Blogroll

At a glance

• AI agents often fail because their instructions, or skills, are manually modified with no guarantee of improvement. SkillOpt turns skill editing into a training process, making agent behavior more reliable without changing model weights.
• SkillOpt treats an agent skill file as a trainable parameter outside a frozen target model, turning skill writing from one-shot prompting into a controlled optimization process.
• Across six benchmarks, seven target models, and three execution modes, SkillOpt is the best or tied-best method in all 52 evaluation cells, improving performance without updating model weights.
• SkillOpt keeps skills compact and auditable through bounded text edits, validation gating, rejected-edit feedback, and slow/meta updates, avoiding uncontrolled prompt drift.
• The optimized skills transfer across model scales, agent harnesses, and related tasks, suggesting that they capture reusable workflow knowledge rather than benchmark-specific instructions.

Large language models (LLMs) are increasingly deployed as agents that gather evidence, call tools, and execute multi-step tasks. For these agents, the hard problem is no longer whether they can call a tool, but whether they can complete tasks reliably and consistently. Today, agent skills typically come from three sources: experts write them by hand, a frontier model generates them one-shot, or the agent loosely revises them after execution. None of these approaches behaves like a deep-learning optimizer. They lack step-size control, held-out validation, and any memory of revisions that failed. As a result, skills tend to grow longer and drift with each rewrite, and a revision that seems perfectly reasonable can quietly degrade real task performance. This uncontrolled skill evolution has become a major obstacle on the path from agent prototype to dependable, production-grade deployment.

In our recent paper, SkillOpt: Executive Strategy for Self-Evolving Agent Skills, we reframe the question from “how do we write a better prompt?” to “how do we train the skill?” SkillOpt treats the skill file as a trainable parameter living outside a frozen target model, bringing a training-style optimization loop, consistent gains across 52 evaluation cells, and a compact skill file that stays readable, auditable, and transferable.

Figure 1. A frozen target model executes tasks while a separate optimizer model trains the skill layer from trajectory feedback, exporting the reusable skill file best_ skill.md through validation gating. How SkillOpt works Video 1. SkillOpt’s optimization loop, from trajectory collection to the exported skill file.

SkillOpt organizes skill editing as a forward–backward–update cycle in text space. In the forward pass, the frozen target model executes a batch of training tasks with the current skill; the rollout batch size controls how much evidence each update receives. In the backward pass, a separate optimizer model reads the resulting trajectories in reflection minibatches, distilling patterns to preserve from successful trajectories and patterns to correct from failures.

In the update step, the optimizer proposes small add, delete, and replace edits; candidate edits are merged, deduplicated, ranked, and clipped by a textual learning rate—a per-step edit budget. Every candidate skill must then pass a strict validation gate: it is adopted only if it scores strictly higher than the current skill on the held-out validation split. Rejected edits are not discarded; they enter a rejected-edit buffer that serves as negative feedback for later optimizer calls in the same epoch. On a slower cadence, an epoch-wise slow/meta update consolidates longer-horizon lessons that single batches cannot reveal (Figure 2). Together, bounded edits, validation gating, and best-version selection keep skill optimization controllable and auditable, so the skill converges instead of drifting.

Figure 2. The SkillOpt pipeline: trajectory collection, minibatch reflection, bounded text updates, validation gating, and epoch-wise slow/meta updates jointly constrain skill training. Consistent gains across benchmarks, models, and execution modes

We evaluated SkillOpt across six benchmarks (SearchQA, SpreadsheetBench, OfficeQA, DocVQA, LiveMathematicianBench, and ALFWorld), seven target models from frontier-scale GPT-5.5 to the small open-weight Qwen3.5-4B, and three execution modes (direct chat, Codex, and Claude Code). Counting each combination as one evaluation cell, When measured against human-written skills, one-shot LLM skills, Trace2Skill, TextGrad, GEPA, and EvoSkill, SkillOpt delivered the best or tied for -best results on all 52 cells. These performance improvements are unusually large for a method that updates no model weights. With GPT-5.5 in direct chat, SkillOpt raises the six-benchmark average from 58.8 to 82.3, a +23.5-point absolute improvement—and +5.4 points above an oracle that picks the single best competing method per cell. The largest gains appear on procedural benchmarks: SpreadsheetBench rises from 41.8 to 80.7, OfficeQA from 33.1 to 72.1, and LiveMathematicianBench from 37.6 to 66.9. The same interface carries over to agentic loops, lifting GPT-5.5 by +24.8 points inside Codex and +19.1 inside Claude Code over no skill.

Spotlight: Event Series

Microsoft Research Forum

Join us for a continuous exchange of ideas about research in the era of general AI. Watch the latest episodes on demand.

Watch on-demand Opens in a new tab A small model plus a skill file

Approaching the next model tier SkillOpt also narrows the gap between small or open-weight models and frontier models—without changing any weights or adding any extra model calls at inference. After optimization, GPT-5.4-mini’s six-benchmark average (64.3) exceeds the no-skill baseline of the larger GPT-5.4 (59.7), and GPT-5.4-nano (57.4) exceeds the no-skill baseline of GPT-5.2 (51.3). Qwen3.5-4B, a 4-billion-parameter open-weight model, surpasses GPT-5.2’s no-skill baseline as well. Gains that once required a larger model can now be approximated by one optimized skill file.

Skills that transfer: train once, reuse everywhere

The optimized skill file captures reusable task-solving procedures rather than instructions overfit to a single model, benchmark, or execution environment. This is why the same skill can still improve performance when transferred across model scales, agent harnesses, and related tasks. In our transfer experiments, skills continued to deliver gains when moved across model scales, across execution harnesses, and to a nearby math benchmark. The clearest example is cross-harness transfer: a spreadsheet skill trained inside Codex, dropped into Claude Code with no further optimization, lifts the no-skill baseline from 22.1 to 81.8 (+59.7)—slightly above the 80.4 achieved by training directly inside Claude Code. Because the two harnesses expose different tool surfaces, this suggests SkillOpt learns general workflow logic, not just harness-specific recipes.

Compact, readable, and built from very few accepted edits

The deployed artifact, best_ skill.md , is neither an opaque parameter blob nor an ever-growing log. Across six case studies, the median final skill length is roughly 920 tokens, and because the validation gate rejects most proposals, only one to four edits are accepted into the final file. OfficeQA’s +39.0-point gain comes from a single accepted edit. The learned rules read like a seasoned practitioner’s advice. Component ablations confirm that the controls do the work: removing the rejected-edit buffer lowers scores on all three ablation benchmarks, and removing both the meta skill and the slow update drops SpreadsheetBench from 77.5 to 55.0. A new adaptation layer for the agent era SkillOpt points to a lighter-weight path for domain-adapting agents: instead of fine-tuning weights, hard-coding task logic, or hand-tuning prompts, teams can train a small, versionable, auditable natural-language skill layer—wherever automatic evaluation or a reliable verifier exists.

By bringing learning rates, schedules, validation splits, rejected samples, and slow updates to agent skills, SkillOpt suggests that training need not be limited to model weights. Procedural knowledge outside the model can also be optimized.

When that process is controlled, validated, and recorded, a natural-language skill becomes a stable, transferable, and reversible adapter between frontier-model capability and real-world workloads. Read the full paper, visit the project page at aka.ms/skillopt (opens in new tab), or explore the SkillOpt GitHub repository at github.com/microsoft/SkillOpt (opens in new tab). Teams building agentic workflows can use SkillOpt as a foundation for training reusable skills against their own tasks and verifiers. See also our companion project, SkillLens.

Paper GitHub SkillLens Project Page Opens in a new tab

The post SkillOpt: Agent skills as trainable parameters appeared first on Microsoft Research.

You've heard of the popular smart home automations: automatic lights, security systems, robot vacuums.. But there are plenty of niche gadgets out there that can streamline your setup and create the ultimate smart home dream—think Tony Stark's house.

In celebration of the 4th of July and while summer is in full swing, now is the perfect time to hit Home Depot and grab some new power tools for DIY projects and home repairs. While select DeWALT and Ryobi tools are up to 40% off and battery deals are available, there are a few even better deals you won't want to miss.

Over 16 years ago, Google released the Nexus One, and it's considered one of the most important phones in Android history. I still own mine. And while a lot has changed since it arrived in January 2010, and the latest Pixel 10 Pro XL is better than ever, Google could still learn a thing or two from those older models.

When most people look at an Ethernet cable, their first thought would be that this is how you transport data. However, you've probably seen devices like security cameras or network switches that aren't only sending data over Ethernet, but also receiving power using that one cable.

SAVE $133: As of June 30, get the Samsung 1TB 9100 PRO SSD for $206.99 at Amazon. That's down from its usual price of $339.99, saving you over $100.

Opens in a new window Credit: Amazon Samsung 1TB 9100 PRO SSD $206.99 at Amazon
$339.99 Save $133   Get Deal

Storage is expensive. You can add that to a veritable treasure trove of other pricey PC parts and components that cost too much, especially RAM.

And it's only getting pricier, so the best time to get a hard drive, especially if you plan on playing some of the year's biggest video games, is now. Though Prime Day may be over, we still found a great deal on an SSD you can use to make sure you're ready for any game that comes your way, as well as photos, videos, or whatever you like to store on your computer.

As of June 30, get the Samsung 1TB 9100 PRO SSD for $206.99 at Amazon, down from its usual price of $339.99. That's $133 off and a discount of 39%.

SEE ALSO: Prime Day is over: We found 10+ deals still live on microSD, portable SSDs, and hard drives to beat RAMageddon

This internal SSD is a whopping 1TB, so it offers plenty of storage space for gaming, photos, videos, files, or backups. It boasts read speeds of up to 14,700MB/s as well as random read/write speeds of up to 1,850K/2,600K IOPS, which means it's zippy enough to recall files and help load content that'll serve you better than an external hard drive, if that's what you're in the market for.

The only downside here is, if you aren't as savvy with computers as you wish you were, you do have to install this SSD. So if you're uncomfortable with popping your case open and putting it in yourself, you may want to opt for an external hard drive instead. If that doesn't bother you though, this is an excellent offer for an SSD that will serve you well, especially if you're planning on future-proofing your setup.

Despite what some people might think, computers run on electricity and not magic fairy dust. Electricity costs money, and generating it usually has some sort of environmental cost too.

Range has basically become the new way people measure efficiency, mostly because it feels more real-world than MPG. MPG tells you how far a car could go on a gallon, but range shows how far you actually get before stopping for fuel, tank size included.

Milwaukee makes hundreds of tools you'll love, but if you're a new homeowner or simply on a budget, certain affordable tools are a must-buy. If you already own a drill or a few other tools, here are four more Milwaukee tools under $120 worth adding to your collection.

For decades, luxury buyers who prioritized reliability had a simple answer: buy a Lexus. The Japanese brand built its reputation on offering premium vehicles with the durability and low ownership costs that many European rivals struggled to match, making it the default choice for cautious shoppers.

If you’re thinking about switching to Linux, you’re probably comparing distros like Ubuntu, Fedora, or Mint and wondering which one is the best fit. But, with my decade-long experience using and recommending Linux, I think that’s the wrong question to ask. Instead, start by comparing GNOME and KDE Plasma, decide which one fits your workflow better, and then choose any distro built around that desktop environment (DE). That narrows the decision down to just two options.

For years, I clicked through the same apps and routines every morning by hand. Then I found a tool buried in Windows that's been there the whole time, quietly waiting to take over.

Samsung is preparing a true onslaught of foldable phones this July, and now we've gotten our first detailed look at all of them.

Android Headlines posted the images on Tuesday, and even though these are technically images of cases that will be offered with these phones at launch, some of them do show the phones inside the case.

Perhaps the most interesting in this bunch is Samsung's rumored wide foldable phone. According to the outlet, this phone will actually be called the Samsung Galaxy Z Fold 8. The successor of the Galaxy Z Fold 7 will be called the Galaxy Z Fold 8 Ultra, and it will be positioned as the most advanced foldable phone Samsung has to offer. Finally, the Galaxy Z Flip 8 is the compact, flip-style foldable that follows the Galaxy Z Flip 7.

In the images, the Galaxy Z Fold 8 has a dual camera system on the back, and a punch-hole style selfie camera on the front. The Samsung Galaxy Z Flip 8 also has a dual camera on the back, while the Galaxy Z Fold 8 Ultra comes with a triple rear camera.

The cases on offer include a dark grey one with a kick-stand and aramid fiber-style pattern (this one comes in several color combos), several solid color cases including a white and a purple option, and a couple of joyful ones with cartoon character prints. The options are similar for all phones, though the aramid fiber cases are missing for the Flip phone.

SEE ALSO: Samsung to take on the iPhone Pro with the new Galaxy S27 Pro?

As for what these phones will actually be like, check out our overview of rumors and reports for both the ultra-wide Samsung Galaxy Z Fold 8, and the Samsung Galaxy Z Fold 8 Ultra.

Plex has been around since 2008, forking from a macOS XBMC port, which is, of course, today's Kodi. I haven't been on the Plex train since day one, but it has been over a decade at this point, and it was my first go at self-hosting streaming media in my home.

Everyone loves a good deal, especially when the car you’re getting feels a lot nicer and drives better than the price would suggest. With affordable cars getting harder to come by, it’s refreshing when a brand actually puts in the effort to stretch value further.

SAVE OVER $100: As of June 30, the Lenovo Slim 7i Aura Edition laptop (Core Ultra 7 Processor 256V, 16GB Memory, 1TB Storage) has returned to its lowest-ever price of $999.99 at Amazon.

Opens in a new window Credit: Lenovo Lenovo Slim 7i Aura Edition $999.99 at Amazon
$1,155.64 Save $155.65   Get Deal

If buying a new laptop has been at the top of your mind — whether for work, school, or fun — Amazon still has some solid discounts available after Prime Day. The Lenovo Slim 7i Aura Edition laptop (Core Ultra 7 Processor 256V, 16GB Memory, 1TB Storage) is one of the best, down to its lowest-ever price at Amazon.

As of June 30, this Lenovo Slim 7i Aura Edition laptop is on sale for $999.99 at Amazon. That's just over $150 off its typical price of $1,155.64, but it won't be at this price for long. A timer has already begun on the store page that's counting down through June 30, so now is the time to take advantage of this deal before it's gone for good.

SEE ALSO: Laptop specs explained: A jargon-free guide to what's inside your computer

If you enjoy the portable feel of a laptop, the Lenovo Slim 7i Aura Edition is the perfect fit for those on-the-go. It starts at just 2.82 pounds and features a thin design, which makes it nice and easy to slip into a bag or backpack to take with you on the commute. This particular model comes with 16GB of memory, a Core Ultra 7 processor 256V, and 1TB of storage, alongside up to 17 hours of battery life.

That's certainly not all, though. The Lenovo Slim 7i Aura Edition also has a beautiful screen, whether you're watching something or scrolling through a project for work. It boasts DisplayHDR True Black 500, 600 nits peak brightness, and WUXGA resolution at 60Hz so you can experience a smooth, sharp, and colorful picture.

This deal on the Lenovo Slim 7i Aura Edition laptop (Core Ultra 7 Processor 256V, 16GB Memory, 1TB Storage) runs out after June 30. Act fast to save at Amazon.

Want to learn more about getting the best out of your tech? Sign up for Mashable's Top Stories and Deals newsletters today.

TL;DR: The LG 27GX790B-B 27-inch UltraGear OLED gaming monitor is on sale for $699.99 at Amazon, down from its $999.99 list price.

Opens in a new window Credit: LG LG 27-Inch Ultragear QHD OLED Gaming Monitor $699.99 at Amazon
$999.99 Save $300 Get Deal

OLED gaming monitors are rarely where you go looking for budget upgrades, but Amazon’s latest LG UltraGear deal gives you a pretty serious price cut on one of the fastest 27-inch screens in LG’s current lineup. 

As of June 30, the LG’s 27-inch UltraGear QHD OLED gaming monitor (the 27GX790B-B model) is on sale for only $699.99 at Amazon — with a limited-time deal cutting $300 from its $999.99 list price. 

Shipped and sold by Amazon directly (instead of a third-party seller), free delivery is set for July 5 at the time of writing, or as soon as July 2 for Prime members. Amazon also shows a used “Like New” option for $664.99 through Amazon Resale, if you want to save as much as possible. Anyone wanting the new model directly from Amazon will want to stick with the $699.99 listing. 

This UltraGear is built around a 27-inch QHD OLED panel with a 2560x1440 resolution, giving you a sharper picture than 1080p without asking your PC to push full 4K. If you’re either new to PC gaming or upgrading from a much older rig, this is a very nice middle ground. 

If you’re the type who plays a lot of competitive games, you’ll be a big fan of LG’s Dual Mode: letting you switch between QHD at up to 540Hz or HD at up to 720Hz, depending on whether you care more about sharpness or pure reaction time. Add in the 0.02ms gray-to-gray response time, and games like Counter-Strike 2, Rocket League, Fortnite, and Valorant should feel seriously quick.

The OLED panel should also help with more cinematic games and streaming, with LG listing 335 nits of typical brightness, DisplayHDR True Black 500, UL Verified Perfect Black, and a 1.5M:1 contrast ratio.

That means if you’ll be playing Assassin’s Creed Black Flag Resynced or Grand Theft Auto VI in the future, you’ll be getting richer blacks, stronger contrast, and better detail in darker scenes than you’d expect from a more standard gaming monitor. 

For keeping frames extra smooth, the monitor also supports NVIDIA G-SYNC compatibility and AMD FreeSync Premium Pro. DisplayPort 2.1, dual HDMI 2.1 ports, USB-C, three USB ports, DTS Headphone:X support, and height, tilt, swivel, and pivot adjustments round out the package — so you have all the versatility you need to complete your new immersive setup and then some. 

If you fancy an extra QHD monitor, the 27-inch Odyssey G5 is now only $149.99 at Amazon. As for OLEDs, you can still grab the Samsung Odyssey OLED G9 curved gaming monitor for its lowest-ever price. 

TL;DR: The Magic: The Gathering Marvel Super Heroes Play Booster Box is on sale for $125.95 at Amazon, down from its $209.70 list price.

Opens in a new window Credit: Magic: The Gathering Magic: The Gathering Marvel Super Heroes Play Booster Box $125.95 at Amazon
$209.70 Save $83.75 Get Deal

Magic: The Gathering and Marvel fans have already seen this Play Booster Box bounce through a few decent discounts, but now that the expansion is officially out, Amazon has pushed even harder with a new record low price for this sealed 30-pack display box. 

As of June 30, the Magic: The Gathering Marvel Super Heroes Play Booster Box has been cut in price at Amazon to just $125.95. Compared to its $209.70 list price, that saves you $83.75 and takes 40% off the box — its lowest-ever price confirmed by price tracker camelcamelcamel. Shipped by Amazon directly, you can get standard free delivery for July 5 (at the time of writing), or as soon as July 1 for Prime members. 

That price brings the box down to about $4.20 per Play Booster before tax, which is a pretty big drop from the launch day and preorder prices we were seeing just a few weeks ago. 

Walmart has the same box listed for $125.95 through marketplace seller Cataclysm Games and fulfilled by Walmart itself, with some other seller listings starting slightly lower. TCGplayer is even cheaper on paper, with listings as low as $119.99 plus $4.99 shipping. That’s a total of $124.98, only $1.24 above the trading card selling platform’s estimated market price of $123.75. 

Mashable Deals Be the first to know! Get editor selected deals texted right to your phone! Get editor selected deals texted right to your phone! Loading... Sign Me Up By signing up, you agree to receive recurring automated SMS marketing messages from Mashable Deals at the number provided. Msg and data rates may apply. Up to 2 messages/day. Reply STOP to opt out, HELP for help. Consent is not a condition of purchase. See our Privacy Policy and Terms of Use. Thanks for signing up!

If you’re only just learning about this new set of cards, the Marvel Super Heroes expansion brings Marvel’s roster into Magic: The Gathering, with heroes and villains jumping from the comics to the battlefield. Players can build around their favourite characters, equip them with recognisable gear, and lean into MTG’s usual strategy-heavy battles across casual games, drafts, and Limited play.

Inside the box, you’ll get 30 Play Boosters, with 14 Magic cards in each pack. Every pack includes one Traditional Foil card and one to four cards of rare or higher rarity, plus a mix of uncommons, commons, and a land card. There’s also a small chance of finding a Foil Borderless Mythic Rare, while Traditional Foil Land replaces the regular land in 20% of boosters.

Preorders for Magic: The Gathering's Reality Fracture Play Booster Box are on sale for $10 off at Amazon. What’s more, the Magic: The Gathering Foundations Jumpstart Booster Box is on offer for under $100. 

TL;DR: Slurpee Day is back this summer, landing on Saturday, July 11. Customers can grab a free small Slurpee drink from participating 7-Eleven, Speedway, and Stripes stores nationwide.

Summer is quite literally heating up, but 7-Eleven is here to help you stay cool with the return of its mega-popular Slurpee Day. And better yet, the iconic drink is turning 60 this year, so 7-Eleven has a couple more surprises in store for customers.

Landing on Saturday, July 11 (that's 7/11, by the way), Slurpee Day will kick off at participating 7-Eleven, Speedway, and Stripes stores nationwide, with each handing out free small Slurpee drinks to celebrate. If you want to get in on the fun earlier, you can actually score a free large Slurpee drink on July 10 if you make a minimum $20 purchase through the 7NOW delivery app.

There's also a brand new flavor to enjoy this year: Mountain Dew Confetti Chill. If you're curious what that tastes like, 7-Eleven says it "blends bright citrus flavor with notes of sweet vanilla icing, giving fans a zesty lemon cake treat in frozen drink form." It's only available until Aug. 29, so summertime will be your big opportunity to taste this chilly new release. The fun certainly doesn't stop there, though.

Select 7-Eleven and Speedway stores will also be hosting a Slurpee Date experience on Slurpee Day. Whether you're planning to share a Slurpee with a partner or grab one with a friend, this experience will include a photo-op bench, commemorative keepsakes, and limited-edition, bedazzled Slurpee “S” double straws that are "inspired by the brand's diamond anniversary," per the press release.

Romantic, right? Credit: 7-Eleven

What better way to spend a summer day? Don't miss out on all the fun of Slurpee Day on July 11.

Want to learn more about getting the best out of your tech? Sign up for Mashable's Top Stories and Deals newsletters today.

I’m always trawling websites like MakerWorld and Printables for models to make and talk about. Checking the reviews before downloading can help avoid disappointment and wasted filament, but too many reviewers seem to blame model authors for problems that have nothing to do with the model itself.