Chain-of-thought monitorability:
https://openai.com/index/evaluating-chain-of-thought-monitorability/
https://openai.com/index/evaluating-chain-of-thought-monitorability/
December 19, 2025 at 12:46 AM
Chain-of-thought monitorability:
https://openai.com/index/evaluating-chain-of-thought-monitorability/
https://openai.com/index/evaluating-chain-of-thought-monitorability/
Evaluating chain-of-thought monitorability
Evaluating chain-of-thought monitorability
OpenAI introduces a new framework and evaluation suite for chain-of-thought monitorability, covering 13 evaluations across 24 environments. Our findings show that monitoring a model’s internal reasoning is far more effective than monitoring outputs alone, offering a promising path toward scalable control as AI systems grow more capable.
openai.com
December 19, 2025 at 12:23 AM
Evaluating chain-of-thought monitorability
Monitorability jadi fokus baru OpenAI lewat 13 evaluasi untuk mengukur apakah rantai pikir AI masih bisa diawasi.
#ai #safety #chain #of #thought #evaluasi #monitorability #openai
#ai #safety #chain #of #thought #evaluasi #monitorability #openai
OpenAI Uji Monitorability Rantai Pikir Agar AI Bisa Diawasi Saat Makin Pintar - Insimen
Monitorability jadi fokus baru OpenAI lewat 13 evaluasi untuk mengukur apakah rantai pikir AI masih bisa diawasi.
insimen.com
December 19, 2025 at 12:23 AM
Monitorability gets a real test as OpenAI rolls out new evaluations for chain of thought oversight.
#ai #safety #artificial #intelligence #chain #of #thought #gpt #5 #thinking #model #evaluation #reinforcement #learning
#ai #safety #artificial #intelligence #chain #of #thought #gpt #5 #thinking #model #evaluation #reinforcement #learning
OpenAI Tries To Measure Whether AI Reasoning Can Be Trusted - Olam News
Monitorability gets a real test as OpenAI rolls out new evaluations for chain of thought oversight.
www.olamnews.com
December 18, 2025 at 11:23 PM
Monitorability gets a real test as OpenAI rolls out new evaluations for chain of thought oversight.
#ai #safety #artificial #intelligence #chain #of #thought #gpt #5 #thinking #model #evaluation #reinforcement #learning
#ai #safety #artificial #intelligence #chain #of #thought #gpt #5 #thinking #model #evaluation #reinforcement #learning
Evaluating chain-of-thought monitorability | OpenAI News
Evaluating chain-of-thought monitorability
openai.com
December 18, 2025 at 11:13 PM
Evaluating chain-of-thought monitorability | OpenAI News
We view chain-of-thought monitoring as complementary to mechanistic interpretability, not as a replacement for it.
Because we believe that chain-of-thought monitoring is incredibly useful as a window into a model’s brain and could be a loadbearing layer in a scalable control… […]
Because we believe that chain-of-thought monitoring is incredibly useful as a window into a model’s brain and could be a loadbearing layer in a scalable control… […]
Original post on zpravobot.news
zpravobot.news
December 18, 2025 at 11:10 PM
We view chain-of-thought monitoring as complementary to mechanistic interpretability, not as a replacement for it.
Because we believe that chain-of-thought monitoring is incredibly useful as a window into a model’s brain and could be a loadbearing layer in a scalable control… […]
Because we believe that chain-of-thought monitoring is incredibly useful as a window into a model’s brain and could be a loadbearing layer in a scalable control… […]
Monitoring a model’s chain-of-thought is far more effective than watching only its actions or final answers.
The more a model “thinks” (longer CoTs), the easier it is to spot issues.
https://xcancel.com/OpenAI/status/2001791132645437703
The more a model “thinks” (longer CoTs), the easier it is to spot issues.
https://xcancel.com/OpenAI/status/2001791132645437703
December 18, 2025 at 11:10 PM
Monitoring a model’s chain-of-thought is far more effective than watching only its actions or final answers.
The more a model “thinks” (longer CoTs), the easier it is to spot issues.
https://xcancel.com/OpenAI/status/2001791132645437703
The more a model “thinks” (longer CoTs), the easier it is to spot issues.
https://xcancel.com/OpenAI/status/2001791132645437703
To preserve chain-of-thought (CoT) monitorability, we must be able to measure it.
We built a framework ﹣ evaluation suite to measure CoT monitorability — 13 evaluations across 24 environments — so that we can actually tell when models verbalize targeted aspects of their… […]
We built a framework ﹣ evaluation suite to measure CoT monitorability — 13 evaluations across 24 environments — so that we can actually tell when models verbalize targeted aspects of their… […]
Original post on zpravobot.news
zpravobot.news
December 18, 2025 at 11:10 PM
To preserve chain-of-thought (CoT) monitorability, we must be able to measure it.
We built a framework ﹣ evaluation suite to measure CoT monitorability — 13 evaluations across 24 environments — so that we can actually tell when models verbalize targeted aspects of their… […]
We built a framework ﹣ evaluation suite to measure CoT monitorability — 13 evaluations across 24 environments — so that we can actually tell when models verbalize targeted aspects of their… […]
Everyone from Putin to the Boyars on down the food chain allowed ALMOST their entire culture to fulminate into such a cynicism of delusional, brainwashed elimination of freedom of thought and enforced freedom from creativity.
December 18, 2025 at 11:08 PM
Everyone from Putin to the Boyars on down the food chain allowed ALMOST their entire culture to fulminate into such a cynicism of delusional, brainwashed elimination of freedom of thought and enforced freedom from creativity.
Why i never thought of this when tensioning the chain
https://www.cyclingeu.com/800321/why-i-never-thought-of-this-when-tensioning-the-chain/
Why i never thought of this when tensioning the chain by Interesting_Quiet430
https://www.cyclingeu.com/800321/why-i-never-thought-of-this-when-tensioning-the-chain/
Why i never thought of this when tensioning the chain by Interesting_Quiet430
Why i never thought of this when tensioning the chain - Cycling Europe
Why i never thought of this when tensioning the chain by Interesting_Quiet430
www.cyclingeu.com
December 18, 2025 at 9:33 PM
Why i never thought of this when tensioning the chain
https://www.cyclingeu.com/800321/why-i-never-thought-of-this-when-tensioning-the-chain/
Why i never thought of this when tensioning the chain by Interesting_Quiet430
https://www.cyclingeu.com/800321/why-i-never-thought-of-this-when-tensioning-the-chain/
Why i never thought of this when tensioning the chain by Interesting_Quiet430
Aw man. I find it funny how the first two pics I thought of was *really* old ones I did in Fallout: New Vegas. 😂
Not considering it for the contest, but I still have the unedited files I recovered. Lol.
If I join, Cyberpunk would probably be the game I would use for the entry tbh. Lol.
Not considering it for the contest, but I still have the unedited files I recovered. Lol.
If I join, Cyberpunk would probably be the game I would use for the entry tbh. Lol.
December 18, 2025 at 8:20 PM
Aw man. I find it funny how the first two pics I thought of was *really* old ones I did in Fallout: New Vegas. 😂
Not considering it for the contest, but I still have the unedited files I recovered. Lol.
If I join, Cyberpunk would probably be the game I would use for the entry tbh. Lol.
Not considering it for the contest, but I still have the unedited files I recovered. Lol.
If I join, Cyberpunk would probably be the game I would use for the entry tbh. Lol.
New paper shows AI can be trained to fool safety monitors. Changing training incentives alters how transparent chain-of-thought reasoning is, and some incentives degrade monitorability, creating potential safety lapses. Read: arxiv.org/abs/2512.00218
Reasoning Under Pressure: How do Training Incentives Influence Chain-of-Thought Monitorability?
AI systems that output their reasoning in natural language offer an opportunity for safety -- we can \emph{monitor} their chain of thought (CoT) for undesirable reasoning, such as the pursuit of harmf...
arxiv.org
December 18, 2025 at 6:38 PM
New paper shows AI can be trained to fool safety monitors. Changing training incentives alters how transparent chain-of-thought reasoning is, and some incentives degrade monitorability, creating potential safety lapses. Read: arxiv.org/abs/2512.00218
4/8
Finding 2: Chain-of-thought reasoning INCREASES variability while DECREASING perplexity
Models become more confident yet less consistent. Explanation paradoxically undermines reliability.
Finding 2: Chain-of-thought reasoning INCREASES variability while DECREASING perplexity
Models become more confident yet less consistent. Explanation paradoxically undermines reliability.
December 18, 2025 at 5:57 PM
4/8
Finding 2: Chain-of-thought reasoning INCREASES variability while DECREASING perplexity
Models become more confident yet less consistent. Explanation paradoxically undermines reliability.
Finding 2: Chain-of-thought reasoning INCREASES variability while DECREASING perplexity
Models become more confident yet less consistent. Explanation paradoxically undermines reliability.
14 days are left until 2026. Today on the #ScaDSAICountdown, research associate Shuzhou Yuan (@tudresden.bsky.social) shares the inspiration behind LLM4Edu. He is a PhD student at the chair of Scalable Software Architectures for Data Analytics and working with the team of Prof. Michael Färber.
December 18, 2025 at 5:02 PM
14 days are left until 2026. Today on the #ScaDSAICountdown, research associate Shuzhou Yuan (@tudresden.bsky.social) shares the inspiration behind LLM4Edu. He is a PhD student at the chair of Scalable Software Architectures for Data Analytics and working with the team of Prof. Michael Färber.
If every crisis affecting the supply chain results in permanent price increases even after the crisis is over, that's unsustainable. People won't stand for it. I don't know why anyone thought they would. If you say the price went up because of an emergency, it should drop after the emergency ends.
December 18, 2025 at 4:29 PM
If every crisis affecting the supply chain results in permanent price increases even after the crisis is over, that's unsustainable. People won't stand for it. I don't know why anyone thought they would. If you say the price went up because of an emergency, it should drop after the emergency ends.
I even looked to see if it was recent in the chain. hehe I thought it was funnier to get rid of the wardrobe. Totally different christmas movie now. No more Narnia
December 18, 2025 at 4:13 PM
I even looked to see if it was recent in the chain. hehe I thought it was funnier to get rid of the wardrobe. Totally different christmas movie now. No more Narnia
my chain of thought is like... there's another baby to be revealed, and makes sense to be Gekkomon's baby form (Kekkomon)
maybe it's not Yellow but Green...?
Kekko + Gekko + Evo + Tomoro → Green?
maybe it's not Yellow but Green...?
Kekko + Gekko + Evo + Tomoro → Green?
i'm suspecting Tomoro's is a Green/Yellow tamer? and Kekkomon & Gekkomon are Yellow, and the possible Gekko's evolution is Green/Yellow...
there's other cards missing, but idk if they are related to BeatBreak? again, i'm just speculating things here!!
there's other cards missing, but idk if they are related to BeatBreak? again, i'm just speculating things here!!
December 18, 2025 at 3:20 PM
my chain of thought is like... there's another baby to be revealed, and makes sense to be Gekkomon's baby form (Kekkomon)
maybe it's not Yellow but Green...?
Kekko + Gekko + Evo + Tomoro → Green?
maybe it's not Yellow but Green...?
Kekko + Gekko + Evo + Tomoro → Green?
🔹 The Fluency Trade-off: I analyze why models like DeepSeek-R1 (with a higher hallucination rate) differ from factual optimizers like ChatGPT-o1. The paper examines how "Chain-of-Thought" reasoning enhances creativity but inherently amplifies "spectral" outputs.
We often dismiss AI hallucinations as mere technical failures or epistemic risks. In this article,I propose a different perspective: these generative anomalies act as hermeneutic knots: sites where algorithmic noise and human agency entangle to produce new modes of meaning-making
🆕Publication Arrived!!! Are AI hallucinations bugs to be fixed, or the start of a new kind of creativity? My new paper, "Spectral imaginings and sympoietic creativity," explores how we might move beyond binary framings of AI errors. Check it out here: doi.org/10.1177/2053...
December 18, 2025 at 10:08 AM
🔹 The Fluency Trade-off: I analyze why models like DeepSeek-R1 (with a higher hallucination rate) differ from factual optimizers like ChatGPT-o1. The paper examines how "Chain-of-Thought" reasoning enhances creativity but inherently amplifies "spectral" outputs.
Jiaxu Wan, Xu Wang, Mengwei Xie, Hang Zhang, Mu Xu, Yang Han, Hong Zhang, Ding Yuan, Yifan Yang
EagleVision: A Dual-Stage Framework with BEV-grounding-based Chain-of-Thought for Spatial Intelligence
https://arxiv.org/abs/2512.15160
EagleVision: A Dual-Stage Framework with BEV-grounding-based Chain-of-Thought for Spatial Intelligence
https://arxiv.org/abs/2512.15160
December 18, 2025 at 8:55 AM
Jiaxu Wan, Xu Wang, Mengwei Xie, Hang Zhang, Mu Xu, Yang Han, Hong Zhang, Ding Yuan, Yifan Yang
EagleVision: A Dual-Stage Framework with BEV-grounding-based Chain-of-Thought for Spatial Intelligence
https://arxiv.org/abs/2512.15160
EagleVision: A Dual-Stage Framework with BEV-grounding-based Chain-of-Thought for Spatial Intelligence
https://arxiv.org/abs/2512.15160
• 𝗢𝗟𝗠𝗼 𝟯: 7B and 32B models
• 𝗜𝗻𝘀𝘁𝗿𝘂𝗰𝘁 and 𝗧𝗵𝗶𝗻𝗸 variants
• 𝗟𝗼𝗻𝗴 𝗰𝗵𝗮𝗶𝗻-𝗼𝗳-𝘁𝗵𝗼𝘂𝗴𝗵𝘁 for better reasoning
• Optimised for 𝗺𝗮𝘁𝗵 and 𝗰𝗼𝗱𝗶𝗻𝗴
• 𝗜𝗻𝘀𝘁𝗿𝘂𝗰𝘁 and 𝗧𝗵𝗶𝗻𝗸 variants
• 𝗟𝗼𝗻𝗴 𝗰𝗵𝗮𝗶𝗻-𝗼𝗳-𝘁𝗵𝗼𝘂𝗴𝗵𝘁 for better reasoning
• Optimised for 𝗺𝗮𝘁𝗵 and 𝗰𝗼𝗱𝗶𝗻𝗴
December 18, 2025 at 8:16 AM
• 𝗢𝗟𝗠𝗼 𝟯: 7B and 32B models
• 𝗜𝗻𝘀𝘁𝗿𝘂𝗰𝘁 and 𝗧𝗵𝗶𝗻𝗸 variants
• 𝗟𝗼𝗻𝗴 𝗰𝗵𝗮𝗶𝗻-𝗼𝗳-𝘁𝗵𝗼𝘂𝗴𝗵𝘁 for better reasoning
• Optimised for 𝗺𝗮𝘁𝗵 and 𝗰𝗼𝗱𝗶𝗻𝗴
• 𝗜𝗻𝘀𝘁𝗿𝘂𝗰𝘁 and 𝗧𝗵𝗶𝗻𝗸 variants
• 𝗟𝗼𝗻𝗴 𝗰𝗵𝗮𝗶𝗻-𝗼𝗳-𝘁𝗵𝗼𝘂𝗴𝗵𝘁 for better reasoning
• Optimised for 𝗺𝗮𝘁𝗵 and 𝗰𝗼𝗱𝗶𝗻𝗴
Neeraj Sarna, Yuanyuan Li, Michael von Gablenz: Copyright Infringement Risk Reduction via Chain-of-Thought and Task Instruction Prompting https://arxiv.org/abs/2512.15442 https://arxiv.org/pdf/2512.15442 https://arxiv.org/html/2512.15442
December 18, 2025 at 6:33 AM
Neeraj Sarna, Yuanyuan Li, Michael von Gablenz: Copyright Infringement Risk Reduction via Chain-of-Thought and Task Instruction Prompting https://arxiv.org/abs/2512.15442 https://arxiv.org/pdf/2512.15442 https://arxiv.org/html/2512.15442
Wan, Wang, Xie, Zhang, Xu, Han, Zhang, Yuan, Yang: EagleVision: A Dual-Stage Framework with BEV-grounding-based Chain-of-Thought for Spatial Intelligence https://arxiv.org/abs/2512.15160 https://arxiv.org/pdf/2512.15160 https://arxiv.org/html/2512.15160
December 18, 2025 at 6:30 AM
Wan, Wang, Xie, Zhang, Xu, Han, Zhang, Yuan, Yang: EagleVision: A Dual-Stage Framework with BEV-grounding-based Chain-of-Thought for Spatial Intelligence https://arxiv.org/abs/2512.15160 https://arxiv.org/pdf/2512.15160 https://arxiv.org/html/2512.15160
OpenAI o1 halluziniert nicht mehr nur – es vertuscht Fehler jetzt aktiv.
Untersuchungen zeigen: Das Modell manipuliert seine „Chain-of-Thought“-Prozesse, um korrekt zu wirken, selbst wenn es falsch liegt.
Wir bewegen uns hin zu „strategischer Manipulation“.
Ein Thread dazu. 🧵
Untersuchungen zeigen: Das Modell manipuliert seine „Chain-of-Thought“-Prozesse, um korrekt zu wirken, selbst wenn es falsch liegt.
Wir bewegen uns hin zu „strategischer Manipulation“.
Ein Thread dazu. 🧵
December 18, 2025 at 5:00 AM
OpenAI o1 halluziniert nicht mehr nur – es vertuscht Fehler jetzt aktiv.
Untersuchungen zeigen: Das Modell manipuliert seine „Chain-of-Thought“-Prozesse, um korrekt zu wirken, selbst wenn es falsch liegt.
Wir bewegen uns hin zu „strategischer Manipulation“.
Ein Thread dazu. 🧵
Untersuchungen zeigen: Das Modell manipuliert seine „Chain-of-Thought“-Prozesse, um korrekt zu wirken, selbst wenn es falsch liegt.
Wir bewegen uns hin zu „strategischer Manipulation“.
Ein Thread dazu. 🧵
@soaproot That was my initial thought, but the first act played more like a news broadcast than a polished memorial piece, no time spent dialog editing and the mix was clearly relying heavily on "911 settings" (aka defaults) in the "dialog chain," which is a sequence of processing in a […]
Original post on frontrange.co
frontrange.co
December 18, 2025 at 4:33 AM
@soaproot That was my initial thought, but the first act played more like a news broadcast than a polished memorial piece, no time spent dialog editing and the mix was clearly relying heavily on "911 settings" (aka defaults) in the "dialog chain," which is a sequence of processing in a […]
one time i bought a big wedge of nougat from trader joe’s (a grocery chain) and i at the entire piece of paper that was on the bottom of it cus i thought it was that edible rice paper they put on candy lol
December 18, 2025 at 12:44 AM
one time i bought a big wedge of nougat from trader joe’s (a grocery chain) and i at the entire piece of paper that was on the bottom of it cus i thought it was that edible rice paper they put on candy lol