Your LLM Agent's Prompt Cache Is Silently Broken — A 571-Project Study Found the Bill You switched on prompt caching, watched your LLM bill drop, and moved on. The dashboard still shows cache hits. The answers still look right. So why does the invoice keep creeping up? A new study of real agent codebases has an uncomfortable answer: the most common caching bugs don't break anything visible. They just quietly stop saving you money — and in some cases, they lie to you about how much you saved. What the researchers did The paper is "Silent and Costly: Characterizing and Detecting Prompt-Caching Mis-Integration in LLM Agents" by Ying Zhang, Ruoqi Guo, Yi Liu, Gelei Deng, Yuekang Li, Yepang Liu, and Leo Zhang, submitted to arXiv on 26 October 2026 and accepted at ICSE 2027. It is the first empirical study of how prompt caching actually goes wrong in downstream applications — not in theory, but in shipped code. The team examined 120 prompt-caching-related pull requests from...

You ship an LLM feature, load-test it, and the dashboard says you can handle 100 requests a second. Then real users arrive — bursty, uneven, with long prompts and longer answers — and the experience falls apart at half that load. The first token takes forever, and nobody can explain why, because the number you planned around was never the right number. A new paper argues the problem is not your model or your GPUs: a single throughput number cannot describe a serving system at all — and it offers a simulator that predicts the full latency picture instead. What the researchers did The paper is “GAUSS: A Workload-Aware Analytical and Monte-Carlo Simulator for LLM Inference Serving” by Srikant Bharadwaj, Ankur Mallick, Spyridon Mastorakis, Anjaly Parayil, Dhruv Deshmukh, and Victor Ruehle, submitted to arXiv on 27 October 2026. Inference is now an operational workload fleets must provision, monitor, and bill against latency service-level objectives (SLOs) — and one throughput number cannot...

Your Deep Research Agent Is Slow Because It Writes One Word at a Time — Diffusion Just Made It 6x Faster Ask a deep research agent a hard question and go make coffee. The answer will probably still be coming. We tend to blame the model — too small, too dumb, too cautious. A new paper says the real culprit is much more mechanical: your research agent writes one word at a time, and every stage of its work inherits that speed limit. Swap out that one mechanism, the researchers show, and the same agent scaffold runs roughly six times faster — with no loss in report quality. What the researchers did The paper is "DiffResearch: Accelerating Deep Research with Diffusion Language Models" by Alan-Barsag Gazzaev, Alexey Gavrilov, Mikhail Mozikov, and Sergey Muravyov, submitted to arXiv on 20 October 2026. It tackles a bottleneck anyone who has used a research agent will recognise: deep research is iterative — plan, gather evidence, synthesise, revise, repeat. When each step is slo...

The bottleneck nobody is benchmarking Watch a computer-use AI agent do a simple office job and you will see the absurdity at once. It types a spreadsheet cell by cell when a single line of code would finish the whole sheet. It prints all two thousand lines of a file through a shell just to find one function. The model is not stupid. The tools it was handed are. That is the starting point of a new paper, "Evolving Agent Interfaces," by Saaket Agashe, Yuzhe Yang, Gonzalo Gonzalez-Pumariega, Bhuvan Chandra Koduru, Chih-Lun Li, Jiachen Yang, Ang Li, and Xin Eric Wang, submitted on 20 October 2026. Their argument is blunt: both the interfaces an agent uses and the knowledge of when to use each one are currently built by hand, and hand-building cannot keep pace with new applications, workflows, and models. So they made interface design itself an agent capability — and the results beat far more expensive frontier models. What the researchers did The team built Agent S4 , an auto...

One Photo Is Now Enough for a Photorealistic 3D Head — EGGHead Spends Its Gaussians Where They Matter Take one ordinary photo and ask a computer to turn it into a head you can rotate, light, and drop into a virtual scene. Until recently, that came with a painful trade-off: a fast, lightweight model that looked like a mannequin, a beautiful model that took far too long to build and render, or something in between that still needed a long, expensive "enrollment" step before it was usable. A new paper, EGGHead: Efficient Generalizable Gaussian Heads from a Single Image , attacks the trade-off directly — not with a bigger model, but by spending its rendering budget more cleverly. If you only have a fixed number of 3D Gaussians, do not spread them evenly. Put them where a face actually needs detail. What the researchers did EGGHead comes from Pilseo Park, Andrew Hou, Michel Sarkis, Ning Bi, Yiying Tong, and Xiaoming Liu. The paper, submitted to arXiv on 16 October 2026 as arXi...

Your Fancy Coding Agent Isn't Smarter — It's Just More Expensive: A 447-Task Study Proves It You picked the expensive coding agent because the demo looked smarter. The polished file search, the slick progress panels, the automatic history compression — surely all that scaffolding buys better code. Two researchers at Shandong University just ran the experiment the vendors never run: hold the model fixed, swap only the harness, and measure what actually changes. The answer is uncomfortable for anyone paying per token. The score barely moves. The bill triples. What the researchers did Yangze Liu and Zhongyi Han define the harness as everything around the model: the system prompt, the tool set, and the context management that turns a chat model into something that can work inside a repository. They ran five models through three production harnesses — Claude Code, mini-SWE-agent, and OpenCode — on SWE-bench Verified, the standard benchmark of real GitHub issue fixes. Then they di...

Your AI Agents Can't See Each Other's Rewards — This Trick Made Them Cooperate Anyway Imagine a team of delivery robots sharing one warehouse. Each robot can see what the others are doing — who carried which box, who waited at the door — but none of them can see the others' scorecards. Robot A doesn't know whether Robot B just got rewarded or penalised. Can a team like that ever learn to cooperate, or is it doomed to selfish chaos? For years, the standard answer in multi-agent reinforcement learning (MARL) has been: give every agent access to everyone else's rewards, and shape those rewards so that selfish behaviour hurts. It works in the lab. It collapses in the real world, because private reward signals are usually exactly that — private. A new paper from Mohamed Ayman Mohamed, Harshil Kotamreddy, and Marcos Menon Jose asks a beautifully human question: what if agents did what people do, and judged other agents' outcomes through their own eyes? What the res...

Let the Model Manage Its Own Memory: Context Language Models Beat Every Hand-Built Strategy Every long-running AI agent eventually drowns in its own context — forgotten plans, re-read files, summaries that ate the one detail that mattered. The standard fix has been better harness engineering: smarter compaction, fixed summarization rules, hand-built memory tools. A new paper from the University of Washington and Meta Superintelligence Labs argues the fix was backwards. Stop building a better harness. Hand the model the keys to its own context — and watch it invent strategies no human wrote down. What the researchers did Rulin Shao and colleagues (University of Washington, Meta Superintelligence Labs, MIT, Trillium Labs) introduce Context Language Models — CLMs . The idea is disarmingly simple. A normal language model can only append to its context: every new token gets concatenated onto what came before. A CLM treats its context as a file it can freely edit . Stale plan? D...

Your AI Agent's Tests Are Lying to You — IBM Studied 2,572 of Them Your AI agent's test suite is green. Every check passes. You ship it — and in production it refunds the wrong order, leaks a tool's internal error to a customer, or quietly hallucinates success while doing nothing at all. This isn't a hypothetical. Researchers from IBM Research, Georgia Tech, and the University of Georgia just published Tangent at ASE '26 (this week, in Munich): the first large-scale empirical study of how developers actually test LLM-based agent applications. They mined 1,190 Python repositories built on LangChain, LangGraph, and MCP, identified 1,723 agents and 4,111 tools , and manually labeled 2,572 test methods across 240 test modules. Their conclusion is uncomfortable: the way the industry tests agents is narrow, shallow, and blind to exactly the failure modes that bite in production. What the researchers did Rangeet Pan and colleagues built a dataset the hard way....

CLIP's zero-shot magic is real — until you deploy it. The moment your test data drifts from the pretraining distribution — different lighting, a new camera, unfamiliar backgrounds — accuracy starts to bleed. The standard fix is retraining on target data, but that needs labels you don't have and a training loop you can't afford at the edge. Test-time adaptation (TTA) promises the middle path: the model quietly learns from the unlabeled data stream as it predicts . Cache-based TTA, the current hot approach, stores past predictions as prototypes and retrieves them to sharpen future ones. It works — but researchers from Tsinghua Shenzhen, Pengcheng Laboratory, and Sun Yat-sen University just showed it has two silent killers. Their fix, RADC (Risk-Aware Dual Caching) , attacks both, and it now holds the state of the art on cross-dataset generalization. What the researchers did First, the diagnosis. Siyu Huang, Yueyong Chen, Xuejiao Li, and Jun Zhou identified two flaw...

AI Just Co-Authored a Published Research Paper. Sakana's Agentic Scientist Runs the Whole Research Loop — and Humans Only Steer. On October 13, 2026, Sakana AI — the Tokyo lab co-founded by David Ha — released an 88-page preprint with a claim that would have sounded like science fiction two years ago: an orchestration layer built on top of modern coding agents conducted end-to-end computational research and co-wrote a paper that was accepted as an oral presentation at ALIFE 2026 . Not a toy benchmark. A real paper, accepted after external peer review. The system is called The Agentic Scientist , it's fully open source under Apache 2.0, and — refreshingly — the authors are blunt about where it still breaks. Here's what they built, what it actually achieved, and the workflow patterns you can steal for your own automation today. What the researchers did Modern coding agents can write code, launch experiments, and analyze results with minimal human input. But the Sakana t...

Every AI agent you have ever used is a language model plus a set of tools. The model does the thinking; the tools do the doing — calculators, search APIs, code runners, database queries. Strip away the tools and the smartest model becomes a very eloquent chatbot that can't touch anything. Here's the problem nobody talks about: your agent can only use tools a human bothered to write. If the perfect tool for your task doesn't exist as an API, the agent is stuck improvising with whatever is on the shelf. Researchers have tried prompting models to write new tools at inference time — but that approach has a fatal split personality: the part that writes the tool is decoupled from the part that uses it, with no signal telling the writer whether its tool is one it could actually call. A team at Appier Research and National Taiwan University (paper submitted August 25, 2026, since accepted at NeurIPS) just closed that loop. Their framework, SMITH — Schema-grounded Multi-task I...