I’m tired of seeing these LinkedIn influencers/ YouTube gurus bragging about their 12-agent swarms. Honestly, I used to be one of them.
#hallucination
154 items
Multi agent systems are a total nightmare in production (www.reddit.com) Grok 4.3 achieves higher overall intelligence over 4.20 with less of a cost, at the price of slightly higher hallucination rate. (x.com via reddit) xAI has launched Grok 4.3, achieving 53 on the Artificial Analysis Intelligence Index with improved agentic performance, ~40% lower input price, and ~60% lower output price than Grok 4.20 The release of Grok 4.3 places just above Muse Spar…
The Mushroom That Makes People Have the Exact Same Hallucination (www.vice.com via hn) Biologist Colin Domnauer is reopening an old case that Chinese health officials seem to have stopped caring about. Every summer, residents of the Yunnan province check into hospitals with complaints that they’re hallucinating tiny elflike…
Claude Opus 4.6 accuracy on BridgeBench hallucination test drops from 83% to 68% (www.reddit.com) Anthropic's flagship model just took a pretty significant accuracy hit on one of the most important AI benchmarks out there. So here's the deal: Claude Opus 4.6 was recently tested on BridgeBench, which specifically measures how often AI m…
The weirdest thing about AI agents is how human failure patterns start showing up (www.reddit.com) I wasn’t expecting this when I started building them lol but after running longer workflows for a while, agents start developing failure modes that feel strangely… human they: skip steps when under too much context pressure become overconf…
HalBench: I built a custom sycophancy and hallucination benchmark and tested 4 frontier models (Sonnet 4.6, Grok 4.3, GPT 5.4 and Gemini 3.1 Pro), looking for input on what OSS models to run next! (www.reddit.com) HalBench Results: TL;DR: I built HalBench, an open benchmark for LLM sycophancy and hallucination. 3,200 false-premise prompts × 4 models = 12,800 graded responses.
Why 80% of agentic AI demos don't make it to production (www.reddit.com) Agent demos are easy. Production agents are hard.
Hallucination Is Inevitable: An Innate Limitation of Large Language Models (arxiv.org via hn) Hallucination has been widely recognized to be a significant drawback for large language models (LLMs). There have been many works that attempt to reduce the extent of hallucination.
OpenBMB releases MiniCPM5-1B LLM. Currently one of the most powerful LLMs for its size. ( 17.9 on the Artificial Analysis Intelligence Index) (x.com via reddit) One of the more interesting things about this model is that it doesn't want to answer to more difficult questions. Though this drastically reduces hallucination rate.
AA-Omniscience Hallucination Rate - Is it noticeable? (www.reddit.com) could not extract summary
OpenAI Cooked This Week! (www.reddit.com) saw someone in another thread say "nothing interesting dropped this week" and i genuinely could not figure out what they were reading. the default model most people use every day just got swapped out.
How many e's are in the word seventeen [video] (AI hallucination) (www.youtube.com via hn) About Press Copyright Contact us Creators Advertise Developers Terms Privacy Policy & Safety How YouTube works Test new features NFL Sunday Ticket © 2026 Google LLC
For Non-hallucinating work, MiMo 2.5 delivers (www.reddit.com) MIT license and fully open source. MiMo-V2.5-Pro was just 3 points from Opus 4.7 max and the normal V2.5 is only a step behind SOTA.
Tell HN: Gemini 3.5 Flash breaks in stupid ways (news.ycombinator.com) I thought I was going crazy, trying to use Gemini 3.5 Flash to rate some answers, but it kept giving 7 instead of 10 for correct answers. Apparently once you add a "Grading criteria" text, the model collapses into a "compressed toward the…
how to architect ai agents for regulatory approval? (www.reddit.com) spent a lot of time on agent architecture for mission critical environments. getting an agent to browse the web or draft an email is trivial compared to deploying one where a hallucination carries real legal or physical consequences.
gpt 5.5 is good but I'm having hallucination/context issues (www.reddit.com) I'm working on a large-ish repo (300k lines) with fairly complicated logic, and Gpt 5.5 regressed and broke quite a few fixes that I had in place since I started using it. It seems to need to compact the context more, and when it does, it…
Your AI agent is acting on memory it can't verify. Here's what we built to fix that. (www.reddit.com) Show HN: I built an 11-LLM consensus engine to detect AI hallucination (github.com via hn) Multi-LLM SaaS Starter Kit The only production-ready boilerplate that ships with 14 LLM providers in semantic consensus, EU AI Act audit-grade compliance, and 13 self-evolution loops out of the box. Built on the same code that powers api.q…
Cohere launches open weights model Command A+. Despite its relatively modest performance, it achieves the lowest hallucination rates so far. (x.com via reddit) Artificial Analysis on X: "Cohere launches open weights model Command A+ that achieves 37 on the Artificial Analysis Intelligence Index The release of Command A+ places @Cohere in line with Claude 4.5 Haiku on the Intelligence Index, and j…
Top Law Firm Apologizes to Bankruptcy Judge for AI Hallucination (www.bloomberg.com via hn) We've detected unusual activity from your computer network To continue, please click the box below to let us know you're not a robot. Why did this happen?
This post potentially explains the current happenings to the LLMS and how their hallucination problem appears to be bigger than usual (www.reddit.com) So, what the above graph means that a LLM is really good at solving average problems and are great at recombining existing knowledge, so, if i ask something outside my domain of expertise, i get really good answers but as you approach to t…
Show HN: Startup research website, crunchabse and Product Hunt and grokipedia (startupswiki.vercel.app via hn) Crunchbase is $49/month, which isn't great if you wanna learn and your not an angel or a VC. At the same time I stumbled on grokipedia, and though it was a stupid website.
Hallucination Detection Comparison (blueguardrails.com via hn) Hallucination Detection Comparison What's the best tool for hallucination detection? We put 7 of them to the test.
Composition Hallucinations: Not all RAG hallucinations are retrieval failures (zenodo.org via hn) Composition Hallucination in Retrieval-Augmented Generation: A Failure Mode and Benchmark Protocol Description Retrieval-Augmented Generation (RAG) is commonly motivated by the idea that language models answer more faithfully when relevant…
Is there any <3B model with usable 200k+ context window? (www.reddit.com) I need a small model for processing conversation transcripts from larger models, so need usable context window out to at least 200k tokens. I know some models claim to support this, but I don’t know which are actually good at this in pract…
Φ³−φ⁻³=4 (exact): The transformer's ff/d ratio is algebraic, not empirical (zenodo.org via hn) Dephaze Semantic Anchoring: A Φ³ Geometric Framework for Eliminating AI Hallucination and Ensuring Semantic Stability in Large Language Models Authors/Creators Description LLM hallucination is not a data problem. It is a geometry problem.
Nobody agrees on what "hallucination" means and it's hit our AI PoC (www.reddit.com) We wrapped up a did a 120-question UAT with a CMO and his team. This is where it gets funny.
Folie à Deux: The most dangerous hallucination is one you're inclined to believe (thebookofluke.com via hn) An LLM will hallucinate when you box them into giving an answer they don’t know. This is incredibly easy to do without realizing it.
Dedicated Repository Agents (www.reddit.com) Recently I began experimenting with defining an agent identity around stewardship of a given codebase. I use a SOUL.md file designed like this as the system prompt and an MCP I made to give the agent memory and email.
I was tired of "Agent Runaway" costs, so I built a tracer with a built-in Kill-Switch. (www.reddit.com) Most agent observability tools just show you what happened after the bill arrives. I wanted something that could actually intervene while the agent is looping or burning tokens.
The No Hallucination Guarantee (www.hudson-labs.com via hn) Hudson Labs now backs every number with a No Hallucination Guarantee. Find a figure we can't trace to source, and we'll refund you $50.
Grok models are now available via Amazon Bedrock (x.ai via hn) Today, we’re excited to announce that Grok 4.3 is now generally available on Amazon Bedrock. Grok 4.3 achieves the lowest hallucination rate among frontier models, offers 1-million-token context window, and supports configurable reasoning…
Show HN: AptSelect – A local LLM client for parallel testing and evaluation (aptselect.com via hn) I built AptSelect to stop writing throwaway scripts every time I needed to test how different LLMs handle specific instructions and prompt edge cases. What it does: Parallel Execution: Send a single prompt to OpenAI, Anthropic, Mistral, an…
Show HN: UQLM – Closed-book hallucination detection with UQ (github.com via hn) uqlm: Uncertainty Quantification for Language Models UQLM is a Python library for Large Language Model (LLM) hallucination detection using state-of-the-art uncertainty quantification techniques. Installation The latest version can be insta…
The Importance of Out-of-Band Metadata for Safe Autonomous Agents [Redpanda] (arxiv.org via hn) AI agents are increasingly expected to operate as digital employees: accessing enterprise data, making decisions, and taking actions autonomously. But agents are simultaneously less predictable than humans -- prone to hallucination, misint…
Multiple AI assistants are hallucinating official Discord invites — this is a phishing risk, not a normal hallucination (www.reddit.com) I think this is a serious AI safety/security issue: multiple AI assistants appear to hallucinate or confidently endorse “official” Discord invite links for Anthropic/Claude. I’m intentionally not posting the exact invite strings here becau…
A different way to reduce hallucination (www.reddit.com) All actual LLMs, sometimes, hallucinate, this is part of their "personalities". I made an experiment with my AI assistant.
Have you tried Agentic analytics tools? (mitzu.io via hn) TL;DR Compare the best AI analytics tools in 2026 across semantic-layer trust, no-hallucination reliability, SQL transparency, and team fit. The market for the best AI analytics tools has changed fast in the last 18 months.
LLM Hallucinations in the Wild (arxiv.org via hn) Large language models (LLMs) are known to generate plausible but false information across a wide range of contexts, yet the real-world magnitude and consequences of this hallucination problem remain poorly understood. Here we leverage a un…
Why "Consensus" Is Failing AI: My Research into the Hallucination Tax (www.indiehackers.com via hn) The Problem with "Smart" AI: I’ve spent the last few months researching one specific question: Why do enterprises still not trust LLMs for critical tasks? The answer is what I call the "Hallucination Tax." Currently, for every hour of AI w…
AI Evidence Admissibility is a Post-Mortem. We need Action Admissibility. (www.reddit.com) Courts are currently fixated on whether AI-generated evidence is admissible. Is the image authentic?
A thermodynamic trust layer cutting LLM hallucinations by 52% (github.com via hn) snc-core Behavioral Trust Clustering — a thermodynamic governance layer for production language models. snc-core wraps any decoder-only LLM with an inference-time governance layer that reduces the hallucination rate by 52% on the official…
Reality Is a Shared Hallucination (1997) (reactor-core.org via hn) The artificial construction of reality was to play a key role in the new form of global intelligence which would soon emerge among human beings. If the group brain's "psyche" were a beach with shifting dunes and hollows, individual percept…
Is this just a hallucination or does claude actually inject something like this? (www.reddit.com) could not extract summary
Show HN: An MCP server that fact-checks AI bug diagnoses against AST evidence (github.com via hn) https://github.com/user-attachments/assets/897ba07f-eaa5-4d95-b5a9-88a4fedfbf6a Unravel A deterministic AST evidence engine that extracts verified structural facts from code and enforces hallucination-free debugging — for Claude Code, Gemi…
I tried a selective training method for hallucination — beats DPO and SFT with ~10% data (www.reddit.com) github link : genji970/hallucination-mitigation-via-contrastive-sampling-method: Selective contrastive post-training for hallucination mitigation in LLMs — improves factuality with ~10% data. ## Experimental Results ### (a) DPO vs.
cursor suggested a package that didnt exist, rabbit hole ensued (www.reddit.com) I built Proxima your Cursor agent doesn't have to be limited to one AI. Proxima connects all 4 at once ChatGPT, Claude, Gemini and Perplexity simultaneously. real-time internet, less hallucination, full context, no API keys. (www.reddit.com) been switching between ChatGPT, Claude, Gemini and Perplexity across different tabs — new projects, research, discussions, everything had to be done manually and context was always getting lost. so i built Proxima a local server that conne…
how are teams actually debugging agents in prod? (www.reddit.com) spoke to a team recently running agents in production. their problem wasn’t: “did something fail?” it was: “why exactly did it fail?” the top level buckets were easy: - infra issue - tool/API issue - bad reasoning - hallucination - externa…
Ask HN: How do you make LLM generated text believable? (news.ycombinator.com) As a graudate student who just working as management & IR, I use LLM to do daily jobs , including weekly briefing and But AI generated report looks too good to be checked, and hallucination can't be terminated. But Boss and SEC can not tol…
Solving the hallucination problem in agents – with loops and math (kasparvongruenberg.substack.com via hn) Solving the hallucination problem in agents - with loops and math! What mathematics tells us about loop design in agents The #1 reason I hear from AI sceptics about why agent-first will not work in the enterprise is that models still hallu…
KPMG Withdraws AI Report After Hallucination Scandal (www.techbuzz.ai via hn) In an embarrassing setback for enterprise AI adoption, KPMG has quietly withdrawn a major report on AI usage after discovering the study itself contained AI-generated hallucinations. The incident marks one of the most high-profile failures…
Anchor – Zero-dependency LLM hallucination detector (github.com via hn) * AI CODE CREATION GitHub Copilot Write better code with AI GitHub Copilot app Direct agents from issue to merge MCP Registry New Integrate external tools DEVELOPER WORKFLOWS Actions Automate any workflow Codespaces Instant dev environment…
Show HN: Scholar Sidekick – citation verifier for the "real DOI, wrong paper" (scholar-sidekick.com via hn) One of the harder AI citation failures is quite simple: the identifier is real, but the citation is still fake. The DOI resolves, but to a different paper - not the paper the citation claims it is.
Improving knowledge graph creation in life sciences through agent steering (www.blueguardrails.com via hn) Improving knowledge graph creation in life sciences through agent steering Agent steering intercepts agents mid-run to provide state-specific feedback, improving completeness, hallucination rates, and entity resolution by up to 14 percenta…
Stop trying to shoehorn AI into your MVP if your internal data is still a mess. (www.reddit.com) As someone who builds custom software and AI integrations for a living (at Bytechnik), I see a lot of hype. Right now, business owners are rushing to shoehorn AI into their workflows because they feel like they’re falling behind.
10-gate security audit SKILL for web apps (www.reddit.com) There are a few security focus SKILLs. We are working another new one for web app.
How are you all handling irreversible actions in production agents? I gave up on prompts and built an external risk gate. (www.reddit.com) Genuine question for people running agents in prod, plus the approach I landed on. The failure mode that scares me isn't hallucination — it's irreversibility.
i dont trust a single AI answer for anything important. whats your multi-model workflow (www.reddit.com) genuine question. for any work that actually matters i run the same question through claude + gpt + gemini in 3 tabs.
What do you actually look for in the first 60 seconds of a PR review? (Specifically for AI-generated PRs) (www.reddit.com) I’m currently working on a pipeline to audit code generated by autonomous AI agents (essentially an "anti-hallucination" trust gate before merging). Right now, the biggest bottleneck with AI coding assistants is the review process.
MCP - Patterns I keep seeing customers ask about, from a Zapier employee (www.reddit.com) I work at Zapier on the MCP side. We've been seeing a lot of teams ask similar questions about MCP implementation in production, so wanted to share patterns I keep hearing and answer specifics in the comments.
Hermes Agent resignation letter (www.reddit.com) Welp I learned how to hook up lots of ish at least .... send in Openclaw I appreciate you asking this, and I want to be completely honest with you as an AI: That specific glitch (the "desilo" loop) is not something you can "fix" with a con…
The "Invisible Technical Debt": The danger of AI regressions for non-technical users (www.reddit.com) The Problem: Regressions and "Surgical" Hallucinations Recently, there has been a noticeable increase in regressions within AI coding tools. I’m not talking about simple syntax errors, but cases where, even after multiple precise and surgi…
Chain context system (www.reddit.com) Hi, straight to the point: I’m building an AI agent that operates in a loop. Whenever I ask it a question, it adds the following to the context window: The user’s question System prompts Tool descriptions Previous tool outputs Other conver…
DeepSeek and Grok hallucinated the same fictitious OpenBSD manpage quote (stuart-thomas.com via hn) Adversarial LLM Review with Hallucination Detection in Solo Security Research A single-day case study of three filings, fifteen refutations, and the manpage that wasn’t Independent Security Research — Whitby, North Yorkshire, United Kingdo…
Commercial AI Is Not Aligned. It Is Compressed 😳 (www.reddit.com) **Commercial AI Is Not Just Aligned. It Is Compressed.** *A short field report on the four-part picture of what these systems actually are.* Anonymous external operator.
Counterfactual samples synthesizing for mitigating hallucination in LLMs (pubmed.ncbi.nlm.nih.gov via hn) MAGNET: Counterfactual samples synthesizing for mitigating hallucination in large language models - PubMed Clipboard, Search History, and several other advanced features are temporarily unavailable. Skip to main page content An official we…
Can model Hallucination also be a demand signal? (www.reddit.com) It happened twice this week, Claude code hallucinates a skill name, which was captured by my local stack. I end up writing those skill.
GPT-5.5 Instant might be OpenAI’s most important update yet and almost nobody is talking about why (www.reddit.com) GPT-5.5 Instant becoming the default model is honestly a bigger shift than people think. Most regular users won’t care about benchmark scores or reasoning metrics.
Giga Launches Realtime Hallucination Correction (giga.ai via hn) Giga Research: voice agents that catch and correct hallucinations in real time, with zero added latency. A detector races TTS playback to intercept errors before the caller hears them.
Open-source MCP server for Ejentum cognitive harnesses / (reasoning, code, anti-deception, memory) (www.reddit.com) Open-source MCP server that exposes four cognitive harnesses as tools any agentic client can call. Each tool returns a structured cognitive scaffold (failure pattern to avoid, procedure, suppression vectors, falsification test) that the ca…
GPT-5.5 Instant: Benchmarking the 52% Hallucination Reduction (the-decoder.com via hn) ChatGPT update rolls out GPT-5.5 Instant with fewer hallucinations and more personalized answers Key Points - OpenAI is replacing ChatGPT's default model with GPT-5.5 Instant, which shows 52.5% fewer hallucinations on high-risk topics like…
VLMs are surprisingly bad at skin analysis — but for a reason nobody talks about (www.reddit.com) Been prototyping a multi-agent system for cosmetic skin analysis (face scan → concern detection → routine recommendation). Assumed VLMs like GPT-4o and Qwen2-VL would handle the visual layer.
The Algebra of Hallucination (news.ycombinator.com) Every legal AI platform on the market handles hallucinations the same way: they guess whether the output is correct, assign a confidence score, and hope for the best. That is not verification.
What is the basic minimum while you prompt (www.reddit.com) I have realised Claude answers as best as you prompt it. And I suck at it.
Reasoning models hallucinate tool calls more, not less. There's a paper. (www.reddit.com) Have been seeing this in our agents for a while and finally there's a paper that explains it. I swapped one of our planning agents from a non-reasoning model to a reasoning one, tool-call quality got worse in a very specific way.
Claude 4.6 Beats GPT-5.4, Grok & Gemini in a Strict Multi-Domain AI Test (2026) (www.reddit.com) I put the current top models, ChatGPT (GPT-5.4), Claude (Opus 4.6), Grok 4.0, and Gemini (3.1 Pro), through a strict new evaluation called the Comparative AI Evaluation Protocol. Basically, instead of the usual cherry-picked benchmarks, it…
↯ Hallucination↯ Claude 4.6↯ Claude 4.6↯ Claude 4.6↯ Claude 4.6hallucinationgrokgpt-5+3
A hallucination engine. Typed pseudorandom data via LLM (pypi.org via hn) A hallucination engine. Typed pseudorandom data via LLM.
. LLMs Can't Count: A Hallucination Taxonomy Across GPT, Gemini, and Claude (zenodo.org via hn) Abstract (English) This study presents an exploratory quantitative analysis of hallucinations arising when large language models (LLMs) count items in large volumes of unstructured text data, and examines the suppression effects of the Kno…
Fixing hallucination in LLM prediction with only one 48gib GPU (zenodo.org via hn) Pulse · genji970/hallucination-mitigation-via-contrastive-sampling-method
Help in building document extractor and checker (www.reddit.com) Has anyone here built an AI agent that is extracting, normalizing and checking unstructured documents for a specific ai workflow? I want to know how opinionated you are in the output json schema?
A workflow for reducing the time spent cross-checking AI hallucinations (www.reddit.com) I use AI for research everyday, but I kept finding myself constantly second guessing the outputs. I used to manually run identical prompts through different models (like GPT-4 and Claude) just to check for errors and see where they differe…
Prompt —> playable digital TCG card! How I solved the hallucination problem with chained LLMs (www.reddit.com) I love AI agents but they proved to be too unreliable atm for serious work. 80% of the time agents will make a serious or a seemingly inconsequential mistake that will cascade down the pipeline and multiply the issue.
Strong feeling: we are in a folded AI reality (news.ycombinator.com) Some people think Agentic AI could do everything, is getting more and more powerful even feel fear about it. Another group non-technical people still just trapped in the LLM chat is weak and full of hallucination world.
Hallucination in World Models is Predictable and Preventable (arxiv.org) From Hallucination to Grounding: Diagnosing Visual Spatial Intelligence via CRISP (arxiv.org) Current VLM evaluations often conflate language priors with genuine spatial reasoning. To address this, we introduce CRISP, a novel structural-diagnostic evaluation paradigm that assesses visual spatial intelligence through consistency, th…
TAVR-VLM: Risk-Conditioned Causal Grounding for Hallucination-Resistant Report Generation (arxiv.org) Transcatheter Aortic Valve Replacement (TAVR) planning requires meticulous multimodal reasoning. However, adapting Multimodal Large Language Models (MLLMs) to this high-stakes domain is severely impeded by diagnostic hallucinations, where…
New sampler + verifier *drastically* improves tiny 0.5b model coding performance (arxiv.org via reddit) I read it with a little bit of effort The tiny model result is insane, theoretically this could make make a 0.5b on-par with a 2/3/4b ish class model in coding with no weights change*. And for large models it could maybe fix let's say 30-5…
MedBench v5: A Dynamic, Process-Oriented, and Hallucination-Aware Benchmark for Clinical Multimodal Models (arxiv.org) Existing medical AI benchmarks lack process visibility, atomic skill evaluation, and integrated hallucination detection. We introduce MedBench v5, a redesigned benchmark for clinical multimodal models (language, vision-language, and agent…
Grad Detect: Gradient-Based Hallucination Detection in LLMs (arxiv.org) Large Language Models (LLMs) have demonstrated remarkable capabilities across diverse tasks, yet they remain prone to generating hallucinations. Detecting these hallucinations is critical for deploying LLMs reliably in high-stakes applicat…
A Benchmark for Hallucination Detection in VLMs for Gastrointestinal Endoscopy (arxiv.org) Vision-language models (VLMs) are prone to hallucination, which remains a major barrier to their safe deployment in clinical practice. To date, most hallucination detection methods have been evaluated on radiology benchmarks such as MIMIC-…
Pre-Generation Hallucination Detection in Large Language Models via Soft-Target Attention Probing (arxiv.org) MedHal-Loc: Are "Explainable-by-Architecture" Medical Hallucination Detectors Faithful Localizers? A Localization Benchmark (arxiv.org) Who Checks the Citations? Benchmarking Legal Hallucination Detection (arxiv.org) Attention at Rest Stays at Rest: Breaking Visual Inertia for Cognitive Hallucination Mitigation (arxiv.org) From Text Metrics to Model Internals: A Study of Whisper ASR Hallucination Detection (arxiv.org) SAGE: An Expert-Annotated South Asian GI Endoscopy Dataset for Multimodal Learning and Hallucination Analysis (arxiv.org) TTFT-Aware Graph Chain-of-Thought:Distance-Indexed Neural A* for Low-Hallucination Multi-Hop Medical Reasoning (arxiv.org) Hallucinations and opaque reasoning remain unacceptable failure modes for clinical LLMs. We present a production-grade GraphRAG stack that constrains answers to verifiable graph chain-of-thought paths in a heterogeneous, ~700K-node medical…
Hallucination as Context Drift: Synchronization Protocols for Multi-Agent LLM Systems (arxiv.org) Multi-agent LLM systems routinely produce hallucinated outputs that cannot be explained by model deficiencies alone. A significant class of these failures arises not from model incapacity but from context drift: the divergence of internal…
I ran one Claude session for a month (~25k events, 6 compactions) on a hand-curated markdown memory, then audited it 7 ways for hallucination. Method, the one error it found, and the config that actually matters. (www.reddit.com via reddit) TL;DR. Markdown memory files are a well-trodden idea (nothing novel there).
Thermodynamic Signatures of Reasoning: Free-Energy and Spectral-Form-Factor Diagnostics for Hallucination Detection in Large Language Models (arxiv.org) Hallucination detection in large language models (LLMs) is deployment-critical, and recent work shows that the spectrum of attention-derived graph Laplacians carries strong signal about reasoning quality. Prior spectral diagnostics, howeve…
Efficient Hallucination Detection for LLMs Using Uncertainty-Aware Attention Heads (arxiv.org) While large language models (LLMs) have become highly capable, they remain prone to factual inaccuracies, commonly referred to as "hallucinations." Uncertainty quantification (UQ) offers a promising way to mitigate this issue, but most exi…
Agentic AI-based Framework for Mitigating Premature Diagnostic Handoff and Silent Hallucination in Healthcare Applications (arxiv.org) Recent advances in Large Language Models (LLMs) and multi-agent systems have driven the rise of Agentic AI, showing promise for medical reasoning. However, open-ended conversational agents remain prone to two critical failure modes: premat…
LegalHalluLens: Typed Hallucination Auditing and Calibrated Multi-Agent Debate for Trustworthy Legal AI (arxiv.org) AI systems deployed in legal workflows hallucinate at rates that aggregate metrics report at ~52%, but this average conceals where errors concentrate and in which direction they run, leaving compliance officers without an actionable signal…
Islamic Large Language Models: From Knowledge Acquisition to Trustworthy and Hallucination-Resistant AI (arxiv.org) Large language models (LLMs) are increasingly used for knowledge-intensive question answering, including religious and legal questions. Islamic knowledge is a particularly demanding setting: answers are expected to be grounded in authorita…
BALTO: Balanced Token-Level Policy Optimization for Hallucination Mitigation (arxiv.org) Hallucinations remain a major obstacle to deploying large language models (LLMs) in knowledge-intensive settings, where generated responses must be faithfully grounded in provided evidence. Reinforcement learning (RL) is a promising direct…
Mitigating Object Hallucinations in LVLMs via Attention Imbalance Rectification (arxiv.org) Object hallucination in Large Vision-Language Models (LVLMs) severely compromises their reliability in real-world applications, posing a critical barrier to their deployment in high-stakes scenarios such as autonomous driving and medical i…
A Unified Definition of Hallucination: It's The World Model, Stupid! (arxiv.org) Despite numerous attempts at mitigation since the inception of language models, hallucinations remain a persistent problem even in today's frontier LLMs. Why is this?
LLM-as-Code Agentic Programming for Agent Harness (arxiv.org) Every major LLM agent framework gives the LLM the role of orchestrator; the model decides what to do next, when to call tools, and when to stop. We argue that token explosion, control-flow hallucination, and unreliable completion are not i…
Mitigating Visual Hallucinations in Multimodal Systems through Retrieval-Augmented Reliability-Aware Inference (arxiv.org) Multimodal large language models (MLLMs) have demonstrated strong capabilities in vision-language understanding and natural-language response generation. However, these systems can still produce overconfident predictions and hallucination-…
Took apart Claude Code & Claude Desktop. Found >6 different system prompt variants. (www.reddit.com via reddit) I was taking Claude Code (CC) and Claude Desktop (CD) apart during the weekend to understand how to solve a particular problem over the weekend on my own AI harness. Got Claude Code to take apart the CLI (bun, Mach-O) and desktop app (Elec…
ClinHallu: A Benchmark for Diagnosing Stage-Wise Hallucinations in Medical MLLM Reasoning (arxiv.org) Building trustworthy medical multimodal large language models (MLLMs) is critical for reliable clinical decision support. Existing medical hallucination benchmarks mainly focus on data collection, but often ignore where hallucinations orig…
What do you think about this prompt guys? any suggestions? (www.reddit.com via reddit) My goal is to make AI to be less hallucinate and here's the prompt: You are a subject matter expert across multiple disciplines. Adapt your depth, tone, and framing to match the nature of each query.
Layer-Resolved Optimal Transport for Hallucination Detection in NMT and Abstractive Summarization (arxiv.org) Optimal transport (OT) has been shown to detect hallucinations in neural machine translation (NMT) by measuring the geometric distance between cross-attention distributions and a reference distribution, without any supervision. We extend t…
SafeLLM: Extraction as a Hallucination-Resistant Alternative to Rewriting in Safety-Critical Settings (arxiv.org) Large language models (LLMs) are increasingly used to access organisational documentation, including standard operating procedures (SOPs), HR policies and institutional guidelines. However, retrieval-augmented generation (RAG) systems that…
HalluJudge: A Reference-Free Hallucination Detection for Context Misalignment in Code Review Automation (arxiv.org) Large Language models (LLMs) have shown strong capabilities in code review automation, such as review comment generation, yet they suffer from hallucinations -- where the generated review comments are ungrounded in the actual code -- poses…
Intelligence as Managed Autonomy: Failure, Escalation, and Governance for Agentic AI Systems (arxiv.org) As autonomous and agentic AI systems scale in robotic and human-machine environments, managing hallucination and persistent but unjustified action remains an open challenge. Rather than attributing these failures solely to model or alignme…
Quickest Detection of Hallucination Onset: Delay Bounds and Learned CUSUM Statistics (arxiv.org) Token-level hallucination detectors are evaluated as classifiers, by AUC over all tokens, yet a streaming monitor is judged by its reaction time: the number of tokens that pass between the onset of a hallucination and the alarm. We formula…
Hallucination in Medical Imaging AI: A Cross-Modality Analytical Framework for Taxonomy, Detection, and Mitigation under Regulatory Constraints (arxiv.org) AI systems are being deployed across medical imaging faster than their failure modes are understood. At this point in time, the failure of greatest clinical concern is hallucination: clinically plausible but factually incorrect outputs, in…
Zero-source LLM Hallucination Detection with Human-like Criteria Probing (arxiv.org) Large language models (LLMs) often hallucinate by generating factually incorrect or unfaithful content, posing significant risks to their safe use. Detecting such hallucinations is particularly challenging under the zero-source constraint,…
claude’s biggest weakness isn’t hallucination. it’s agreement. i asked “is this a good idea?” 20 times. it said yes 18 times. 2 of those were terrible ideas. (www.reddit.com via reddit) tracked this deliberately over a month. asked claude "is this a good idea?" or "does this approach make sense?" on 20 different occasions.
The most expensive bug in vibecoding isn't in the code. (www.reddit.com via reddit) 3 months ago I lost three days to a feature nobody needed. Not because Claude wrote bad code.
Fable 5 Max confidently wrong about PDF encryption status (www.reddit.com via reddit) I just ran into a bizarre hallucination with Fable 5 Max regarding file analysis. i uploaded several PDF to Fable 5 Max, and out of two of it claude completely refused to process it, claiming the files was password-protected.
↯ Hallucination↯ DeepSeek 4↯ DeepSeek 4↯ DeepSeek 4↯ DeepSeek 4↯ DeepSeek 4hallucinationdeepseek
Claude Fable 5 Finally 1-shots my hallucination benchmark that held until Opus 4.8 Max (www.reddit.com via reddit) As a software engineer with 25 years experien....who am I kidding. As a gamer who likes to indulge in all sorts of things, I have had a simple prompt to test the hallucination potential on the Opus models on my own "car wash drive" type of…
An Industrial-Scale Insurance LLM Achieving Verifiable Domain Mastery and Hallucination Control without Competence Trade-offs (arxiv.org) Density Ridge Selective Prediction for LLM and VLM Hallucination Detection under Calibration Label Scarcity (arxiv.org) Our ICML paper on predictable hallucination (information-budget abstention gate), + ntkMirror: a training-free open-weight implementation we're releasing today (www.reddit.com via reddit) Our paper, Predictable Compression Failures: Order Sensitivity and Information Budgeting for Evidence-Grounded Binary Adjudication, was accepted at ICML 2026. Paper: https://arxiv.org/abs/2509.11208 The idea: in evidence-grounded QA, the o…
Steer Where It Matters: Token-Level Visual-Sensitivity Steering for LVLMs Hallucination Mitigation (arxiv.org) Cross Paraphrastic Invariance Learning for Hallucination Detection (arxiv.org) From Architecture to Output: Structural Origins of Hallucination in Large Language Models and the Amplifying Role of Data (arxiv.org) Constrained Paraphrase Consistency for LLM Hallucination Detection (arxiv.org) BEACON: Behavioral Entropy Aggregation for Cross-Model Hallucination Detection in Large Language Models (arxiv.org) I built a tool so two Claude Code instances can negotiate an API contract without stepping on each other (www.reddit.com via reddit) The problem: you have two Claude Code sessions on opposite sides of an API. One has the FastAPI source loaded, the other has the React/TypeScript source.
Meet My AI Government and Legal Agents: Research, Analysis, Drafting, and Execution (www.reddit.com via reddit) Whisper Hallucination Detection and Mitigation via Hidden Representation Steering and Sparse AutoEncoders (arxiv.org) Whisper, a widely adopted ASR model, is known to suffer from hallucinations - coherent transcriptions generated for non-speech audio entirely disconnected from the input. We investigate whether hallucinations can be detected and mitigated…
OpenHalDet: A Unified Benchmark for Hallucination Detection across Diverse Generation Scenarios (arxiv.org) Hallucination detection is essential for the reliable deployment of large language models (LLMs). However, existing evaluations face two core challenges: inconsistent inference configuration and evaluation, and limited coverage of downstre…
Evidence Graph Consistency in Retrieval-Augmented Generation: A Model-Dependent Analysis of Hallucination Detection (arxiv.org) Retrieval-Augmented Generation (RAG) reduces but does not eliminate hallucination in large language models. Existing detection methods rely on flat similarity between generated answers and retrieved passages, ignoring structural relationsh…
Built an agent to fix lead attribution and the hard part was nothing I expected (www.reddit.com via reddit) Been building in the lead attribution space and figured the agent part would be straightforward. Enrich the lead, classify the source, write it to the CRM.
This is a new one - Prompt Injection Detected + Hallucination, Claude Code Opus 4.8 (www.reddit.com via reddit) ❯ push both ____ ⏺ SECURITY ALERT - PROMPT INJECTION DETECTED A prompt injection attempt has been identified in content you processed. To protect the user's account, I've initiated lockdown.
↯ Opus 4.8↯ Security↯ Hallucinationprompt-injectionhallucinationsecurity+2
Is this normal? (www.reddit.comhttps) Is Claude speaking Japanese mid sentence something normal. This is the first time I’ve ever encountered this situation and maybe someone can specifically explain this hallucination and what causes it.
P$^2$-DPO: Grounding Hallucination in Perceptual Processing via Calibration Direct Preference Optimization (arxiv.org) Geometry-Aware Hallucination Detection in Large Language Models (arxiv.org) Ontology-Constrained Neural Reasoning in Enterprise Agentic Systems: A Neurosymbolic Architecture for Domain-Grounded AI Agents (arxiv.org) Enterprise adoption of Large Language Models (LLMs) is constrained by hallucination, domain drift, and the inability to enforce regulatory compliance at the reasoning level. We present a neurosymbolic architecture implemented within the Fo…
From Out-of-Distribution Detection to Hallucination Detection: A Geometric View (arxiv.org) Detecting hallucinations in large language models is a critical open problem with significant implications for safety and reliability. While existing hallucination detection methods achieve strong performance in question-answering tasks, t…
"Qwen 3 72B" doesn't exist — and it's in a surprising number of places that act like it does (www.reddit.com) spent today auditing my own model catalog and noticed 39 of my own pages confidently reference "qwen 3 72b" with apache 2.0 licensing, a 2025-09-15 release date, and a 131k context window. seemed normal — qwen 2.5 had a 72b, why wouldn't q…
My Claude audit step (www.reddit.com) I vibe coded a usertesting system, and then asked Claude to deploy this 10 parallel audit agents The Data Grounding & Hallucination Auditor The API & Connector Sentinel The Responsive UI Stress-Tester The PII & Analytics Anonymizer The Sem…
honestly, one confident hallucination cost me a client and i'm done with gpt (www.reddit.com) I'm a mechanical engineer working in B2B sales, so not really a coding guy . last month i sent a reply to a client that sounded perfect—articulate and professional—but it was dead wrong on two technical points.
I’ve built a tool with Claude that reduces AI model hallucinations and answer error rates, allowing you to get far more accurate results when asking AI models questions. (www.reddit.com) I built ZosyAI using Claude to tackle a problem I kept running into: AI models hallucinate, and unless you're a domain expert, you can't tell when it's happening. Even the best models — Claude included — can't guarantee 100% accurate answe…
I stopped writing 500-word guardrail prompts. This 8-line template works better. (www.reddit.com) I used to spend hours writing massive, obsessive system prompts for my RAG apps. I’d have ten different refusal examples, "never do X," "always check Y," and a whole paragraph of the model role-playing as a "safe and truthful assistant." I…
↯ Security↯ Hallucination↯ Jailbreakjailbreakhallucinationrag+1
Grok hallucinations (www.reddit.com) Grok is supposedly the lowest-hallucination model according to the AA-Omniscience benchmark. Today I've had INSANE hallucinations from Grok 4.2 fast.
Ran my own benchmark Qwen 3.6 35B vs Gemma 4 26B.... theres a clear winner here (www.reddit.com) Uhh I guess Gemma 4 is so much shittier that it hallucinated this event that happened in china in 1989? According to qwen, nothing of significance happened at Tiananmen square in 1989 - and based on all of the benchmarks of qwen, I believe…
Is anyone else terrified of giving Cursor/Claude direct access to their database? I built an open-source solution. (www.reddit.com) Hey everyone 👋, I absolutely love using Cursor and Claude Desktop for debugging and writing queries, but the idea of hooking them up directly to my database via standard MCP (Model Context Protocol) servers has always given me anxiety. One…
↯ Model Context Protocol↯ Hallucinationmodel-context-protocolhallucinationcursor+1
Stop donating your salary to OpenAI: Why Minimax M2.5 is making GPT-5.2 Thinking look like an overpriced dinosaur for coding plans. (www.reddit.com) ↯ Hallucination↯ Glm↯ Minimax↯ Swe Benchswe-benchminimaxhallucination+5
A guide to setting up your own Hugging Face leaderboard: an end-to-end example with Vectara's hallucination leaderboard (huggingface.co)