agent systems that use language models to evaluate each other's outputs are gaining traction in automated research, code review, and content moderation pipelines. But when one agent's bias influences another's, errors can compound silently across the network. This paper formalizes that risk with the Contagion Networks framework, measuring how systematically biased evaluators propagate their...

Security teams are increasingly exploring whether large language models can automatically detect vulnerabilities in source code — a task with serious consequences if done poorly. This paper delivers a sobering assessment: even fine-tuned models that score well on benchmarks may be learning surface-level patterns rather than genuine security reasoning. Using carefully curated Linux kernel samples w...

Generative image models are increasingly asked to do something cognitively demanding: take the content of one image and the style of another, and fuse them seamlessly without letting either bleed into the wrong dimension. This is harder than it sounds — style references tend to smuggle in unwanted structural or semantic content. FreeStyle approaches this challenge by mining the large community eco...

Jailbreaks via in-context examples are a known vulnerability of language models, but the underlying mechanics have remained murky. Why does showing a model a few harmful exchanges cause it to comply with further harmful requests? This paper dissects the phenomenon carefully, mixing benign and harmful demonstrations to isolate what models actually extract. Surprisingly, benign demonstrations can ei...

AI agents operating in enterprise environments — browsing the web, calling APIs, reading files — must be constrained by security policies. Prior work on policy enforcement assumed those policies were deterministic, but real tools like PII detectors or content classifiers have inherent failure probabilities. This paper introduces a framework grounded in distributionally robust optimization that pro...