PAPER PLAINE

Fresh research, simply explained. Updates twice daily.

Fusion Training for Mathematical Generalization in Large Language Models

Training AI to think deeply and answer quickly at the same time

Teaching language models to both reason through hard problems and give quick answers creates a tension between the two skills—more training on fast answers actually weakens the model's ability to think carefully. The order and balance of training matter: the right schedule can reduce this damage, but the best approach depends on how much of each type of training data you use.

As AI assistants take on both simple customer-service tasks and complex problem-solving work, companies need both speed and accuracy. This research shows they can't maximize both equally in one model, and gives engineers concrete guidance on how to balance this trade-off depending on their actual needs.

Bias-robust causal inference for panel data

Measuring treatment effects while accounting for guesswork in missing data

Economists often estimate what would have happened to people without a treatment by filling in missing data—but that guesswork can distort the final answer. This paper introduces a method that catches and corrects for this hidden error, reporting wider but more honest confidence intervals. The approach outperforms standard alternatives, especially when data is limited, keeping its accuracy promise even when the unobserved data patterns are partially misspecified.

Policy decisions about job training, tax credits, or health programs often rest on estimates from observational data where the counterfactual is guessed. Traditional methods claim narrow confidence intervals but deliver false certainty—the coverage is nearly zero when data is sparse. This method trades some precision for honesty: its stated margins actually contain the true answer, making it safer for policymakers to rely on.

Context Is Not Authority: Structured Runtime Governance for Financial Market Agents

Stopping financial AI from acting on outdated or unauthorized information

Financial trading systems can misinterpret instructions or act on stale data, turning correct information into wrong trades. SAGE-Fin is a control system that verifies what a financial agent actually proposes to do—not just what it claims—and rechecks permissions whenever market conditions, account status, or policies change. Real-world testing at a digital-asset platform confirmed it caught cases where agents would have acted on outdated or missing authorization.

Financial errors move fast and cost real money. A trading system that silently acts on yesterday's permissions or misread instructions can execute unauthorized trades before anyone notices. SAGE-Fin forces the system to prove at runtime that each proposed action is still authorized, catching drift and stale data before execution. The platform's independent operations team found it practical enough to deploy on live customer requests, suggesting it can reduce the window where bad trades slip through.

Multimodal Model Diffing for Feature Discovery and Control

Opening the black box of AI vision models to see and control what they learn

Researchers created a method to identify which internal features of AI vision-and-language models drive specific behaviors—and then selectively turn those features on or off. By comparing how a base language model changes when trained on images, they discovered sparse, interpretable features that control tasks like reading text in images and understanding spatial layouts, and showed they could degrade unwanted behaviors by up to 24% without breaking general performance.

As AI systems become more powerful and are deployed in real applications, being able to audit what's actually driving their decisions matters for safety and trust. This work provides a concrete toolkit for developers to find problematic features (like those underlying adversarial attacks) and remove them, rather than guessing blindly at how to make these systems behave. It's the difference between adjusting dials you can see versus hoping a sledgehammer approach doesn't break something important.

FinRank: An Evidence-Grounded Benchmark for Financial Question Answering and Retrieval over SEC Filings

Spotting when financial AI finds the right answer in the wrong place

A new test for financial AI reveals a hidden problem: systems can answer questions about company finances correctly while pointing to the wrong evidence in SEC filings. The benchmark includes 1,185 real questions about company reports, with deliberately tricky wrong answers drawn from similar facts elsewhere in the same filing, earlier periods, or competitor companies. Even advanced AI systems struggle, with the best reaching only 45% accuracy when forced to find the right evidence, and dropping 13–20 percentage points when tested on hard-to-distinguish wrong answers.

Investors, regulators, and analysts increasingly rely on AI to search financial documents for facts about companies. If an AI finds the correct number but attributes it to the wrong quarter or the wrong company, someone making a million-dollar decision based on that answer could lose everything. This benchmark forces developers to build systems that not only get the right answer but prove it came from the right place—making AI-powered financial research trustworthy enough to act on.

PsychoAgent: An Affect-Sensitive Cognitive Architecture for Conflict-Aware Memory in LLM Agents

Making AI remember emotional weight, not just matching keywords

A new system for AI assistants splits memory into facts and emotions, then retrieves past information based on what matters psychologically—not just topical similarity. In three conflict scenarios, this approach recovered conflict-critical details 40% more often than systems that treat all memories equally, though human raters couldn't consistently tell the difference in final conversation quality.

AI assistants today retrieve information like search engines: if you mention a topic, they pull up matching facts. But humans retrieve memories shaped by unresolved conflicts and emotional weight—what still bothers us. This work demonstrates a mechanism for building AI that prioritizes emotionally significant past events, potentially creating assistants that navigate complex, sensitive situations with better context awareness and more natural conversation flow.

SoK: Cryptographic Key Recovery for Cryptoasset Custody and Financial Technologies

How to safely recover access to cryptocurrency wallets when keys are lost

When someone loses access to a cryptocurrency wallet, recovery mechanisms let them regain control—but they vary widely in how they work and where they place risk. This paper analyzed 77 different recovery systems across custody platforms and financial technologies, revealing that most recovery methods shift trust to third parties, create new abuse vulnerabilities, and leave recovery metadata unprotected.

Cryptocurrency users routinely lose access to accounts containing real money, and recovery is often their only option—yet most recovery systems force a dangerous choice between security and recoverability. This taxonomy helps wallet providers and developers understand the tradeoffs in existing approaches and identifies six specific gaps (like unprotected recovery metadata and poor lifecycle management) that create theft vulnerabilities. Better recovery design could prevent both devastating user losses and the security breaches that plague custody platforms today.

On the Optimal Laplacian Jordan Structure for Synchronizability

Finding the best network structure for synchronized systems to settle faster

When multiple systems are connected in a network, they can synchronize — lock into matching patterns — but how fast they sync depends on the network's mathematical structure. This paper identifies the precise network configurations that guarantee synchronization happens both reliably and as quickly as possible, with the added benefit of resisting small perturbations.

Synchronized networks appear everywhere: power grids must stay in sync to avoid blackouts, groups of robots coordinate through networked communication, and biological systems like fireflies flashing together rely on synchronization. By determining which network structures sync fastest and most robustly, engineers can design more stable and responsive systems across these applications.

SkillProx: Self-Evolving Agent Skills via Proximal Textual Gradient Descent

Teaching AI agents to build and refine their own instruction libraries

When AI language models tackle similar tasks repeatedly, they can build up a library of reusable instructions—called skills—that live in the model's working memory rather than changing its weights. A new system called SkillProx improves how these skills evolve by adding explicit diagnosis of what goes wrong and a structured way to prune unhelpful knowledge, achieving 3 percentage points higher accuracy than previous methods across multiple benchmark tests.

As AI agents take on more complex real-world tasks, the ability to learn and refine their own strategies becomes critical. SkillProx makes this learning process more transparent and efficient—you can audit which knowledge pieces actually help—and shows the approach generalizes to tasks the system has never seen before. This points toward AI systems that improve themselves more reliably without expensive retraining.

Efficient Discrete Position Design for Movable Antenna Systems: Low Complexity and Robustness

Making wireless antennas that move around work without slowing down networks

Wireless antennas that can physically shift position promise better communication, but figuring out where to place them requires enormous computational effort. Researchers proved that the positioning problem has a special mathematical structure that lets them find near-optimal placements 34 times faster than existing methods, while still achieving 90% of the best possible performance even when channel measurements are slightly wrong.

Movable antenna systems could significantly boost wireless network capacity and reliability, but they've been impractical because the positioning calculations would overwhelm real systems. This algorithm makes the computation fast enough for actual deployment without sacrificing performance—a necessary step before movable antennas appear in real phones, base stations, and 5G/6G networks.

FedDOSE: Federated Learning Framework Decomposing Site Effects for Modeling Brain Dynamic Functional Connectivity

How hospitals can share brain scans without revealing patient data

Researchers developed FedDOSE, a new method that lets hospitals collaborate on analyzing brain scans while keeping patient information private. The system accounts for differences between hospital scanners and captures how brain connectivity changes over time, improving accuracy in diagnosing autism and ADHD by outperforming existing approaches.

Brain disorder diagnosis currently requires pooling data from multiple hospitals—but sharing raw scans raises privacy concerns. FedDOSE lets hospitals train AI models together on their local data without sending scans elsewhere, while still achieving better diagnostic accuracy. This could accelerate research into neurodevelopmental disorders without compromising patient privacy.

Tracing the Heart: An Evidence-Linked Pipeline for Heart-Failure Feature Engineering

Automating the detective work that turns messy medical records into AI-ready data

Researchers built an AI system that automatically extracts and structures heart-failure data from fragmented medical records, a task that currently consumes nearly half of clinical data scientists' time. Testing on 500 patient records, the system created features that boosted predictive accuracy from 89.5% to 96.3% for one heart-failure type, while leaving an auditable trail showing exactly where each piece of data came from and why it was included.

Heart failure affects 6.7 million Americans, and developing better predictive AI requires months of tedious manual data work that slows research. If this approach scales, it could cut months off the time needed to build and validate heart-failure detection tools, freeing data scientists to focus on clinical strategy rather than data plumbing. The system's transparency—showing evidence and reasoning for every data choice—also matters for hospitals that need to defend AI decisions to regulators and doctors.