Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Clear out junk files and repair common Windows errors3Fix the driver behind crashes, sound loss and screen glitchesSome links on this page are affiliate links: if you buy through them we may earn a commission, at no extra cost to you.
A research team at Sungkyunkwan University reports that its Self-improving Agent with Gated critique (SAG) framework raised success on the ALFWorld virtual-task benchmark from 24.6% to 78.4% while keeping token use comparable to the ReAct baseline. The authors calculate a 3.1× improvement in normalized token efficiency. A university account summarizes the resource use as approximately one-third of conventional approaches, but that wording should not be read as a measured reduction in electricity, hardware cost or total operating expense.
SAG achieves the trade-off by asking an external language model for critique only when the agent’s action appears uncertain, then learning from those critic-assisted decisions so it can rely on the critic less often. The same research release discusses a second method, VAN-Flow, but that is a separate offline-reinforcement-learning contribution rather than the source of the paired compute-and-success headline.
As an Amazon Associate I earn from qualifying purchases.
Free tools Windows power users keep installed
One-click scans. No signup required.
How does SAG reduce computing while improving task success?
SAG treats a critique request as a value-of-information decision. Instead of sending every proposed action to an external critic, it estimates whether critique is likely to change the outcome. The gate uses two action-level ambiguity signals over the admissible actions: global entropy and the margin between the two most likely actions. High uncertainty or a small top-two gap indicates that additional advice may be useful; a confident choice can proceed without another model call.
The method is training-free at the gating stage, so it can be placed around an existing long-horizon language-model agent. It also performs online bootstrapped self-improvement: after receiving useful critique, the agent incorporates that behavior into its own subsequent decisions. In principle, this lets the number of expensive consultations fall as the agent becomes better at recognizing and resolving situations itself.
Popular on Amazon
- 【IP68 Waterproof Multifunctinal Smart Watch】Torjalph fitness activity tracker smart watch is IP68 waterproof technology ,you can wear it when swimming,handwashing,it's rain-proof which means that you perform multiple exercise modes outside without worrying about sweat or rain corroding your fitness smart watch,and this watch equipped with a 230 mAh large capacity battery magnetic charging ,and this sports activity tracker watch can be used 5-7 days If fully charged.
- 【Larger 1.83'' Full Screen Watch & Multi-dial Selection】1.83 inch full HD color touch screen smart watch for men / women provides excellent 240*284 resolution image quality ,This Bluetooth smart watch offers a better touch experience and remarkable immersion and ,more than 200 watch faces for you to choose from App"Da Fit ", you can change the dial style according to your mood to wear keeping the world on your wrist
- 【Accurate Activity Tracker Smart Watch 】sports watch activity tracker uses sweat and dust resistant breathable silicone strap, our Fitness Tracker Smart Watch for Android iOS Phones supports 12 sports modes,walking,running,bicycle,swimming,badminton,basketball,football,rope skipping. This bluetooth calls smart watch for men women is lightweight and stylish, bringing more convenient and relaxing sports experience while also providing more protection.
- 【Your 24 Hour Personal Daily Health Monitor】Fitness watch for women men full touch screen IP67 waterproof bluetooth smartwatch for Android iOS Phones with heart rate blood pressure monitor multifunction Andriod smart watch with steps counter sleep monitor sports fitness watch, monitoring blood oxygen saturation allows you to understand your physical condition in real time to track fitness progress ,to know your health status in time.
- 【More Functions Bluetooth Call Smart Watch】This fitness activity smart watch full screen for iPhone and Android phones includes a bunch of new features including:The bluetooth smart watch can hang up the phone, receive and display WhatsAPP /Text /Facebook /Twitter /Messenger reminder.Support sedentary reminder, alarm clock, stopwatch, timer,weather information,music control, remote camera control compatible with Android 5.0 and above or iOS 9.0 and above smartphones.
What did the authors report on ALFWorld?
In the SAG paper, “Selective Critique for Cost-Aware LLM Agents in Long-Horizon Decision Making,” Heewon Park, Somin Im and Minhae Kwon report ALFWorld task success increasing from 24.6% to 78.4%. The abstract says SAG used a token budget comparable to ReAct and achieved a 3.1× improvement in normalized token efficiency. The evaluation covered three long-horizon interactive benchmarks and multiple backbone models, according to the paper.
The university’s TechXplore report gives the endpoint as 78.3% and says computational-resource consumption fell to approximately one-third of conventional approaches. The 0.1-point discrepancy is a reporting difference between the news account and the paper abstract. More importantly, the paper’s quantified cost language concerns token budget and normalized token efficiency. Neither source establishes a universal two-thirds reduction in electricity, dollar cost or data-center compute for every deployment.
ALFWorld is a simulated household-task environment. Its result is therefore a benchmark outcome, not a 78.4% success guarantee for general-purpose agents, robots or production workflows. The reported numbers come from the authors’ experiments and have not been independently reproduced by this publication.
What is the critique gate actually measuring?
Global entropy captures how spread out the model’s probability distribution is across legal actions. A diffuse distribution means the agent is unsure. The local top-2 margin measures the probability gap between the best and second-best admissible actions; a narrow gap signals a close decision. SAG combines these signals to decide when an external critique is worth its token cost.
Also popular on Amazon
- [Low Noise High Fidelity] New bee guitar cord with 22AWG oxygen-free-copper center conductors provides enhanced signal transfer, the PVC insulation material and the high-density stranded copper double shield provide noise rejection for quiet operation
- [Ideal for Choice] New bee 1/4 straight and right angle metal connectors, ensures a secure connection in any instrument, pedal, or amp jack. L shape great fro on stage guitar stands
- [Wide Compatible] Suitable for mixer, power amplifier, microphone, effector, decoder, equalizer, electoronic organ, electric guitar, electric bass and other electronic equipments. Widely used in the installation of smart home, karaoke speaker, professional stage audio, etc
- [Easy Disassemble] Making incidental repairs quick and easy! Very nice, flexible cable
- [Dont Hesite to Order] 24 hours customer services and professional technology team are standing by
This design targets a common failure mode in long-horizon interaction: an early mistake can make many later steps irrelevant. Selective intervention spends additional reasoning where a wrong branch is most likely or most damaging, rather than paying the critique cost at every step.
How is VAN-Flow different from SAG?
VAN-Flow addresses a different problem and should not be presented as an alternative measurement of SAG’s efficiency result. In “Variance-Averse n-Step Offline Reinforcement Learning for Sparse Long-Horizon Environments,” Guhyeon Kang and Minhae Kwon combine three components:
- a categorical, distributional critic that represents a range of possible returns;
- a variance-averse expectation operator that favors actions with high expected return and lower dispersion; and
- a flow-matching generative actor guided by rejection sampling.
The authors report evaluations on more than 40 D4RL and OGBench tasks, with the largest gains in long-horizon, high-variance settings. The abstract does not provide one aggregate success percentage comparable to SAG’s ALFWorld figure. VAN-Flow is offline reinforcement learning, whereas SAG is a selective-critique mechanism for language-model agents making interactive decisions.
| Method | Primary setting | Core idea | Reported evidence |
|---|---|---|---|
| SAG | Long-horizon interactive LLM agents | Request critique only when ambiguity suggests high value, then self-improve | ALFWorld success 24.6% to 78.4%; 3.1× normalized token efficiency |
| VAN-Flow | Offline reinforcement learning | Prefer high-return, lower-variance actions with a distributional critic and generative actor | More than 40 D4RL and OGBench tasks |
What do these results mean for real-world AI systems?
The findings suggest a practical strategy for controlling inference overhead: make extra reasoning conditional on uncertainty instead of treating it as a mandatory step. That could matter in agents with long action sequences, where repeated critique calls accumulate rapidly. However, the evidence currently establishes benchmark research results, not validated performance in deployed robotics, autonomous driving or other safety-critical systems.
Also popular on Amazon
- Digital Family Calendar: BSIMB digital calendar wall touch screen puts an end to your chaotic life. Designed for busy families, the digital calendar keeps you organised and in step with your family. Assign different colours and tasks to each family member to make family scheduling easier and more intuitive.
- Easy Setup and Auto-Sync: Plug in, connect to WLAN and auto-syn your electric calendar in minutes. Digital wall calendar compatible with Google calendar, iCloud calendar, Outlook calendar, Cozi calendar and Yahoo calendar. Add events directly on the device or through the ecalendar app. Electronic calendar stays current whether you're at work or traveling—no manual updates needed.
- All-In-One Home Assistant: Smart calendar builds habits and encourage independence with personalised chores. Display daily meal plan to reduce the daily ‘What's for dinner?’ stress. To-do lists keep track of everything you want. Turning chores into valuable achievements with stars and rewards, making life easier and more enjoyable.
- Manage Anytime Anywhere: BSIMB electronic calendar for family stay connected with the powerful ecalendar app. Access and manage your digital wall calendar anytime, anywhere with free features like stars & rewards, photo screensaver, chore charts, meals plans, to-do lists etc. Electronic planner paid features like Magic Import tailor advanced tools for users with special needs.
- Stay in Touch with Loved Ones: Share photos or videos from your phone to electronic wall calendar with app, instantly update your life moment, and keep in touch with your family, especially those who are far away; you can also surprise your smart calendar digital wall by sending him a wish card; it's also great for keeping tech-illiterate elders in touch with the youngsters.
The institutional report quotes Minhae Kwon saying, “It is particularly meaningful that our student researchers took the lead in developing practical foundational technologies that can be readily applied to real-world and industrial environments.” That is Kwon’s characterization of the work, not evidence that commercial or safety-critical deployments have been demonstrated. The reviewed records also do not establish code availability, commercial licensing or a production-ready package.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.When were the studies published?
Both papers were submitted to arXiv in October 2026: SAG on October 5 and VAN-Flow on October 6. The university report, published by TechXplore on October 26, identifies both studies as accepted to NeurIPS 2026. Readers should check the linked records for later revisions, since arXiv manuscripts can change after their initial submission.
PC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchSources: SAG paper (arXiv:2610.07335); VAN-Flow paper (arXiv:2610.07899); Sungkyunkwan University report via TechXplore.
Frequently Asked Questions
Does SAG reduce electricity use by exactly two-thirds?
Not according to the quantified paper result. The authors report token-budget and normalized-token-efficiency changes, while the university news account describes resource consumption as approximately one-third of conventional approaches; neither is a direct electricity or cloud-bill measurement.
Also popular on Amazon
Is the 78.4% figure a general success rate for AI agents?
No. It is the SAG authors’ result on ALFWorld, a simulated long-horizon household-task benchmark. Performance can differ substantially with another environment, model backbone, prompt, action space or failure definition.
Can I download SAG or VAN-Flow for production use?
The cited arXiv abstracts and institutional report do not establish a public code release, commercial package or deployment support. They document research methods and benchmark evaluations rather than a production distribution.
Why are entropy and the top-two margin useful together?
Entropy reflects uncertainty across the full set of admissible actions, while the top-two margin focuses on how close the leading decision is to its nearest competitor. Combining broad uncertainty with local decisiveness gives the gate more information than either signal alone.
