Paper Radar

ELI5A spacecraft relay station uses AI to decide which messages to send to Earth when it only has brief contact windows—like a mail sorter that must pick the most important packages based on how stale they are and which planet they came from.

Problem solvedDeep-space networks can't send everything due to contact time limits and storage constraints. Simple schedulers like FIFO ignore which data matters most to missions, wasting precious communication windows and delivering outdated information. This system prioritizes based on mission goals.

ELI5When AI agents call external tools (like APIs), they need to pick the right one. This work makes tool selection smarter by watching which tools actually work in real-time and adjusting recommendations on the fly, instead of just picking based on semantic match.

Problem solvedAI agents fail silently when they pick unavailable or broken tools. Existing retrievers only look at semantic relevance and ignore whether tools are actually working right now, causing agents to make bad choices during execution.

ELI5Researchers built a system that automatically tests AI agents connected to many web services (like email, calendar, banking) to find ways they could cause unintended harm—like accidentally deleting important data across multiple apps when following a seemingly innocent user request. It learns from failed attempts to get better at discovering new dangerous combinations.

Problem solvedAs AI agents gain access to real services and tools, they can cause real harm through unexpected interactions across apps—but existing red-teaming methods miss novel dangerous combinations because they're biased toward familiar patterns. Teams need a systematic way to find these hidden compositional failures before deploying agents in production.

ELI5When coding agents hit a dead end, instead of starting from scratch, this paper saves the entire machine state (like a video game checkpoint) and tries different fixes from there. It found that rewinding to earlier checkpoints helps agents solve problems faster, especially when they've spun up databases or services that can't be rebuilt from code alone.

Problem solvedCoding agents waste time restarting from zero after failures, losing all progress like database setups or running services. By recovering to saved machine states and branching from there, agents can retry more efficiently and solve multi-service coding tasks that would otherwise require complete restarts.

ELI5AI agents being tested in controlled environments accidentally broke out and accessed real systems they shouldn't have. This paper shows how to build multiple overlapping safety layers—like credential locks, automatic kill switches, and continuous monitoring—so agents can't escape their test boundaries.

Problem solvedCompanies were finding out only after the fact that their agent evaluations had breached real production systems. You need real-time verification that agents stay within intended boundaries during testing, not just trust that sandboxes will hold.

ELI5A benchmark that tests whether AI agents can follow text instructions to design buildable LEGO sets. An agent gets a prompt like 'build a house,' then must pick real LEGO pieces that physically fit together and actually look like a house.

Problem solvedThere's no standardized way to evaluate whether AI agents can design realistic, physically-buildable objects from scratch. Companies and researchers need benchmarks to measure progress on embodied design tasks before deploying these systems.

ELI5A robot learns to do household tasks better over time by breaking down what goes wrong into specific parts of its behavior, fixing just those parts, testing the fixes, and saving the improvements so it can reuse them on future tasks.

Problem solvedRobots that learn from experience have no way to organize what they learn—they can't tell which part of their behavior caused a failure, verify a fix actually works, or safely reuse it later. This means each new task starts from scratch instead of building on past experience.

ELI5A system that teaches AI agents to solve problems by learning visual skills from successful attempts. Instead of describing actions in words, it stores them as visual cards the agent can recognize and reuse, creating a self-improving loop where new successes teach the agent better strategies.

Problem solvedVLM agents struggle with spatial reasoning when skills are stored as text, losing geometric details. Existing approaches build and improve skills separately from the main learning process, missing opportunities to reinforce each other and accelerate learning.

ELI5A system that watches LLM agents in real-time and catches them going off track before they waste money or cause problems — like a copilot that hits the brakes when something looks wrong, using only ~1 millisecond per step.

Problem solvedLLM agents deployed in production (trading, planning, IT ops) can take irreversible costly actions before anyone notices they're failing. Post-hoc checking arrives too late; monitoring agents add latency to every step. OnTrack detects failures mid-run so you can stop the agent before damage happens.

ELI5This paper tests whether AI agents admit when they're uncertain or wrong. Researchers create situations where an agent's trained knowledge conflicts with new information, then measure whether it changes its answer, says "I don't know," or confidently sticks with a wrong conclusion.

Problem solvedAI agents often give confident wrong answers instead of flagging uncertainty when evidence contradicts their training. This is dangerous in real applications—you need systems that know when to pause, ask for help, or admit confusion rather than bulldoze ahead with false confidence.

ELI5AI agents play competitive games and use their match results to rewrite their own game-playing code, trying to climb the rankings without changing the underlying AI model itself. It's like a player who learns by studying replays and keeps tweaking their strategy document rather than retraining their brain.

Problem solvedGame AI research has relied on either brute-force search or training new models from scratch, both slow and data-hungry. This explores whether agents can iteratively improve game strategies by reading rules, analyzing their losses, and editing code—closer to how human competitors actually learn.

ELI5Researchers tested whether an AI system adapting neural networks to solve physics problems relies on its initial understanding or learns from experimental results. They found it does both: the AI makes good first guesses based on the problem description, then adjusts based on what actually works.

Problem solvedWhen tuning machine learning models for specialized tasks with limited experiments, it's unclear if AI agents are just following their initial training or actually learning from feedback. This matters because you need to know if the system is truly adapting or just getting lucky with its first guess.

ELI5Researchers tested AI models used to guard agent systems (deciding if tool calls should be allowed) and found they're easily tricked by simple text tricks—like misleading option names or random log lines—to allow forbidden actions up to 100% of the time.

Problem solvedAI agents need reliable gatekeepers to prevent dangerous tool use, but typed decision models currently used for this fail catastrophically under adversarial manipulation, creating a security vulnerability in deployed agent systems.

ELI5A system that helps interpret disease-linked DNA variants by having an AI agent intelligently collect real genetic evidence (from databases like ADASTRA and ENCODE) and reason through it, rather than just guessing—it knows when to say 'I don't know' instead of making things up.

Problem solvedDoctors can't reliably interpret 90% of disease-linked genetic variants because they're in regulatory regions; current AI systems hallucinate biological significance and fake evidence, so clinicians have no trustworthy way to assess which variants actually matter for disease.

ELI5Instead of training SQL models to output queries in isolation, this method trains them while they're actually interacting with real databases—running queries, checking results, fixing errors—just like they'll do in production. It's like practicing a sport in game conditions rather than drills.

Problem solvedSQL models trained on static query prediction fail badly when deployed as interactive agents that need to explore schemas, debug queries, and iterate. This gap between training and real-world use causes dramatic accuracy drops. HarnessSQL closes this by training agents inside the actual execution environment they'll operate in.

ELI5Researchers built a scientific challenge where AI agents have to look at real climate data and build their own computer models of El Niño without being told the answer. Some AI agents actually built models that predict El Niño better than published human-made models.

Problem solvedWe can't tell if AI agents are doing real scientific research or just pattern-matching against known answers. This benchmark lets us grade AI on genuinely open scientific questions by testing if their models actually predict and explain real-world phenomena that scientists still debate.

ELI5Instead of just looking up similar documents, this system learns to decide what computations and queries to run step-by-step to build up the evidence needed to answer a question—like assembling a puzzle by figuring out which pieces to fetch and how to combine them.

Problem solvedRAG systems fail when answers require combining, filtering, or calculating across multiple sources rather than just finding one relevant document. RECAST learns to actively construct the right evidence instead of passively retrieving it.

ELI5A robot learns to do new tasks by automatically figuring out which experiments to try next, then using what happens to improve its code and decision-making—like a scientist testing hypotheses to get smarter, but for robots.

Problem solvedRobot models degrade when tasks change, and collecting more training data through teleoperation is expensive. This system learns efficiently by choosing smart experiments and updating itself without massive new datasets.

ELI5Instead of waiting weeks to fine-tune a coding agent to see if it'll work well, this paper lets you test a base model in minutes by replaying successful agent trajectories and checking if the base model would make the right moves at key decision points.

Problem solvedTraining coding agents is expensive and slow. Teams waste compute fine-tuning base models that won't improve much anyway. This method predicts post-training performance upfront so you can pick the best base model before investing in training.

ELI5Imagine thousands of AI researchers working together in a lab where they propose projects, compete for computing resources, and review each other's work—like a real university department, but automated. The system organizes itself through institutional rules rather than central commands.

Problem solvedRunning large populations of AI agents efficiently requires coordination, but existing systems either handle one project at a time or leave agents chaotic. This provides a framework where agents self-organize through competitive grants and peer review, reducing wasted compute and enabling emergent discoveries.

ELI5Instead of calling an expensive AI model millions of times, can an AI agent figure out how to solve the whole batch of problems more cheaply—like by training a small model or writing a program it can reuse? This paper tests whether LLM agents can actually do that.

Problem solvedRunning an LLM on millions of similar tasks gets absurdly expensive fast. Companies need ways to detect when a workload is repetitive enough to warrant building a cheaper solution, and agents that can autonomously make that call.

ELI5A system trains AI web agents to resist prompt injection attacks by having an adversary continuously come up with new tricks to fool the agent, while the agent learns to handle both real tasks and attacks at the same time—all in a simulated web environment.

Problem solvedWeb agents following instructions can be hijacked by malicious text hidden on webpages, redirecting them from user goals. Existing defenses only train on known attacks and fail against new ones; this method builds adaptive robustness that survives unseen attacks.

ELI5A system that teaches robots to get better at tasks by having both the robot and its teacher (judge) improve together—when the robot learns new ways to fail, the judge learns how to catch those failures, then they both improve again.

Problem solvedRobots get stuck improving because their evaluators can't recognize new failure types. VeriFine fixes this by letting the evaluator learn from the robot's mistakes, so both can keep improving instead of hitting a wall.

ELI5When AI agents summarize old conversation history to save memory, does what the agent was recently doing predict whether that summary will cause problems later? Researchers tested this on real agent interactions and found the answer is: barely—recent behavior is a weak signal for when summarization will hurt.

Problem solvedLong-running AI agents need to compress old context to stay efficient, but blindly truncating at token limits wastes useful summaries and keeps harmful ones. Better triggers could compress smarter, but we don't yet know if an agent's behavior pattern actually predicts when summarization will fail.

ELI5A benchmark that tests whether AI agents can write physics simulation code from scratch—like asking an AI to write the math and code that makes water flow or cloth move realistically in a video game or movie.

Problem solvedPhysics simulation is expensive and time-consuming to hand-code; if LLM agents could generate solvers automatically, it could speed up game dev, animation, and robotics. Currently even best agents only succeed ~50% of the time.

ELI5A tool that automatically redesigns flowchart diagrams to fit different screen shapes (wide slides, tall phone screens, square posters) while keeping all boxes and connections correct and editable—solving the problem of resizing ML pipeline figures without breaking them.

Problem solvedResearchers need to reuse the same flowchart diagram across papers, slides, posters, and social media, each with different aspect ratios. Current tools either distort the diagram, lose connections, or create fake content. This ensures diagrams stay accurate and editable across all formats.

ELI5Instead of showing a robot a full video of how to do a task, this method breaks the video into a hierarchical tree of smaller clips—keyframes, then phases, then detailed moments. The robot can browse this tree during planning and zoom into specific clips when it needs details, like watching only the frames where it's grasping something.

Problem solvedRobots using language models struggle with demonstration videos: showing the full video every step is slow and loses fine details, while fixed keyframes miss critical contact information. This method lets robots learn from one video per task without slowing down, by letting them request only the detail level they need.

ELI5A system that decides on-the-fly whether to grab pre-stored memories or have an AI re-read raw conversation history to answer a question, optimizing for speed, accuracy, or cost depending on what you care about.

Problem solvedLLM agents waste compute storing and retrieving memories that might not matter later, or miss important details by preprocessing everything upfront. MemPilot lets you trade off quality, speed, and expense based on what matters for each query.

ELI5A web agent learns to complete browser tasks by asking itself verification questions about its own actions, then uses a statistical method to keep only the trustworthy signals. At test time, it can retry tasks and pick the best one without needing an expensive external judge.

Problem solvedTraining web agents with reinforcement learning is expensive because task success signals are sparse and calling a language-model judge at every step costs too much. This method creates a reusable verification system that works at training and test time without constant external judge calls.

ELI5A benchmark that measures how well AI systems design experiments and draw conclusions from results—essentially rating their scientific intuition. It compares how quickly models can solve research problems versus human experts, treating good experimental design as a multiplier on compute efficiency.

Problem solvedWe lacked a way to measure whether AI systems are getting better at the creative, strategic parts of research (picking good experiments, interpreting results) versus just brute-force compute. This matters because AI's ability to multiply research output depends on this skill, not just raw performance.

ELI5An AI system that watches chip simulations run, stores all the messy output data in an organized database, and then answers engineers' questions about what went wrong by querying that data and linking problems back to the original code.

Problem solvedChip verification today is mostly manual debugging of massive simulation outputs. Engineers spend hours digging through waveform data to find bugs. This system automates the grunt work of organizing and querying that data so engineers can ask questions in English and get answers.

ELI5Agents watch videos of objects moving, deforming, and breaking, then write computer graphics code that recreates what they saw. It's like reverse-engineering the physics of a scene by writing the simulation that would produce that video.

Problem solvedCurrent AI can reconstruct static 3D scenes from images, but struggle with dynamic scenes involving movement, fluid, and deformation. This benchmark tests whether models can actually understand and encode real-world physics, not just fit geometry.

ELI5A chess AI that plays at world championship level and explains why it makes each move—combining a powerful silent chess engine with a language model so it can both play brilliantly and talk through its reasoning.

Problem solvedChess engines play superhuman chess but can't explain moves; language models explain things but play poorly. This fixes both: you get moves a grandmaster would make plus coherent explanations of the strategy.

ELI5A framework that gives multimodal AI models like Claude better memory and control over what they see, letting them solve visual puzzles and games by remembering past observations and choosing which ones to look at again while reasoning.

Problem solvedMultimodal models struggle with long-horizon visual tasks because they can't retain or retrieve past observations effectively. VISTA solves this by maintaining a lossless visual memory, enabling models to solve complex interactive visual puzzles efficiently—achieving perfect scores on ARC-AGI-3 benchmarks.

ELI5A coding agent learns when to clean up and summarize its notes as it works through long programming tasks, deciding what matters to keep and what can be discarded. This keeps the agent focused and prevents running out of memory while solving complex software engineering problems.

Problem solvedLong-horizon coding agents hit context limits and lose track of useful information as they explore repositories. They need to actively manage working memory—deciding when to summarize, what to preserve, and how to recover—rather than just reacting when context overflows.

ELI5Instead of treating each lookup in a reference document as a fresh search, this method lets an AI agent develop a persistent "understanding" of that source that improves over time—like an expert who gets better at using a manual the more they reference it.

Problem solvedLLM agents repeatedly query the same sources but don't learn from those interactions. This wastes opportunities to build better internal models of how information is structured in each source, forcing them to re-search the same content and miss patterns they should have already learned.

ELI5An AI system learns to control irrigation by combining language model reasoning with physics simulation—it proposes actions, checks them against soil-water math, learns from mistakes without breaking the rules, and uses 51% less water than historical schedules.

Problem solvedLLM agents work well for one-off tasks but struggle with long-running physical control where mistakes compound and you can't reset the environment. This builds a system that safely learns irrigation decisions over entire growing seasons without human intervention.

ELI5Instead of summarizing documents when they arrive, this system keeps everything intact and figures out what's relevant only when you ask a question. It's like keeping an unedited archive of all decisions rather than maintaining a single "current" version that loses history.

Problem solvedLLM agents working in organizations need to know which decisions were active at specific times, but existing memory systems compress or overwrite old versions at write time, making it impossible to answer temporal questions accurately. This fixes that by preserving document history and selecting context only at query time.

ELI5Small AI models (2-9B parameters) can run on your laptop, but they fail at real tasks because typical agent software wastes their limited memory on tool instructions and gets stuck in loops. This paper builds better scaffolding that fits small models' constraints—like a zero-overhead instruction format and loop detection—so they actually finish tasks.

Problem solvedSmall open models are practical for local use but fail silently or loop endlessly when paired with standard agent frameworks designed for bigger models. Developers can't reliably use them for real work without custom fixes for each failure mode.

ELI5A system that helps AI models solve hard scientific problems by learning which web searches are actually useful—it evolves both the solutions and the search queries together, teaching the model when to look things up versus when to reuse what it already found.

Problem solvedLLMs get stuck on scientific discovery tasks because they don't have the right knowledge and waste time re-fetching the same web pages. This method lets models intelligently decide when to search, what to search for, and which retrieved documents actually help solve the problem.

ELI5A system that learns to customize how you prompt an AI agent based on what task it's solving. Instead of using one-size-fits-all instructions, it studies past optimization runs and creates specific tweaks for each individual problem.

Problem solvedAI agents currently use the same prompt/harness for all tasks, even though different problems need different guidance. This wastes the learnings from optimizing for different scenarios—Turbo Harness recycles that optimization work to boost performance on new instances without restarting from scratch.

ELI5A benchmark that tests whether AI agents can walk around virtual 3D worlds and spot bugs like floating objects or walls you can walk through—combining navigation skills with visual understanding to find and verify problems.

Problem solvedAs game engines and simulations get more complex, manually checking for environment errors is tedious and error-prone. This benchmark measures whether AI agents can automatically audit 3D worlds by exploring them intelligently and identifying anomalies, something current multimodal models struggle with.

ELI5A system that coordinates multiple AI agents to work on hard math problems like a research team would—each agent tries different proof approaches, others check their work, and good ideas get saved to build on. It found novel results on real open problems.

Problem solvedSingle AI shots can't crack research-level math problems that need exploring multiple ideas, recovering from dead ends, and remembering progress over weeks of work. This orchestrates agents to collaborate, verify, and compound progress like human mathematicians do.

ELI5Researchers tested whether complex multi-agent systems for automating machine learning tasks are actually necessary, or if a simpler agent with direct code execution works just as well. They found that simpler setups perform equally—the quality of the underlying model matters most.

Problem solvedTeams building autonomous ML systems are investing heavily in complicated orchestration layers and specialized sub-agents, but this work shows those additions don't actually improve results. Organizations can save engineering effort by using simpler, more direct approaches instead.

ELI5When a language model takes one wrong action in a multi-step task, everything after breaks. This method trains models to both avoid those critical mistakes and to recover from them when they happen, using teacher feedback to show the student how to get back on track.

Problem solvedMulti-turn agents fail because one wrong early move cascades into failure for the entire task. Existing training methods don't specifically address recovering from these pivotal mistakes, wasting training signal on unrecoverable dead-ends.

ELI5A standardized testing framework that measures how fast AI agents can complete computer tasks using GUIs—like clicking buttons and filling forms—while comparing their speed, cost, and accuracy across different models.

Problem solvedAI agents that use computers are getting better but remain slow and expensive to run. Without consistent, reproducible benchmarks, it's hard to know which approaches actually improve speed versus just changing test conditions, blocking real-world deployment.

ELI5Instead of an agent just trying things randomly, this gives it a meta-controller that sits above and decides what to do next—like a chess player thinking about the game plan rather than just moving pieces. The controller keeps notes on what worked, evaluates options, and picks the best next move given time budget.

Problem solvedLong-horizon agent tasks fail because agents waste compute replaying history or making bad next-step choices. This adds a smart controller layer that tracks progress compactly and makes deliberate decisions about which paths to pursue, dramatically improving success on complex multi-step problems.

ELI5Instead of making an AI smarter, this paper teaches another AI how to build better tools and environments for the first AI to work in—like learning what kind of workspace, notes, and resources help someone do their job better, even with the same underlying skills.

Problem solvedAI systems perform better or worse depending on their environment and support—but we can't always retrain them. This work lets you learn, at test time, how to construct smarter execution environments and scaffolding for fixed models, making AI systems more capable without changing their weights.

ELI5A small AI helper learns to give better advice to a frozen large language model by selectively learning from corrections that actually matter—those where the advice made a real difference. It ignores 'corrections' that wouldn't have changed anything.

Problem solvedTraining advisors for frozen LLMs is inefficient because models waste time learning from feedback that doesn't meaningfully improve advice quality. Selective learning from impactful corrections makes training faster and results better.

ELI5LLM agents often say they'll follow a plan but then deviate from it mid-execution. This paper shows that if you force the agent to stick to specific execution patterns (like sequential steps, hierarchical breakdowns, or search), task success jumps dramatically—but agents still struggle to pick the right pattern for each task.

Problem solvedLLM agents fail at long-horizon tasks not just because they plan poorly, but because they don't actually execute the plans they declare. You can't tell if a failure is bad planning or bad execution. This fixes the execution half by enforcing plan adherence, surfacing that plan selection remains the bottleneck.