Research, distilled.
Research explained clearly, with lessons that connect the ideas.
AgentValve caps Sybil-driven payment exposure by trust capacity
In permissionless AI markets, cheap fake identities can inflate provider reputation, leaving clients unsure how much a manipulated payment decision could risk. AgentValve ties payment approval to finite trust capacity; tests found exposure stayed at the honest-to-adversarial boundary capacity even as 10,000 such identities were added in selected graphs.
Latest added:
Latest lessonRoboticsWhen a Robot Struggles, What Should Its Training Change?
A robot may close a laptop yet drop an object it has grasped; another must feel whether an object is soft, or complete a task when the tabletop changes. Improving these behaviors is not one problem with one remedy. The studies here change different parts of a robot system: training from human touch, rewarding distinct successful actions during practice, or turning human movement into commands a humanoid can use. Each reports a physical-robot outcome, but the tasks and comparisons differ. The useful question is which change addresses the difficulty a particular deployment presents.
Latest added:
Latest across topics
View all papersSeparating search from reading improves document question answering
Finding evidence in a document collection is not the same as understanding it. When one agent can both search and read, does it keep searching instead of opening the table, figure, or text that contains the answer? KnowForge separates them: a search-only Locator gathers evidence, a read-only Deep Reader examines it, and a bridge check seeks missing documents. Its scores rose over one-shot search-and-answer, while reading helped most when answers depended on figures or tables.
AI & AgentsWritten skills improve agent procedures without changing weights
Agent procedures often need domain-specific tool use and answer formats, but changing model weights may be unavailable or costly. Can the procedure itself be improved without rewriting the model? SkillOpt edits a portable text skill, tests each revision on separate validation examples, and keeps only validated changes. Across 52 measured combinations of model, task, and execution mode, it was best or tied-best against the strongest measured alternative in every combination.
Audio & SpeechOpen-vocabulary speech runs on a low-cost microcontroller
Cheap microcontrollers can recognize wake words, but open-vocabulary speech usually needs cloud processing or more memory than the chip has. Can a small device understand arbitrary English without sending audio away? Oído combines integer arithmetic with a memory-aware execution schedule on an ESP32-S3. Its engine keeps word error rate within 0.12 percentage points of full precision on LibriSpeech, but one-core processing still takes longer than the audio itself.
Trending this week
View all papersTop trending scores among papers added in the last seven days.
AgentValve caps Sybil-driven payment exposure by trust capacity
In permissionless AI markets, cheap fake identities can inflate provider reputation, leaving clients unsure how much a manipulated payment decision could risk. AgentValve ties payment approval to finite trust capacity; tests found exposure stayed at the honest-to-adversarial boundary capacity even as 10,000 such identities were added in selected graphs.
#1 · Trending 59.8Search & DataCurvature-guided compression avoids rotating model weights
At two bits per weight, trellis codecs compress model weights, but leading approaches rotate them first and must undo that transform during generation. Can curvature information guide the search directly, without rotation, while retaining accuracy on language tasks? CurveTQ is a rotation-free codec that weights its search by curvature. On three tested 4–8B Instruct models, it scored higher on average than two competing codecs after they received the same stored block-start information.
#2 · Trending 52.5NLP & LanguageA crisis-support test shows recognition is not enough
Most safety tests ask whether a language model refuses unsafe text. That misses a harder question in a suicide or self-harm crisis: can it recognize distress, calm the situation, and connect someone with human help? DistressBench is a clinician-authored English test that scores those actions separately across simulated conversations. Across 25 models, 35.3% of conversations recognized as crises still lacked a route to help, showing that low-harm replies can remain incomplete.
#3 · Trending 46.6Lessons
View all lessonsGuided paths that connect several papers into one idea you can keep.
When a Robot Struggles, What Should Its Training Change?
A robot may close a laptop yet drop an object it has grasped; another must feel whether an object is soft, or complete a task when the tabletop changes. Improving these behaviors is not one problem with one remedy. The studies here change different parts of a robot system: training from human touch, rewarding distinct successful actions during practice, or turning human movement into commands a humanoid can use. Each reports a physical-robot outcome, but the tasks and comparisons differ. The useful question is which change addresses the difficulty a particular deployment presents.
Machine LearningReaching the Goal, Responding in Time, and Staying Safe
A robot can complete a task and still respond too late to a moving object. It can also reach a destination after a collision. Saying that it “succeeded” leaves out what was measured and what happened while it decided. The selected studies offer a bounded answer: adapting computation was checked against task completion, rewarding varied successful behavior was checked under shifted simulation conditions, and a moving-world test separately counted safe arrivals. Each claim belongs to its own test.
Machine LearningWhat Else Changes When You Fix a Model?
A model can learn from a new corpus while losing old behavior, receive privacy-protecting noise while losing task accuracy, or refuse a harmful prompt while also refusing a harmless neighbor. Success at the intended target leaves a nearby question unanswered. The studies suggest a bounded way to investigate: locate what changes, choose a comparator, and inspect the closest relevant boundary. That boundary is different for training updates, privacy calibration, and prompt repair. Neither a plausible mechanism nor a reassuring broad score settles the other studies’ questions.