Tootfinder

xRouter: Training Cost-Aware LLMs Orchestration System via Reinforcement Learning
Cheng Qian, Zuxin Liu, Shirley Kokane, Akshara Prabhakar, Jielin Qiu, Haolin Chen, Zhiwei Liu, Heng Ji, Weiran Yao, Shelby Heinecke, Silvio Savarese, Caiming Xiong, Huan Wang
https://arxiv.org/abs/2510.08439

xRouter: Training Cost-Aware LLMs Orchestration System via Reinforcement Learning
Modern LLM deployments confront a widening cost-performance spectrum: premium models deliver strong reasoning but are expensive, while lightweight models are economical yet brittle on complex tasks. Static escalation rules and keyword heuristics under-utilize this spectrum and fail to adapt across task types. We present xRouter, a tool-calling-based routing system in which a learned router can either answer directly or invoke one or more external models. The router is trained end-to-end with re…

Extending the low-$Z$ "border'' of the $A=100$ region of deformation with precision mass spectrometry of $^{96-98}$Kr
D. Lunney, V. Manea, M. Mougeot, L. Nies, N. A. Althubiti, D. Atanasov, K. Blaum, A. Herlert, W. -J. Huang, J. Karthein, I. Kulikov, Yu. A. Litvinov, D. Neidherr, T. R. Rodr\'iguez, M. Rosenbusch, L. Schweikhard, T. Steinsberger, A. Welker, F. Wienholtz, R. N. Wolf

Extending the low-$Z$ "border'' of the $A=100$ region of deformation with precision mass spectrometry of $^{96-98}$Kr
The onset of collective nuclear behavior in the ${N=60}$, ${A\sim100}$, region is examined through high-precision mass measurements of $^{96-98}$Kr, performed with the ISOLTRAP mass spectrometer at ISOLDE, CERN. Our results for $^{96-97}$Kr agree with previous measurements, with our new $^{97}$Kr Penning-trap mass value three times more precise. The mass value of $^{98}$Kr is measured for the first time. The new mass surface, together with comparisons to beyond-mean-field theoretical prediction…

UniVoice: Unifying Autoregressive ASR and Flow-Matching based TTS with Large Language Models
Wenhao Guan, Zhikang Niu, Ziyue Jiang, Kaidi Wang, Peijie Chen, Qingyang Hong, Lin Li, Xie Chen
https://arxiv.org/abs/2510.04593

UniVoice: Unifying Autoregressive ASR and Flow-Matching based TTS with Large Language Models
Large language models (LLMs) have demonstrated promising performance in both automatic speech recognition (ASR) and text-to-speech (TTS) systems, gradually becoming the mainstream approach. However, most current approaches address these tasks separately rather than through a unified framework. This work aims to integrate these two tasks into one unified model. Although discrete speech tokenization enables joint modeling, its inherent information loss limits performance in both recognition and g…

The co-founder of landlord app Airbnb is now Chief Design Officer of the USA. And the UX commentariat cheers, because the vibes of a "seat at the table" play better than the reality: the same government that appointed Joe Gebbia has just demolished 18F, an actual design studio that did actual UX work for the USA, rather than meaningless posturing.
This is not the first time UX has made this bargain: have a seat at the table, only if you can pledge to do nothing with it.

The aesthetics of success come at the cost of failure
It's easy to get a seat at the table. Just agree to become part of the problem.

About 40 to 50 Pentagon reporters turned in access badges and exited the Pentagon rather than agree to new US government-imposed restrictions on their work (David Bauder/Associated Press)
https://apnews.com/article/pentagon-press-a…

Journalists exit Pentagon rather than agree to new reporting rules
Journalists at the Pentagon turned in access badges and cleaned out their workspaces, the price for refusing to agree to new restrictions on their jobs at the seat of U.S. military power. The refusal was near-unanimous, from trade publications, wire services, television networks and newspapers, and included outlets that appeal largely to conservatives like Fox News Channel and Newsmax. Many of the reporters chose to turn in their badges together at the 4 p.m. deadline set by the Defense Departm…

Dynamic Speculative Agent Planning
Yilin Guan, Wenyue Hua, Qingfeng Lan, Sun Fei, Dujian Ding, Devang Acharya, Chi Wang, William Yang Wang
https://arxiv.org/abs/2509.01920 https…

Dynamic Speculative Agent Planning
Despite their remarkable success in complex tasks propelling widespread adoption, large language-model-based agents still face critical deployment challenges due to prohibitive latency and inference costs. While recent work has explored various methods to accelerate inference, existing approaches suffer from significant limitations: they either fail to preserve performance fidelity, require extensive offline training of router modules, or incur excessive operational costs. Moreover, they provid…

The fracturing of the Dutch far-right, after Wilder's reminded everyone that bigots are bad at compromise, is definitely a relief. Dutch folks I've talked to definitely see D66 as progressive, <strike>so there's no question this is a hard turn to the left (even if it's not a total flip to the far-left)</strike> a lot of folks don't agree. I'm going to let the comments speak rather than editorialize myself..
While this is a useful example of how a democracy can be far more resilient to fascism than the US, that is, perhaps, not the most interesting thing about Dutch politics. The most interesting thing is something Dutch folks take for granted and never think of as such: there are two "governments."
The election was for the Tweede Kamer. This is a house of representatives. The Dutch use proportional representation, so people can (more or less) vote for the parties they actually want. Parties <strike>rarely</strike> never actually get a ruling majority, so they have to form coalition governments. This forces compromise, which is something Wilders was extremely bad at. He was actually responsible for collapsing the coalition his party put together, which triggered this election... and a massive loss of seats for his party.
Dutch folks do still vote strategically, since a larger party has an easier time building the governing coalition and the PM tends to come from the largest party. This will likely be D66, which is really good for the EU. D66 has a pretty radical plan to solve the housing crisis, and it will be really interesting to see if they can pull it off. But that's not the government I want to talk about right now.
In the Netherlands, failure to control water can destroy entire towns. A good chunk of the country is below sea level. Both floods and land reclamation have been critical parts of Dutch history. So in the 1200's or so, the Dutch realized that some things are too important to mix with normal politics.
You see, if there's an incompetent government that isn't able to actually *do* anything (see Dick Schoof and the PVV/VVD/NSC/BBB coalition) you don't want your dikes to collapse and poulders to flood. So the Dutch created a parallel "government" that exists only to manage water: waterschap or heemraadschap (roughly "Water Board" in English). These are regional bureaucracies that exist only to manage water. They exist completely outside the thing we usually talk about as a "government" but they have some of the same properties as a government. They can, for example, levy taxes. The central government contributes funds to them, but lacks authority over them. Water boards are democratically elected and can operate more-or-less independent of the central government.
Controlling water is a common problem, so water boards were created to fulfill the role of commons management. Meanwhile, so many other things in politics run into the very same "Tragedy of the Commons" problems. The right wing solution to commons management is to let corporations ruin everything. The left-state solution is to move everything into the government so it can be undermined and destroyed by the right. The Dutch solution to this specific problem has been to move commons management out of the domain of the central government into something else.
And when I say "government" here, I'm speaking more to the liberal definition of the term than to an anarchist definition. A democratically controlled authority that facilitates resource management lacks the capacity for coercive violence that anarchists define as "government." (Though I assume they might leverage police or something if folks refuse to pay their taxes, but I can't imagine anyone choosing not to.)
As the US federal government destroys the social fabric of the US, as Trump guts programs critical to people's survival, it might be worth thinking about this model. These authorities weren't created by any central authority, they evolved from the people. Nothing stops Americans from building similar institutions that are both democratic and outside of the authority of a government that could choose to defund and abolish them... nothing but the realization that yes, you actually can.
#USPol #NLPol

These days, anti-GenAI screeds cross my radar every week. Even when I agree with what they’re saying (I often do), they’re mostly saying the same things. So, FWIW, here’s one with a fresh PoV: https://going-medieval.com/2025/10/21/on-ai-and-the-golem/
With an entertain…

On AI and the golem
Recently I, like a lot of people who make their living by creating things, have been thinking a lot about so-called AI. I have been forced to do this because so many companies have been stealing th…

Drax: Speech Recognition with Discrete Flow Matching
Aviv Navon, Aviv Shamsian, Neta Glazer, Yael Segal-Feldman, Gill Hetz, Joseph Keshet, Ethan Fetaya
https://arxiv.org/abs/2510.04162

Drax: Speech Recognition with Discrete Flow Matching
Diffusion and flow-based non-autoregressive (NAR) models have shown strong promise in large language modeling, however, their potential for automatic speech recognition (ASR) remains largely unexplored. We propose Drax, a discrete flow matching framework for ASR that enables efficient parallel decoding. To better align training with inference, we construct an audio-conditioned probability path that guides the model through trajectories resembling likely intermediate inference errors, rather tha…

Series B, Episode 02 - Shadow
BLAKE: Don't tell me you agree with Gan.
AVON: No, no, no, I'm just...interested. It seems to me that we have identified the source of that threat rather easily.
BLAKE: So why haven't they?
AVON: Exactly.
https://blake.torpidity.net/m/202/405