When the Agent Stumbles: Turning Failure into Recovery Rules with Agent-Reflex
AI agents fall apart when a tool goes dark or a user cuts in mid-task. A team at Renmin University turned those failures into per-environment recovery rules and baked them into a small model beforehand. It lifted ordinary tool work, but handling noise remains unfinished homework.
Shuffle the Options, the Probabilities Hold — Jeva, a Decision-Only Neural Network
Jeva is a decision-only neural network that keeps each option's probability intact no matter how you reorder the options. It reuses candidates yet re-weighs them when the set changes, and when no option fits, it flags "none" separately. Research from NUS.
Why AI That Reads Your History Still Can't Call Your Next Move: Inside Vela
Vela is a framework that sharpens an LLM's grasp of individual users by turning their own behavior logs into "predict what they did next" problems. We unpack how it works, what it delivers, and where the headline numbers get muddied by distillation.
Will Your Coding Agent Build the Agent You Actually Asked For? — The AgentSWE Benchmark
We're entering an era where you hand over requirements and a coding agent builds the agent for you. AgentSWE is the first benchmark to measure whether that output actually *works* rather than just *looks right* — and we dig into what it found.
Tell It "Just the Answer" and It Still Shows Its Work — How to Measure an LLM's "Thinking Inertia"
Even when told "answer only," LLMs tack on their reasoning anyway. An analysis of arXiv 2610.11765, which dissects this behavior using three metrics that measure only the explanation you can actually see.
How Do You Teach an AI to Work for Hours? The Pitfalls of Long Task Trajectories and TerminalHorizon
Getting a terminal AI to work for hours requires long task trajectories. But a long record doesn't mean there's a lot to learn from it. TerminalHorizon builds and verifies only the tasks where earlier results actually change the next decision.
Letting the Model Decide Where to Look Closely — VisionWeave
Alibaba's Qwen team built VisionWeave, which lets a model decide for itself which parts of an image to inspect in detail and which to skim — cutting visual tokens by 43% on average while retaining 98.9% of the original model's performance.
Synthesizing an Entire Training Ground for AI Agents — A Look at SoftwareGym
A study that automatically synthesized 182 pieces of software to teach AI how to operate a computer. It pushed success on a real-app benchmark (OSWorld) from 41.39% to 56.97%.
The AI That Knows Where Things Are in an Image — So Why Does It Stumble When You Change the Viewpoint?
A multimodal AI gets viewpoint questions wrong, yet the correct location information was inside it all along. This study separates the failure: it isn't about "not seeing," but about "the computation that rotates what it saw to match a new viewpoint."
Fixing the Generation Process from Score Gaps Between Finished Outputs — DiffusionART
From the score differences across several outputs, it estimates a single correction direction that every generation step shares. We read DiffusionART, which learns on modest GPU budgets with no per-task re-tuning.
Five at Once — Why an AI PM Sets the 'Order of Inspection' Before Doing Anything
When requests pile up, the moment you queue them by appearance you lose the most expensive one. The order only becomes visible once you open each one — here's how an AI PM separates inspecting from doing to set priority, and which decisions a human has to keep hold of to the end.
Code Only, No Model at Runtime: A Dissection of the COAP Robot Policy
At runtime, no AI model is involved—pre-written code measures the robot's state and decides what to do. It reports a 70.24% success rate across RoboDojo's 42 bimanual tasks.