For more ways to supercharge your workflow, check out more articles in ourTech for Architects series, which includes our recommendations of Top Laptops for Architects and Designers.While Artificial Intelligence (AI) has entered the mainstream conversation, its application in architecture remains mostly unfamiliar to those outside the AEC professional or academic circles. Terms like parametric design and computer algorithms sound like technical jargon, creating a disconnect between these technological processes and the real-world constructions that people interact with. As a result, AIs potential to analyze data, optimize energy use and predict environmental challenges goes unnoticed by the broader public. People engage with architecture daily, yet often without awareness of how these design innovations impact their lives.Iconic buildings are easily recognizable and often admired for their eye-catching aesthetic or shape. However, the connection between these constructions and the AI-powered design processes behind them isnt always clear to the public. People may appreciate a buildings form, size, materials or distinctive features without realizing that many of these elements result from algorithmic design processes that respond to specific parameters. To the average observer, such buildings appear as products of architectural progress, while the role of advanced technologies remains hidden. The lack of visibility of AI tools role in the architectural design process creates a disconnect, leaving people unaware of how these tools are used to create adaptive, efficient and sustainable designs benefits they experience without realizing their origin. This gap between design intent and public perception underscores a broader challenge: bridging the gap between architectural innovation and public understanding, particularly regarding technologies like parametric design and AI tools.The Public Perception Gap in Architectural Innovation170 Amsterdam designed by Handel Architects, New York, New York | Photo by Bruce DamonteArchitects use AI-powered and parametric design tools to create innovative solutions that optimize functionality, efficiency and sustainability, yet these advancements often go unnoticed or are poorly understood by the public. This disconnect limits engagement and appreciation for the transformative role of modern architecture. People may admire or react positively or negatively to visually striking buildings but are often unaware that the parametric design they observe spans beyond aesthetics, responding to site constraints, structural performance, energy efficiency, user behavior and environmental parameters. For example, the 170 Amsterdam residential building in Manhattan, New York, features an exoskeleton that serves structural and shading purposes, addressing functional and environmental challenges. The average passerby, unfamiliar with the designs parametric principles, would interpret it as a bold architectural statement. They would likely recognize the dynamic character of the buildings exoskeleton and the contrast between the robust structure and the expansive glazed surfaces as the buildings most distinctive features. Yet they may be unaware that this design optimizes usable floor area and energy efficiency.Design That Speaks: How Design Features Connect with the PublicSun Shading System designed by Tilt Industrial Design for the University of Technology (UTS) Central building, Ultimo, Australia | Photo by Brett BoardmanWhile the complexities of AI technologies and parametric design often go unnoticed, the rising awareness of sustainability has made some architectural features more recognizable and relatable to the public, especially when they visibly and effectively showcase their environmental purpose. This connection is significant because AI tools and parametric design play an important role in achieving sustainable goals. They enable architects to analyze data, optimize energy efficiency and design thoughtful solutions that respond to environmental challenges. The growing environmental awareness among the general population has made many people more conscious of the efforts in the architectural industry to address sustainability. There might not be a complete understanding of the complex strategies that architects use to achieve sustainable goals. However, when visually striking, some features are more likely to be recognized and associated with sustainability and effectively communicate their purpose to the public. Such features include green roofs and living walls, solar panels, dynamic shading systems and materials such as bamboo or recycled wood. For example, the central building of the University of Technology Sydney (UTS) in Ultimo, Australia, features an automated sun-shading system that regulates solar heat gain and minimizes glare. In addition to playing a critical role in the buildings energy-saving design, the shading system also creates a striking visual impact. At the same time, it informs of its functional purpose as an integral part of the overall architectural design. Even if people are not fully versed in sustainable design principles, the dynamic shading system on the UTS building conveys its purpose through its visible functionality. The movement of the louvers in response to sunlight and their integration into the buildings faade make their role as a solar heat regulator intuitive. This blend of practicality and visual impact allows the shading device to communicate its environmental intent clearly, making sustainability tangible and relatable to the public.Enhancing Public Awareness to Shape User-Centered EnvironmentsThe visibility of sustainability through AI-powered architectural designs can be extended to other areas of the built environment. Improving public awareness of how AI and parametric design tools shape the built environment can help people recognize the positive impact these technologies have on their daily lives. This awareness, in turn, can promote a stronger societal interaction with architecture and greater support for functional, efficient and sustainable development. The benefits can be particularly evident in high-traffic public areas, like open public spaces, transportation hubs, schools, universities and medical facilities, where efficiency and functionality are paramount.AI is a valuable tool for analyzing human movement, usage patterns and parameters to optimize layouts for circulation and accessibility. This strategy ensures that design outcomes are user-centered while optimizing functionality and efficiency.A Public Education Imperative: How AI Shapes the Built EnvironmentThe Airport Typology Reconfigured. Study Area: Des Moines, Iowa. Created by Blake Gallagher at NBBJThe gap between AI-powered design processes in the architecture, engineering and construction (AEC) industry and the general public is significant. However, it also offers an opportunity for a greater and more productive connection. If the AEC community takes an educational approach to bridge this divide, the interaction between the public and architecture can be transformed. Educational initiatives whether through exhibitions, talks by experts, interactive websites or collaboration with schools and continuous cooperation between architects and the public structured around participatory workshops can demystify design choices, making architecture more relatable. These efforts can showcase how elements like form, materials and layout are thoughtfully considered to optimize functionality, efficiency and sustainability.As AI and parametric tools continue to revolutionize how the built environment is designed and experienced, educating the public about their impact will encourage a deeper connection to architecture. This understanding can lead to more inclusive, adaptive solutions aligned with peoples needs.For more ways to supercharge your workflow, check out more articles in ourTech for Architects series, which includes our recommendations of Top Laptops for Architects and Designers.The post From Algorithms to Architecture: Why Architects Are Turning to AI for Smarter, Greener Designs appeared first on Journal.

Sony has announced the next wave of titles coming to the PlayStation Plus Game Catalog. Starting January 21st, PS Plus Extra and Premium subscribers can play God of War Ragnark, Like a Dragon Gaiden: The Man Who Erased His Name, SD Gundam Battle Alliance, ANNO: Mutationem, Orcs Must Die! 3, and Citizen Sleeper on PS5 and PS5.PS5 players can look forward to Atlas Fallen: Reign Of Sand while PS4 players get Sayonara Wild Hearts and, oddly, enough, Poker Club. Of course, the latter is backwards compatible on PS5, further expanding the line-up for current-gen owners.Exclusive to Premium subscribers are two PlayStation Classics Indiana Jones and the Staff of Kings from the PS2 era and MediEvil 2 on the PS One with rewind, up-rendering, and quick save. Both are playable on PS4 and PS5, and the former is a great way to prepare for the arrival of Indiana Jones and the Great Circle this Spring for PS5.

A new crayfish species calledCherax pulverulentus has a bright blue color. Ji PatokaScientists have described a new species of colorful crayfishand it might already be sitting in your aquarium. The researchers detailed their discovery in the journal Zootaxa last week.Though crayfishalso called crawdads or crawfishmay not be everyones idea of the perfect animal companion, scientists in the Czech Republic found the newly named Cherax pulverulentus in a shipment of Indonesian pet crayfish theyd purchased for their lab in March 2023.The lobster-like crustacean is a burrowing freshwater crayfish endemic to the Indonesian part of New Guinea, according to the study. Specimens have large eyes and claws, as well as smooth bodies. The species comes in two color forms: Blue form crayfish are a dark hue with orange joints and tails, while purple form crayfish are turquoise with purple spots and white along their joints and tails. ACherax pulverulentuscrayfish raises its claw toward the camera. Ji PatokaLike the other members of this genus, this crayfish is aggressive, strictly freshwater (inhabiting streams and maybe also lake ecosystem), probably moderately burrowing, omnivorous and gonochorist, meaning it has distinct male and female individuals,Ji Patoka, a zoologist at the Czech University of Life Sciences Prague and a co-author of the study, tells Popular Sciences Laura Baisas. They grow more than four inches long and might live between five and ten years.Despite the fact that Cherax pulverulentus has been sold as an aquarium pet for more than two decades in places including Europe, Japan, Indonesia and the United States, it was never differentiated from similar species. It was referred to as Hoa Creek, Irian Jaya or Blue Moon crayfish, names that were also used for other types of crayfish, according to the study.The new formal name, Cherax pulverulentus, means dusty crayfish in Latin, in homage to the crustaceans dotted exoskeleton. The crayfish's dotted exoskeleton gave it its Latin namepulverulentus, which means "dusty." Ji PatokaThis recent find coincides with another crayfish discovery: Two other species were newly described in North Carolina this month. Like thedusty aquarium crayfish, they had been mistakenly lumped together with another species, as Frank Graff reports for PBS North Carolina.Taxonomy is essential for conservation, Bronwyn Williams, a research curator at the North Carolina Museum of Natural Sciences, tells PBS. If a unique creature does not have a formal name, it is not eligible for the resources needed to manage and protect it.The New Guinea crayfish discovery has also been tied to conservation. It highlights the need for better management and identification in the aquatic pet trade as well as recognition and protection for the creatures in their natural habitat, writes Micah Hanks for the Debrief.To confirm their finding, the researchers conducted morphological and genetic analyses of six specimens. Broadly, however, not much is known about the species, since it has been mostly observed in captivity. One exception is a blue form specimen discovered in a Hungarian thermal spring, which the researchers suspect was previously a pet, dumped there by its former owner.They now aim to study the newly identified species in its natural environment.Crayfish are just one group of awesome animals native to New Guinea, the island facing many environmental risks. In this regard, we believe that our findings will help to highlight the importance of this island from a biodiversity perspective, and more conservation activities will apply there, Patoka tells Popular Science. Last but not least, I would like to alert aquarium owners: Do not release your pets outdoors!Get the latest stories in your inbox every weekday.Filed Under: Animals, Biology, Cool Finds, DNA, Genetics, New Research, Pets, Weird Animals

https://youtu.be/2tt1bNPFk0UStray Kite Studios, an indie team composed of veteran developers, announced their first original game called Wartorn.Read More

Baldur's Gate 3 exceeds 100m mod downloads | News-in-briefLarian introduced mod support in June 2024, surpassed 50 million mod downloads last NovemberImage credit: Larian Studios News by Sophie McEvoy Staff Writer Published on Jan. 15, 2025 This is a News-in-brief article, our short format linking to an official source for more information. Read more about this story by following the link below:Baldur's Gate 3 exceeds 100m mod downloads

The case were about to discuss is none other than the beautiful and critically acclaimed The Star Named EOS. In this article about console porting, well discuss specifically the solutions we found to translate the save system. Without further ado, lets jump right in..In The Star Named EOS, the system operations are performed directly and synchronously without a hint of asynchronous interaction. This is quite an exciting task that presents many challenges. For example, PlayStation 5 works with saves through memory mount, which does not happen instantly. The critical point is that the saves will only be completed by unmounting the memory area. In fact, for PlayStation, you can use PlayerPrefs, which works similarly to memory mount, but all of this happens behind the scenes, out of our control. However, this approach has a significant drawback the available memory volume is limited since the primary purpose of this feature is to save game settings. Therefore, the limits are pretty expected. However, since screenshots are used for saves, this limit will be insufficient, so the first save option remains primary.What about Xbox? Xbox currently uses GDK API as the main one, and to use it at the start of the project, synchronization with cloud data always takes place. This already affects another element of the project initialization. But thats not what we are here to discuss. The main idea of working with saves on Xbox is that each time you write or read, you need to open a container, perform the necessary operations, notify the GDK API about the changes (if any), and close the container.What about the Switch? It is almost the same as on PlayStation: mounting and unmounting take time.How is data saved in the original game we are working on? Saves are created as follows: data is saved, a screenshot of the screen is taken, and it is recorded in the save. Each time a save is deleted or a new one is created, the data is reread.We created a unified save system for this project as a single entry point for any platform. Each platform has its own SDK and methods for working with saves, so creating a unified system became essential for ensuring consistency. As a result, we made a single entry point script that works using the Adapter pattern and controls entities for each platform, determining which platform is in use and running the appropriate script.Now, we have a more or less complete picture of our main challenges. What are they?The project code must be able to release the main thread without breaking the core execution logic.We need to minimize the number of calls to our save system, as even asynchronous calls can cause freezes.Since we have asynchronous calls, we must ensure the main condition only one call to the save system at a time.A simple and quite effective solution is to use Task. Async. Why? Because it allows you to pause the original logic and resume it when needed. Is this the best solution in terms of project performance? No. Will it provide the fastest and most expected result? Yes.Of course, this approach generates much additional code after compilation, but it gives us precisely the expected result. Now, we need to remove all direct calls to the file system and replace them with new calls to the "new save system implementation" that we developed on Task. Async.After that, we rework all methods that call our save system methods to async so they can pause their execution until all save or load actions are completed. We also partially rework higher-level methods in the call hierarchy if necessary. Thus, the first most critical issue has been resolved.A screenshot of The Star Named EOS. Image via Pingle Studio.What next?At this stage, we encountered an issue with Unitys Player Prefs. To address it, we created a custom analog that works similarly but saves data to a file, allowing it to be used on any platform, including Switch. This solution was necessary because Nintendo Switch does not support Unity Player Prefs and can only save using the native Nintendo SDK.Then, we had to minimize calls to the save system. The original project was implemented as follows: We have an analog of PlayerPrefs that is written to a filea dictionary with save names that are used to access screenshots and save data. Every time reading, writing, or deleting files occurs, they are read from scratch. On PCs, especially with SSDs, this is not a problem, so optimization can be ignored, but on consoles with more than a dozen saves, this can lead to serious problems.There are several solutions to this problem:Bundle a large amount of data and access the saves with a batch of operations (which requires reworking the original logic and may take a lot of time).Create a cache for already loaded saves and use cached data for repeated access.The second option was much more convenient to implement, so we chose it as the main one. It does not create additional interaction scenarios when there are dozens rewrites of one object or simultaneous read-and-write operations. Such situations might not be very obvious, but avoiding them from the start is better so they do not become a significant challenge later on.So, we have one final challenge left multiple simultaneous write, read, and delete operations. This is pretty easy to manage when using asynchronous calls. Each time we are about to work with a file, we can use an indicator, such as a semaphore or a simple variable that signals that the queue is still busy. Since we reworked the original logic for Task, async, our code awaits further calls, and a simple variable where we do an increment when the save operation starts and a decrement when it ends is sufficient. This way, we can ensure that multiple operations do not happen simultaneously, and the interaction logic eliminates competition at the entry point.Hope youve had a good read and maybe learn something. Until next time...

This week is quickly becoming a sea change moment for Sonos as the company looks to undo the damage done to its reputation since last May. It all began on Monday with the departure of CEO Patrick Spence, who was replaced by board member Tom Conrad. Then came news that chief product officer Maxime Bouvat-Merlin would also be leaving the company another indication that Sonos is serious about correcting course and taking accountability for its new app woes.In a third shakeup within the companys leadership ranks, I can report that chief commercial officer Deirdre Findlay is also stepping down. Sonos not-yet-updated corporate governance page says Findlay oversees all marketing, revenue, and customer experience organizations at Sonos. She is responsible for integrated brand strategy, geographic expansion strategies, and all go to market execution. By now, theres no arguing that Sonos go-to-market strategy for its rebuilt mobile app was deeply flawed and rushed. Before he lost his job, Spence eventually conceded that the company shouldve taken a far more cautious approach and offered the new software as a beta release while keeping the previous, more stable version in place. Instead, Sonos pushed a buggy experience on all customers and has spent the months since dealing with the resulting fallout. Do you know more about what went wrong at Sonos under Patrick Spence? In my time at The Verge, Ive covered Sonos more comprehensively (and Id like to think fairly) than any other company. Yes, that includes a long list of product leaks, but Im far more interested in shining a light on all the frustrations caused by the new app both for customers and staff and the bad decisions that led Sonos off the tracks. Those choices have had repercussions for ordinary employees who gave their best to the brand.If you have more to share about the last year at Sonos, please reach out to me confidentially and securely over Signal at (845) 445-8455 or chriswelch.01. I can also be reached via DM on Bluesky, X, or Instagram. As it relates to marketing, some Sonos employees have expressed their dismay to me over just how much money the company dumped into advertising last year. The big spends included an expansive New York City subway campaign for the Sonos Ace headphones and a holiday elves campaign that cost a staggering amount. The Ace headphones, which I maintain are a very good product, were quickly forgotten when the gravity of Sonos app problems came into focus, so the marketing had little effect. None of that sat well internally especially after layoffs in the summer.But that was then. In the span of 48 hours, interim CEO Tom Conrad has demonstrated a clear objective to get Sonos back on the right path. Im told that the moves have immediately boosted morale inside the company, with employees sensing that the new regime is serious about getting back to doing what Sonos does best.

Advanced Hallucination Mitigation Techniques in LLMs RAG, knowledge editing, contrastive decoding, self-refinement, uncertainty-aware beamsearch 0 like January 15, 2025Share this postLast Updated on January 15, 2025 by Editorial TeamAuthor(s): Mohit Sewak, Ph.D. Originally published on Towards AI. Advanced Hallucination Mitigation Techniques in LLMs: Beyond RAG to a World of RichesStuck in the RAG rut? Break free and discover other innovative and powerful techniques like knowledge editing, contrastive decoding, self-refinement, uncertainty-aware beam search, and many more.Unlocking the Riches: Mitigating Hallucinations in LLMs.Related Articles (Recommended PreReads)Hallucinations in LLMs: Can You Even Measure the Problem?Unmasking the Surprising Diversity of AI HallucinationsHallucination is like Autism, it has types and spectrum Prepare to be surprised by the Wide Spectrum of AI HallucinationsIntroduction: The Problem with HallucinationsPicture this: Youre asking your Large Language Model (LLM) to summarize the history of tea. With confidence, it replies, Tea was invented by a squirrel in 2500 BC to stay awake during winter hibernation. Wait, what? Unless squirrels were ancient alchemists, thats a bona fide hallucination a moment where the LLM concocts something so wild it might make Hollywood writers jealous.Hallucinations in LLMs arent just about quirky trivia gone rogue. In serious contexts like healthcare, finance, or legal advice they can lead to misinformation with real-world consequences. The irony? These hallucinations emerge from the same power that makes LLMs so good: their ability to predict coherent text by analyzing vast patterns in data.Now, you might say, But RAG (Retrieval-Augmented Generation) has got this under control, right? Sure, RAG has been the knight in shining armor, retrieving external facts to ground responses. But relying solely on RAG is like patching a leaky boat with duct tape it works, but only up to a point. Its time we explore newer, shinier tools to tame the hallucination beast.In this article, well venture beyond RAG into the uncharted territories of advanced mitigation techniques. From Knowledge Editing (the surgeons scalpel) to Contrastive Decoding (the ultimate fact-checker), each method is a superhero in the making. Buckle up, grab your masala chai, and lets dive into a world where hallucinations meet their match.Section 1: RAG A Safety Net with HolesRetrieval-Augmented Generation, or RAG, is the dependable workhorse of hallucination mitigation. Its charm lies in its simplicity: fetch relevant information from external sources (like knowledge bases or vector databases) and hand it over to the LLM as context for text generation. Think of it as giving your LLM a cheat sheet before the big test.But heres the catch: RAG isnt infallible. For starters, its effectiveness hinges on the quality of the retrieved data. If your vector database is filled with outdated or irrelevant information, the LLM might as well be guessing answers from a magic 8-ball. Worse, RAG doesnt inherently verify its sources, leaving room for sneaky hallucinations to creep in.Then theres the issue of dependency. By relying solely on external retrieval, we risk turning the LLM into an overconfident intern who parrots whatever they read online hardly the paragon of reliability were aiming for. Plus, the process of retrieving and integrating external data can slow things down, making RAG less practical for real-time applications.But dont get me wrong; RAG isnt obsolete. Its the dependable friend you call when youre in a pinch. However, as the AI landscape evolves, we need to pair it with more sophisticated tools. Luckily, the next-gen techniques were about to explore promise just that.Pro Tip: While RAG works wonders in knowledge-heavy domains, always double-check the quality of your retrieval sources. Garbage in, garbage out!RAG: A valiant effort, but not always enough.Section 2: Knowledge Editing: The Surgeons ScalpelImagine youre editing an encyclopedia, and you find an entry claiming Shakespeare was a software developer. Instead of rewriting the whole article, youd fix just that line. Thats what Knowledge Editing does for LLMs. Its a precision tool that allows us to update specific facts within the models parameters without disrupting its broader capabilities.How It WorksKnowledge Editing techniques like ROME (Rank-One Model Editing) and MEMIT (Model Editing via Memory) identify the specific neural connections responsible for a factual association. Then, like a neurosurgeon with a steady hand, they tweak those connections to replace incorrect information with the truth. MEMIT goes a step further by storing these edits in a memory module, ensuring the model remembers the corrections over time.For example, if your model claims the Earth is flat (thanks, internet conspiracies), you can use ROME to pinpoint the weight matrices responsible for that knowledge and adjust them to say, The Earth is an oblate spheroid. Voil, your model is now a geography nerd.Why Its CoolUnlike traditional fine-tuning, which retrains the entire model and risks breaking unrelated functionalities, Knowledge Editing is surgical. It minimizes collateral damage, preserves the models performance on other tasks, and saves computational resources.The Laugh TrackLets picture ROME and MEMIT as doctors in a sitcom:ROME: The straight-laced, no-nonsense surgeon who gets the job done in one clean swoop.MEMIT: The quirky neurologist who insists on adding a memory bank for every patient. Together, they keep your LLM in top shape.Rome and MEMIT performing a neural operation precision at its finest.Section 3: Contrastive Decoding: The Ultimate Fact-CheckerImagine youre at a party, and two friends start debating who invented the telephone. One says its Alexander Graham Bell; the other insists it was Thomas Edison. As the mediator, you compare their arguments, fact-check their claims, and declare the winner. That, in a nutshell, is what contrastive decoding does but for LLMs.What Is Contrastive Decoding?Contrastive decoding pits two models against each other during text generation. One is the primary LLM (the know-it-all), and the other is a contrastive model (the skeptic). For every token the LLM generates, the contrastive model raises its eyebrow and goes, Are you sure about that? The final output is a blend of probabilities that favor tokens deemed factual and coherent by both models.The math behind it boils down to tweaking probabilities based on a weighted difference between the primary model and the contrastive model. Think of it as having a grammar teacher who double-checks your work before you hit submit.Why It WorksLLMs can sometimes generate tokens that seem plausible but lack factual grounding. By using a smaller, less hallucination-prone model for contrast, this method curbs the primary models overconfidence. Its like having a devils advocate who questions everything, ensuring only reliable tokens make the cut.Real-Life AnalogyPicture a talkative parrot that loves to ad-lib stories. Next to it, you place a parrot trainer who knows the facts. Every time the parrot squawks nonsense, the trainer nudges it back on track. The result? A parrot that sounds not just entertaining but also accurate.The Humor FactorLets imagine the models as siblings in a family dinner debate:The LLM: Im pretty sure Napoleon won the Battle of Waterloo.The Contrastive Model: Bruh, thats not even close. Google it!Together, they settle on the truth and avoid embarrassing the family.Contrastive decoding where every word gets a fair trial.Section 4: Self-Refinement: The Models Self-Help JourneyIf youve ever written a first draft and then gone back to refine it, congratulations! Youve already practiced self-refinement, one of the most exciting techniques in hallucination mitigation. Here, the model essentially becomes its own editor, reviewing and improving its initial output.How It WorksSelf-refinement operates in a loop:The LLM generates a response.It evaluates its own output for inconsistencies or hallucinations.Based on that evaluation, it revises and improves the response.This process mirrors how humans refine their thoughts before speaking. For example, if the model initially says, The moon is made of cheese, self-refinement kicks in to correct it to, The moon consists of rock and dust.Why Its BrilliantSelf-refinement empowers the model to use its internal knowledge more effectively. By iterating on its output, it learns to identify flaws and generate more accurate text. Its like giving the model a journal where it can write, reflect, and grow wiser.The Self-Help AnalogyImagine an LLM attending a motivational workshop:Speaker: Every hallucination is a stepping stone to the truth!LLM: furiously taking notesBy the end of the session, the model has not only improved its answers but also gained a newfound sense of purpose.Fun Fact: Self-refinement aligns with the concept of reinforcement learning from human feedback (RLHF). Its like having a personal trainer who claps every time you do a perfect push-up.Self-refinement: When an LLM turns inward for better answers.Section 5: Uncertainty-Aware Beam Search: Playing It SafeIf LLMs were explorers, uncertainty-aware beam search would be their map and compass. This method helps models steer clear of risky terrain aka hallucinations by favoring safer, more reliable paths during text generation.The BasicsBeam search is a decoding strategy where multiple sequences (or beams) are explored simultaneously. In uncertainty-aware beam search, each beam is assigned a confidence score. Beams with high uncertainty likely to lead to hallucinations are pruned, leaving only the trustworthy ones.Why Its SafeThis method acts as the cautious driver who double-checks directions before making a turn. By avoiding paths with high uncertainty, it reduces the chances of generating wild, unsupported claims. Sure, it might sacrifice some creativity, but when accuracy is the goal, caution wins the day.Analogy TimeThink of beam search as a group of hikers exploring a forest. Uncertainty-aware beam search is the guide who says, That path looks sketchy; lets not go there. Thanks to this guide, the group avoids getting lost or in the case of LLMs, generating bizarre answers.Pro Tip: Use uncertainty-aware beam search when deploying LLMs in critical domains like healthcare or law. Its better to be safe than sorry!Uncertainty-aware beam search navigating safely through the forest of text.Section 6: Iterative Querying and Reasoning: Detective WorkIf Sherlock Holmes were an LLM, hed use iterative querying and reasoning to solve every mystery. This technique enables models to interrogate themselves repeatedly, poking holes in their own logic until only the truth remains. Its the ultimate self-skepticism toolkit for LLMs, ensuring theyre not just saying something plausible but also correct.How It WorksThe LLM generates an initial response.It asks itself follow-up questions or attempts alternative explanations to test the validity of its own output.Based on this internal cross-examination, it refines the response to make it more accurate and consistent.This method mirrors how detectives build their cases. They gather clues, test theories, and refine conclusions until the truth comes to light. For example, if the model generates, Unicorns are real, iterative querying might prompt it to ask, Whats the scientific evidence for unicorns? forcing it to confront the fallacy.Why Its CleverIterative querying and reasoning exploit the LLMs internal knowledge and reasoning capabilities. By engaging in self-dialogue, the model becomes more critical of its own outputs, reducing the chances of hallucination. Its like having an inner Watson double-checking Sherlocks deductions.Detective HumorImagine the LLM as a trench-coat-wearing sleuth:LLM: Elementary, my dear user. The capital of Brazil is Buenos Aires.Watson (aka Iterative Querying): Hold on, old chap. Isnt it Braslia?Cue a dramatic pause, followed by a corrected response. Mystery solved!Real-World ApplicationsThis technique shines in scenarios requiring logical reasoning or evidence-based answers. From academic research to medical diagnosis, iterative querying ensures outputs are less prone to fanciful detours.Iterative querying where LLMs play detective to find the truth.Section 7: Decoding Strategy Optimization: The Compass for TruthDecoding strategies are the unsung heroes of text generation. They determine how an LLM picks its words, and when optimized, they can steer the model away from hallucinations. Think of them as the compass guiding the LLM through the vast terrain of possible outputs.What Are Decoding Strategies?Decoding is the process of selecting the next token in a sequence. Strategies like greedy decoding, beam search, and nucleus sampling dictate how this selection happens. Optimized decoding strategies tweak these processes to balance fluency, diversity, and factual accuracy.Key TechniquesContrastive DecodingWeve already covered this one a tug-of-war between two models that ensures factuality.Factual Nucleus SamplingStandard nucleus sampling selects tokens from the top percentile of probabilities. Factual nucleus sampling adds a twist: it prioritizes tokens backed by factual evidence. Its like curating a guest list where only the well-informed are invited.Monte Carlo DropoutThis technique generates multiple outputs by applying dropout at inference time, then selects the most reliable one. Picture a robot running simulations and picking the most sensible outcome.Why It WorksDecoding strategies operate directly at the generation level, so they dont require retraining or external resources. Theyre lightweight, flexible, and powerful tools for improving output quality.Humor BreakImagine an LLM navigating a treacherous mountain trail of possible outputs:LLM: Ill take this shortcut wait, no, it leads to The moon is cheese! Lets try the main path.Thanks to decoding optimization, it always finds the safest route.Decoding optimization charting the safest route to factuality.Section 8: Combining Forces: The Avengers of Hallucination MitigationWhat happens when you bring together self-refinement, contrastive decoding, iterative querying, and beam search? You get an elite squad of techniques ready to obliterate hallucinations. Think of it as assembling the Avengers for the AI world, where each technique brings unique strengths to the table.How It WorksCombining techniques means leveraging their complementary strengths. For example:Contrastive Decoding + RAG: Retrieve external knowledge to fill gaps, then use contrastive decoding to verify and refine the response.Self-Refinement + Uncertainty-Aware Beam Search: Let the model refine its outputs while steering clear of uncertain paths.Iterative Querying + Knowledge Editing: Combine detective-like self-interrogation with precise fact updates for ironclad reliability.The Movie PitchPicture this:RAG: The resourceful genius with a database of knowledge.Contrastive Decoding: The sharp-tongued fact-checker.Self-Refinement: The introspective philosopher.Iterative Querying: The inquisitive detective.Together, theyre a blockbuster team saving the world from misinformation.Why Its the FutureNo single method can eliminate hallucinations entirely. By combining forces, we can cover each techniques blind spots, creating a system thats robust, reliable, and ready for real-world deployment.The Avengers of Hallucination Mitigation stronger together.Conclusion: Towards a World of RichesAs we wrap up this wild ride through the realm of hallucination mitigation, one thing is clear: tackling hallucinations in LLMs isnt a one-size-fits-all endeavor. Its more like assembling a toolbox, where each method from RAGs trusty retrieval skills to the philosophical self-refinement plays a vital role in ensuring our models are not just eloquent but also accurate.Mitigating hallucinations is about balance. While some techniques, like Knowledge Editing, offer precision, others, like Iterative Querying, provide introspection. Together, they form a symphony of strategies that make LLMs safer, smarter, and more reliable.But lets not forget the human element in this journey. Whether its the developers refining these methods, researchers exploring uncharted territories, or users like us questioning the answers we receive, the fight against hallucination is a collaborative effort. After all, the goal isnt just to make models less wrong its to make them tools we can trust.So, next time you sip your cardamom tea and marvel at the wonders of generative AI, remember the heroes working behind the scenes those clever algorithms ensuring your LLM doesnt claim that squirrels invented tea. Heres to a future where hallucinations are less of a nuisance and more of a curiosity we can laugh about, all while unlocking the true potential of AI.References & Further ReadingKnowledge EditingMeng, K., Bau, D., Andonian, A., & Belinkov, Y. (2022). Locating and editing factual associations in GPT. Advances in Neural Information Processing Systems, 35, 1735917372.Meng, K., Sharma, A. S., Andonian, A., Belinkov, Y., & Bau, D. (2022). Mass-editing memory in a transformer. arXiv preprint arXiv:2210.07229.Contrastive DecodingLi, X. L., Holtzman, A., Fried, D., Liang, P., Eisner, J., Hashimoto, T., & Lewis, M. (2022). Contrastive decoding: Open-ended text generation as optimization. arXiv preprint arXiv:2210.15097.Self-RefinementNiu, M., Li, H., Shi, J., Haddadi, H., & Mo, F. (2024). Mitigating Hallucinations in Large Language Models via Self-Refinement-Enhanced Knowledge Retrieval. arXiv preprint arXiv:2405.06545.Uncertainty-Aware Beam SearchZeng, H., Zhi, Z., Liu, J., & Wei, B. (2021). Improving paragraph-level question generation with extended answer network and uncertainty-aware beam search. Information Sciences, 571, 5064.Iterative Querying and ReasoningLi, W., Wu, W., Chen, M., Liu, J., Xiao, X., & Wu, H. (2022). Faithfulness in natural language generation: A systematic survey of analysis, evaluation and optimization methods. arXiv preprint arXiv:2203.05227.Qi, P., Lin, X., Mehr, L., Wang, Z., & Manning, C. D. (2019). Answering complex open-domain questions through iterative query generation. arXiv preprint arXiv:1910.07000.General TechniquesHuang, L., Yu, W., Ma, W., Zhong, W., Feng, Z., Wang, H., & Liu, T. (2023). A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Information Systems.Perkovi, G., Drobnjak, A., & Botiki, I. (2024, May). Hallucinations in llms: Understanding and addressing challenges. In 2024 47th MIPRO ICT and Electronics Convention (MIPRO) (pp. 20842088). IEEE.Mishra, A., Asai, A., Balachandran, V., Wang, Y., Neubig, G., Tsvetkov, Y., & Hajishirzi, H. (2024). Fine-grained hallucination detection and editing for language models. arXiv preprint arXiv:2401.06855.Disclaimers and DisclosuresThis article combines the theoretical insights of leading researchers with practical examples, and offers my opinionated exploration of AIs ethical dilemmas, and may not represent the views or claims of my present or past organizations and their products or my other associations.Use of AI Assistance: In preparation for this article, AI assistance has been used for generating/ refining the images, and for styling/ linguistic enhancements of parts of content.Follow me on: | Medium | LinkedIn | SubStack | X | YouTube |Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming asponsor. Published via Towards AITowards AI - Medium Share this post

Marvel is reportedly bringing back a baddie from the first Marvel Cinematic Universe film Iron Man in the upcoming Vision Quest series.According to Deadline, Faran Tahir will reprise his role of Raza Hamidmi al-Wazar from the opening sequences of Iron Man after almost two decades. He was the leader of an Afghanistan terrorist group who kept Robert Downey Jr.'s Tony Stark captive in a cave before being betrayed by Jeff Bridges' Obadiah Stane.Raza Hamidmi al-Wazar hasn't appeared since that opening 30 minutes of 2008's Iron Man but, just like The Incredible Hulk's Samuel Sterns showing up in Captain America: Brave New World, he'll make his MCU return in the near future. Vision Quest, which stars Paul Bettany as White Vision following the events of WandaVision, has no release window as of yet.Faran Tahir in 2008. Image credit: Jeffrey Mayer/WireImage.While Hamidmi al-Wazar was the head of a seemingly generic terrorist group at the time, his lore deepened years later in Phase 4 of the MCU. The group was briefly mentioned as the Ten Rings, and while Marvel likely didn't plan much beyond this subtle reference to comic book lore at the time, 2021's Shang-Chi and the Legend of the Ten Rings expanded upon it significantly.Hamidmi al-Wazar was therefore (presumably) retroactively written in as a commander of the Ten Rings organization operating its Afghanistan terrorist group, and as Shang-Chi itself left plenty open for a continuation, it's therefore possible it will be connected with Vision Quest through the returning character.Just like how Deadpool & Wolverine explored the wackier parts of the scrapped Fox Marvel universe, however, it's possible Vision Quest could be looking to do the same with forgotten elements of the official MCU.Ultron actor James Spader is also reportedly set to return for the first time since the second Avenger's film, though practically nothing else is known about the show as of yet.Image Credit: Jeffrey Mayer/WireImageRyan Dinsdale is an IGN freelance reporter. He'll talk about The Witcher all day.