Create accountLogin

Recent searches

No recent searches

Search options

Only available when logged in.
c.im is one of the many independent Mastodon servers you can use to participate in the fediverse.
C.IM is a general, mainly English-speaking Mastodon instance.

Administered by:

Server stats:

2.8K
active users

c.im: AboutStatusProfiles directoryPrivacy policy

Mastodon: AboutGet the appKeyboard shortcutsView source codev4.3.6

#LLMs

90 posts78 participants1 post today
Public
Curated Hacker News

Search-R1: Training LLMs to Reason and Leverage Search Engines with RL

arxiv.org/abs/2503.09516

arXiv logo
arXiv.orgSearch-R1: Training LLMs to Reason and Leverage Search Engines with Reinforcement LearningEfficiently acquiring external knowledge and up-to-date information is essential for effective reasoning and text generation in large language models (LLMs). Prompting advanced LLMs with reasoning capabilities during inference to use search engines is not optimal, since the LLM does not learn how to optimally interact with the search engine. This paper introduces Search-R1, an extension of the DeepSeek-R1 model where the LLM learns -- solely through reinforcement learning (RL) -- to autonomously generate (multiple) search queries during step-by-step reasoning with real-time retrieval. Search-R1 optimizes LLM rollouts with multi-turn search interactions, leveraging retrieved token masking for stable RL training and a simple outcome-based reward function. Experiments on seven question-answering datasets show that Search-R1 improves performance by 26% (Qwen2.5-7B), 21% (Qwen2.5-3B), and 10% (LLaMA3.2-3B) over strong baselines. This paper further provides empirical insights into RL optimization methods, LLM choices, and response length dynamics in retrieval-augmented reasoning. The code and model checkpoints are available at https://github.com/PeterGriffinJin/Search-R1.
#arxiv#llm#llms
Public
Miguel Afonso Caetano

"Now consider the chatbot therapist: what are its privacy safeguards? Well, the companies may make some promises about what they will and won't do with the transcripts of your AI sessions, but they are lying. Of course they're lying! AI companies lie about what their technology can do (of course). They lie about what their technologies will do. They lie about money. But most of all, they lie about data.

There is no subject on which AI companies have been more consistently, flagrantly, grotesquely dishonest than training data. When it comes to getting more data, AI companies will lie, cheat and steal in ways that would seem hacky if you wrote them into fiction, like they were pulp-novel dope fiends:
(...)
But it's not just people struggling with their mental health who shouldn't be sharing sensitive data with chatbots – it's everyone. All those business applications that AI companies are pushing, the kind where you entrust an AI with your firm's most commercially sensitive data? Are you crazy? These companies will not only leak that data, they'll sell it to your competition. Hell, Microsoft already does this with Office365 analytics:
(...)
These companies lie all the time about everything, but the thing they lie most about is how they handle sensitive data. It's wild that anyone has to be reminded of this. Letting AI companies handle your sensitive data is like turning arsonists loose in your library with a can of gasoline, a book of matches, and a pinky-promise that this time, they won't set anything on fire."

pluralistic.net/2025/04/01/doc

pluralistic.netPluralistic: Anyone who trusts an AI therapist needs their head examined (01 Apr 2025) – Pluralistic: Daily links from Cory Doctorow
#AI#GenerativeAI#LLMs
Public
Miguel Afonso Caetano

In other words, Generative AI and LLMs lack a sound epistemology and that's very problematic...:

"Bullshit and generative AI are not the same. They are similar, however, in the sense that both mix true, false, and ambiguous statements in ways that make it difficult or impossible to distinguish which is which. ChatGPT has been designed to sound convincing, whether right or wrong. As such, current AI is more about rhetoric and persuasiveness than about truth. Current AI is therefore closer to bullshit than it is to truth. This is a problem because it means that AI will produce faulty and ignorant results, even if unintentionally.
(...)
Judging by the available evidence, current AI – which is generative AI based on large language models – entails artificial ignorance more than artificial intelligence. That needs to change for AI to become a trusted and effective tool in science, technology, policy, and management. AI needs criteria for what truth is and what gets to count as truth. It is not enough to sound right, like current AI does. You need to be right. And to be right, you need to know the truth about things, like AI does not. This is a core problem with today's AI: it is surprisingly bad at distinguishing between truth and untruth – exactly like bullshit – producing artificial ignorance as much as artificial intelligence with little ability to discriminate between the two.
(...)
Nevertheless, the perhaps most fundamental question we can ask of AI is that if it succeeds in getting better than humans, as already happens in some areas, like playing AlphaZero, would that represent the advancement of knowledge, even when humans do not understand how the AI works, which is typical? Or would it represent knowledge receding from humans? If the latter, is that desirable and can we afford it?"

papers.ssrn.com/sol3/papers.cf

papers.ssrn.comAI as Artificial IgnoranceAI and bullshit (in the strong philosophical sense of Harry Frankfurt) are similar in the sense that both prioritize rhetoric over truth. They mix true, false,
#AI#GenerativeAI#Chatbots
Public
Holle Meding

📰 Classifying Genre in Historical Medical Periodicals

Next in line: Vera Danilova presents her work on genre classification in digitized periodicals from European patient organizations (1951–1990) using #LLMs as part of the #ActDisease project.

🔹 XLM-RoBERTa (UDM) led Q&A tasks with 32% more correct answers than mBERT/hmBERT.
🔹 hmBERT (UDM) topped Administrative classification (+16%)
🔹 CORE-based models excelled in legal genre prediction.

#DigitalHumanities @tuberlin #classification #NLP

Public
5h15h

#LLMs Pass the #TuringTest: Interrogators mistook GPT-4.5 for a human 73% of the time—far more than they did the actual human participant🤖🔍 arxiv.org/abs/2503.23674

arXiv logo
arXiv.orgLarge Language Models Pass the Turing TestWe evaluated 4 systems (ELIZA, GPT-4o, LLaMa-3.1-405B, and GPT-4.5) in two randomised, controlled, and pre-registered Turing tests on independent populations. Participants had 5 minute conversations simultaneously with another human participant and one of these systems before judging which conversational partner they thought was human. When prompted to adopt a humanlike persona, GPT-4.5 was judged to be the human 73% of the time: significantly more often than interrogators selected the real human participant. LLaMa-3.1, with the same prompt, was judged to be the human 56% of the time -- not significantly more or less often than the humans they were being compared to -- while baseline models (ELIZA and GPT-4o) achieved win rates significantly below chance (23% and 21% respectively). The results constitute the first empirical evidence that any artificial system passes a standard three-party Turing test. The results have implications for debates about what kind of intelligence is exhibited by Large Language Models (LLMs), and the social and economic impacts these systems are likely to have.
#AI#LLM#GenAI
Public
Holle Meding

🔍 Large-Scale Text Analysis & Cultural Change

In their talk at the workshop “Large Language Models for the HPSS” @tuberlin Pierluigi Cassotti and Nina Tahmasebi presented a multi-method approach to studying cultural and societal change through large-scale text analysis.

By combining close reading with computational techniques, including but not limited to #LLMs , they demonstrate how diverse tools can be integrated to uncover shifts in language. #DigitalHumanities

Public
Nishant Kaushik

The insatiable hunger to feed #LLMs and #AI is parasitically draining the commons and public internet. Bandwidth costs are spiking as crawlers take data for training and information. For Wikipedia, the lack of attribution means no visitors, no donors, just cost. The #ethics of AI are failing here.

I saw Tim Karr on bluesky suggest that AIs should pay fees or a tax (should that be tariffs?) into a fund that supports public content. Services like Cloudflare and Fastly that defend against bots are evolving for crawlers. In #identity, the implications for #AgenticAI, #AI, and #NHI are vast.

diff.wikimedia.org/2025/04/01/

Diff · How crawlers impact the operations of the Wikimedia projectsSince the beginning of 2024, the demand for the content created by the Wikimedia volunteer community – especially for the 144 million images, videos, and other files on Wikimedia Commons – has grow…
Public
José A. Alonso

Proof or bluff? Evaluating LLMs on 2025 USA math olympiad. ~ Ivo Petrov et als. arxiv.org/abs/2503.21934 #LLMs #Math

arXiv logo
arXiv.orgProof or Bluff? Evaluating LLMs on 2025 USA Math OlympiadRecent math benchmarks for large language models (LLMs) such as MathArena indicate that state-of-the-art reasoning models achieve impressive performance on mathematical competitions like AIME, with the leading model, o3-mini, achieving scores comparable to top human competitors. However, these benchmarks evaluate models solely based on final numerical answers, neglecting rigorous reasoning and proof generation which are essential for real-world mathematical tasks. To address this, we introduce the first comprehensive evaluation of full-solution reasoning for challenging mathematical problems. Using expert human annotators, we evaluated several state-of-the-art reasoning models on the six problems from the 2025 USAMO within hours of their release. Our results reveal that all tested models struggled significantly, achieving less than 5% on average. Through detailed analysis of reasoning traces, we identify the most common failure modes and find several unwanted artifacts arising from the optimization strategies employed during model training. Overall, our results suggest that current LLMs are inadequate for rigorous mathematical reasoning tasks, highlighting the need for substantial improvements in reasoning and proof generation capabilities.
Public *
Misinformation-Superhighwayman

"Prompt Engineering" for AI is this today's version of "Don't hold it that way" for the iPhone 4.

Users are misassigned blame for fundamental flaws in the technology, and are instructed to adopt behavioural workarounds. These improvised habits lack the causal power to fix underlying problems in the tech, but they serve to reinforce the notion that this new tech is superior to the tech it's trying to replace or "disrupt". Furthermore, users are taught, "Just keep trying and you'll get it right," without questioning whether the new tech is the problem, or to ask if the new tech has the potential to ever deliver on its promises.

A crucial difference between early smartphones and wishing that LLMs are a route to "Thinking Machines" is: later models of phones successfully matured the engineering of antennas and improved mobile reception, but LLMs are a dead-end that can never lead to real Artificial Intelligence.

This can be summarised by the AM/FM Principal: Actual Machines in contrast to Fucking Magic.

#PromptEngineering#AI#ArtificialIntelligence
Public
Miguel Afonso Caetano

"In a new joint study, researchers with OpenAI and the MIT Media Lab found that this small subset of ChatGPT users engaged in more "problematic use," defined in the paper as "indicators of addiction... including preoccupation, withdrawal symptoms, loss of control, and mood modification."

To get there, the MIT and OpenAI team surveyed thousands of ChatGPT users to glean not only how they felt about the chatbot, but also to study what kinds of "affective cues," which was defined in a joint summary of the research as "aspects of interactions that indicate empathy, affection, or support," they used when chatting with it.

Though the vast majority of people surveyed didn't engage emotionally with ChatGPT, those who used the chatbot for longer periods of time seemed to start considering it to be a "friend." The survey participants who chatted with ChatGPT the longest tended to be lonelier and get more stressed out over subtle changes in the model's behavior, too."

futurism.com/the-byte/chatgpt-

Futurism · Something Bizarre Is Happening to People Who Use ChatGPT a LotBy Noor Al-Sibai
#AI#GenerativeAI#LLMs
Public
Bibliolater 📚 📜 🖋

🔴 💻 **Are chatbots reliable text annotators? Sometimes**

“_Given the unreliable performance of ChatGPT and the significant challenges it poses to Open Science, we advise caution when using ChatGPT for substantive text annotation tasks._”

Ross Deans Kristensen-McLachlan, Miceal Canavan, Marton Kárdos, Mia Jacobsen, Lene Aarøe, Are chatbots reliable text annotators? Sometimes, PNAS Nexus, Volume 4, Issue 4, April 2025, pgaf069, doi.org/10.1093/pnasnexus/pgaf.

#OpenAccess #OA #Article #AI #ArtificialIntelligence #LargeLanguageModels #LLMS #Chatbots #Technology #Tech #Data #Annotation #Academia #Academics @ai

ExploreLive feeds
About