DEV Community: Fox

IBM's new Token Maturation paper

Fox — Sun, 04 Oct 2026 17:43:04 +0000

IBM's new "Token Maturation" paper is a very intriguing look into the world of token synthesis rather than token sampling in language models.

Token maturation takes a pipelined autoregressive generation approach via continuous vector refinement. Put crudely, it's like a K-token sliding-window autoregressive sort of "diffusion".

Its basic idea is to maintain a K-token vector buffer where tokens get to develop into vectors that are good argmax choices once they exit the buffer.

It's pipelined, in the sense that K tokens enter an "assembly line" (called the liquid tail) where they get to evolve cascadingly in latent space, before getting snapped to the model vocabulary one-by-one as new tokens enter the liquid tail.

It's still autoregressive due to the fact that tokens still enter one-by-one and exit the liquid tail one-by-one. Pipelining does not affect this nature.

It achieves this cascading evolution by representing the current K tokens in the liquid tail as continuous vectors where each vector goes through K passes of prediction until it reaches the head of the liquid tail and gets discretised into the vocabulary.

These passes happen latently in the entire continuous vector space where the discrete vocabulary embedding vector group is defined. So, rather than having to develop in the liquid tail into certain tokens in vocabulary, tokens in the liquid tail get to develop near relevant token areas before getting discretised.

The Token Maturation process did not fall into entropy collapse. The lack of entropy collapse means the model isn't just simulating standard discrete sampling inside a continuous latent space—where it would have effectively chosen the token at step one. Without premature (pre discretisation) collapse, the model actively steers tokens into the correct semantic region with each pass, making every iteration meaningful rather than just a nudge toward a pre-determined pick. Because liquid tail tokens aren't forced to collapse near vocabulary early on, the model avoids the error propagation that comes from locking into a suboptimal choice. Furthermore, maintaining higher entropy prevents the model from settling into bland, generic output.

This new approach also exposes a new inference modifier, the guidance scale s, it acts as a steering weight, where higher values steer the liquid tail tokens closer to the prompt but restricts trajectory. The liquid tail length K acts as the model's refinement horizon (or convergence budget), controlling how many passes a token gets to resolve ambiguity and integrate context before committing.

While using standard LLM samplers could be possible by swapping the snapping step from an argmax choice to a softmax sampling, this would subvert the point of using the liquid tail, where token stochasticity is gained from the maturation itself rather than random sampling in standard language models. It is also not guaranteed to function well with maturation due to the possibility of unexpected jumps at discretisation.

Source:

IBM Paper

How to fix inconsistent Play-Pause behaviour with Bluetooth Headsets on Linux

Fox — Sun, 15 Feb 2026 20:25:23 +0000

The Problem

Some Bluetooth headsets (or from my limited experience, all of them) handle Play-Pause behaviour internally instead of relying on the OS to tell whether media is playing or not. This can easily lead to a de-sync between the host and the device, where the device will either end up constantly trying to Play while media is playing, or Pause while media is paused.

This can be verified by running xev (sorry Wayland users, get gud :p) with and without media playing, then seeing if eventually the wrong signals are sent (XF86AudioPlay when the device is supposed to pause or XF86AudioPause when it's supposed to play).

It is worth mentioning that this issue does not occur on Android, I wasn't able to manually verify the reason myself as it was too complex for me to look into, but investigating with Google Gemini reveals that Android has a way to communicate with Bluetooth devices to sync playback state.

The Solution

The solution is to redirect both XF86AudioPause & XF86AudioPlay to XF86AudioPlayPause, bypassing the whole internal state issue, this is most easily done through your Desktop Environment's shortcut settings.

Setup in KDE Plasma:

Go to:

System Settings
Keyboard
Shortcuts
Media Controller
Disable shortcuts for "Pause media playback" & "Play media playback"
Set the following two keys for the "Play/Pause media playback" shortcut (you might need to manually play or pause media, or even reconnect the device, to get the right key to fire when you press on your device):
- "Media Play" (This one is the default shortcut, so you can just enable it using the check mark)
- "Media Pause"

Some programs might not respond to the shortcuts correctly until they're relaunched.

How to get virtual display outputs on Linux without using a dummy plug

Fox — Fri, 09 Jan 2026 21:53:22 +0000

This guide is mainly for Xorg, on Wayland there are easier way of achieving this, for example KDE's virtual monitor implementation in Krfb: krfb-virtualmonitor | tldr InBrowser.App.

A useful way to troubleshoot while working on this is to use OBS Studio to take a peek at the virtual output by adding a new source screen.

Method #1: Sacrificing one of your existing outputs

Use Xrandr to check for free outputs on your system.

Follow How to Create a Virtual Display for Sunshine on Arch Linux - Anton Ždanov and stop before the Sunshine section:

Regarding EDID files, you can add modes later using xrandr, so you don't need to choose one that matches your desired specifications.

Step 1: Get an EDID File

An EDID (Extended Display Identification Data) file simulates a monitor. You can download pre-made EDID files from the v4l-utils repository. Choose one that matches your desired resolution and features (e.g., 4K, HDR).

Step 2: Place the EDID File

Create a directory for EDID files:
sudo mkdir -p /usr/lib/firmware/edid
Copy your chosen EDID file to this directory. For example:
 sudo cp samsung-q800t-hdmi2.1 /usr/lib/firmware/edid/
Step 3: Configure Kernel Parameters

Identify your GPU's free HDMI or DP output:
for p in /sys/class/drm/*/status; do con=${p%/status}; echo -n "${con#*/card?-}: "; cat $p; done
Replace HDMI-A-1 in the next step with the appropriate output.

Add the following kernel parameters (see arch wiki or endeavour os discovery):
drm.edid_firmware=HDMI-A-1:edid/samsung-q800t-hdmi2.1 video=HDMI-A-1:e
Update your mkinitcpio.conf file to include the EDID file:
FILES=(/usr/lib/firmware/edid/samsung-q800t-hdmi2.1)
Regenerate the initramfs:
sudo mkinitcpio -P
Reboot your system.

You might afterwards need to enable the display in your Display Settings before it starts working.

Method #2: Abusing the EVDI Linux driver before adding the output in kernel parameters

From DisplayLink/evdi: Extensible Virtual Display Interface on Github:

The Extensible Virtual Display Interface (EVDI) is a Linux® kernel module that enables management of multiple screens, allowing user-space programs to take control over what happens with the image. It is essentially a virtual display you can add, remove and receive screen updates for, in an application that uses the libevdi library.

This driver is usually used for DisplayLink devices, which are basically external USB graphics cards used to obtain more display outputs from a system than it physically has, it even works for Macbooks. The cost being the software requirement and additional CPU overhead.

In Using the EVDI driver - Extreme Multihead - ArchWiki, it is suggested that this driver can be (ab)used to create virtual outputs without a DisplayLink dock. You will lose out on GPU acceleration that way, but the result is still usable for general tasks. The ArchWiki links to nvidia - Unable to add a VIRTUAL display to Xorg - Unix & Linux Stack Exchange, but the Etam Software - Virtual screen guide is clearer and lets you know that the current Xorg session will freeze during initial configuration. Although I've had luck deploying this solution for a long time on my machine, an update to EVDI suddenly now results in an unusable black screen, even tho this was supposedly patched in Monitor blank but detected as connected · Issue #523 · DisplayLink/evdi, I was never able to get it working again just using EVDI.

Either way, the pure EVDI solution is less elegant the solution presented here, as it does not play nicely with Desktop Environment screen management (eg. having to disable the Screen Management background service KScreen2 in KDE Plasma using the Background Services GUI Program, rendering the Display Configuration page in KDE System Settings unusable.)

From Etam Software - Virtual screen:

For Arch Linux the EVDI module is available in the AUR
You can get more than one extra output this way by changing the number for initial_device_count

Install evdi kernel module
git clone https://github.com/DisplayLink/evdi.git
cd evdi/module

make && sudo make install
# or
sudo make install_dkms
Add evdi to /etc/modules-load.d/local.conf.

Add options evdi initial_device_count=1 to /etc/modprobe.d/local.conf.

WARNING! This next step will freeze your Xorg, so you might want to reboot instead or be clever about killing your Xorg or something.
sudo modprobe evdi

After rebooting there should be new DVI outputs present in Xrandr

You will probably see that your new outputs don't have any modes listed in Xrandr, you should add those manually.

Now that you have extra outputs, you are able to use them to follow the steps in Method #1.

Using your newly obtained virtual outputs

It is straightforward now to use VNC or other screen sharing solutions to get your virtual outputs on your desired devices, remember that by default VNC traffic is unencrypted.

If you're planning on using an Android device, consider using adb reversing for better latency and easier connection without having to set up the target IP:

From Etam Software - Virtual screen:

4) (optional) Attach Android device through USB cable
adb reverse tcp:5900 tcp:5900
Now you can point your VNC client application like MultiVNC to localhost:5900.

DEV Community: Fox

IBM's new Token Maturation paper

How to fix inconsistent Play-Pause behaviour with Bluetooth Headsets on Linux

The Problem

The Solution

Setup in KDE Plasma:

How to get virtual display outputs on Linux without using a dummy plug

Method #1: Sacrificing one of your existing outputs

Step 1: Get an EDID File

Step 2: Place the EDID File

Step 3: Configure Kernel Parameters

Method #2: Abusing the EVDI Linux driver before adding the output in kernel parameters

Install evdi kernel module

Add `evdi` to `/etc/modules-load.d/local.conf`.

Add `options evdi initial_device_count=1` to `/etc/modprobe.d/local.conf`.

Using your newly obtained virtual outputs

DEV Community: Fox

IBM's new Token Maturation paper

How to fix inconsistent Play-Pause behaviour with Bluetooth Headsets on Linux

The Problem

The Solution

Setup in KDE Plasma:

How to get virtual display outputs on Linux without using a dummy plug

Method #1: Sacrificing one of your existing outputs

Step 1: Get an EDID File

Step 2: Place the EDID File

Step 3: Configure Kernel Parameters

Method #2: Abusing the EVDI Linux driver before adding the output in kernel parameters

Install evdi kernel module

Add evdi to /etc/modules-load.d/local.conf.

Add options evdi initial_device_count=1 to /etc/modprobe.d/local.conf.

Using your newly obtained virtual outputs

Add `evdi` to `/etc/modules-load.d/local.conf`.

Add `options evdi initial_device_count=1` to `/etc/modprobe.d/local.conf`.