7/
Finding 1: simplification. BCSS programs explaining human and LRM rules tend to be simpler than the target rule. For LRMs, more reasoning effort means less simplification.
Finding 1: simplification. BCSS programs explaining human and LRM rules tend to be simpler than the target rule. For LRMs, more reasoning effort means less simplification.
November 9, 2026 at 5:39 PM
7/
Finding 1: simplification. BCSS programs explaining human and LRM rules tend to be simpler than the target rule. For LRMs, more reasoning effort means less simplification.
Finding 1: simplification. BCSS programs explaining human and LRM rules tend to be simpler than the target rule. For LRMs, more reasoning effort means less simplification.
The 3K went the other way; reduced the LRMs to 10 racks, which has more reloads per ton, hence giving it more endurance in exchange for immediate damage. Retains the SRM-6s and the inherent ammo problems, but now with 16 sinks, it can fight effectively.
November 10, 2026 at 2:00 PM
The 3K went the other way; reduced the LRMs to 10 racks, which has more reloads per ton, hence giving it more endurance in exchange for immediate damage. Retains the SRM-6s and the inherent ammo problems, but now with 16 sinks, it can fight effectively.
I'm always trying out different stuff, but I'm very much a Hunchback pilot. I love my 3052 nissan altima.
Other faves are the Awesome, & Bushwhacker (RAC-5), and Crab, and been having fun with a modernized Mackie withb a shit ton of hot-loaded LRMs and clan medium+large lasers
Other faves are the Awesome, & Bushwhacker (RAC-5), and Crab, and been having fun with a modernized Mackie withb a shit ton of hot-loaded LRMs and clan medium+large lasers
November 12, 2026 at 5:47 AM
I'm always trying out different stuff, but I'm very much a Hunchback pilot. I love my 3052 nissan altima.
Other faves are the Awesome, & Bushwhacker (RAC-5), and Crab, and been having fun with a modernized Mackie withb a shit ton of hot-loaded LRMs and clan medium+large lasers
Other faves are the Awesome, & Bushwhacker (RAC-5), and Crab, and been having fun with a modernized Mackie withb a shit ton of hot-loaded LRMs and clan medium+large lasers
10/
Finding 4: people vs. machines. LRMs err on different tasks than people, and classifiers trained on program-level features can discriminate between them. The stronger the LRM, the more human-like, even though the best LRM outperforms people on average.
Finding 4: people vs. machines. LRMs err on different tasks than people, and classifiers trained on program-level features can discriminate between them. The stronger the LRM, the more human-like, even though the best LRM outperforms people on average.
November 9, 2026 at 5:39 PM
10/
Finding 4: people vs. machines. LRMs err on different tasks than people, and classifiers trained on program-level features can discriminate between them. The stronger the LRM, the more human-like, even though the best LRM outperforms people on average.
Finding 4: people vs. machines. LRMs err on different tasks than people, and classifiers trained on program-level features can discriminate between them. The stronger the LRM, the more human-like, even though the best LRM outperforms people on average.
4/
What do these programs reveal? In both people and LRMs, inferred rules a) are oversimplified, b) only explain part of the evidence, and c) use incorrect feature dependencies. Further, we find that reasoning effort predicts the extent of simplification and partial explanation.
What do these programs reveal? In both people and LRMs, inferred rules a) are oversimplified, b) only explain part of the evidence, and c) use incorrect feature dependencies. Further, we find that reasoning effort predicts the extent of simplification and partial explanation.
November 9, 2026 at 5:39 PM
4/
What do these programs reveal? In both people and LRMs, inferred rules a) are oversimplified, b) only explain part of the evidence, and c) use incorrect feature dependencies. Further, we find that reasoning effort predicts the extent of simplification and partial explanation.
What do these programs reveal? In both people and LRMs, inferred rules a) are oversimplified, b) only explain part of the evidence, and c) use incorrect feature dependencies. Further, we find that reasoning effort predicts the extent of simplification and partial explanation.
9/
Finding 3: wrong dependencies. We built a taxonomy of error types and labeled errors with an LLM judge conditioned on the program, task and response. Both people and LRMs mostly misinfer which problem features matter (attribute misselection errors). Weaker models produce more “ad-hoc” errors.
Finding 3: wrong dependencies. We built a taxonomy of error types and labeled errors with an LLM judge conditioned on the program, task and response. Both people and LRMs mostly misinfer which problem features matter (attribute misselection errors). Weaker models produce more “ad-hoc” errors.
November 9, 2026 at 5:39 PM
9/
Finding 3: wrong dependencies. We built a taxonomy of error types and labeled errors with an LLM judge conditioned on the program, task and response. Both people and LRMs mostly misinfer which problem features matter (attribute misselection errors). Weaker models produce more “ad-hoc” errors.
Finding 3: wrong dependencies. We built a taxonomy of error types and labeled errors with an LLM judge conditioned on the program, task and response. Both people and LRMs mostly misinfer which problem features matter (attribute misselection errors). Weaker models produce more “ad-hoc” errors.
Xingyu Liu, Xiangyang Yin, Tianhua Xia, Haiyu Wang, Sai Qian Zhang
CodeQuant2: Enhanced Outlier Mitigation and Adaptive Clustering for Low-Precision LRMs
https://arxiv.org/abs/2611.09868
CodeQuant2: Enhanced Outlier Mitigation and Adaptive Clustering for Low-Precision LRMs
https://arxiv.org/abs/2611.09868
November 10, 2026 at 6:30 AM
Xingyu Liu, Xiangyang Yin, Tianhua Xia, Haiyu Wang, Sai Qian Zhang
CodeQuant2: Enhanced Outlier Mitigation and Adaptive Clustering for Low-Precision LRMs
https://arxiv.org/abs/2611.09868
CodeQuant2: Enhanced Outlier Mitigation and Adaptive Clustering for Low-Precision LRMs
https://arxiv.org/abs/2611.09868
SepLRM: A Foundation Model Paradigm for Recommendation
ByteDance splits large recommendation models into one shared model that encodes long user histories once and exports a KV cache, plus vertical models that query it.
📝 arxiv.org/abs/2611.10356
ByteDance splits large recommendation models into one shared model that encodes long user histories once and exports a KV cache, plus vertical models that query it.
📝 arxiv.org/abs/2611.10356
SepLRM: A Foundation Model Paradigm for Recommendation
Large recommendation models (LRMs) improve recommendation performance by scaling behavior sequence length, model capacity, and training compute. However, the return on investment (ROI) of continued sc...
arxiv.org
November 11, 2026 at 12:58 AM
SepLRM: A Foundation Model Paradigm for Recommendation
ByteDance splits large recommendation models into one shared model that encodes long user histories once and exports a KV cache, plus vertical models that query it.
📝 arxiv.org/abs/2611.10356
ByteDance splits large recommendation models into one shared model that encodes long user histories once and exports a KV cache, plus vertical models that query it.
📝 arxiv.org/abs/2611.10356
Report sadly claiming non-deterministic string prediction like in recent LLMs/LRMs as AI; NOT mentioning at least batch (in)variance or floating point problems or other arguments against use in serious contexts: work, health or even security. But at least they put more focus on nongenerative stuff
The One Hundred Year Study on Artificial Intelligence (#AI100) is a long-term initiative administered by the Stanford Institute for Human-Centered Artificial Intelligence (HAI). Every five years, an international panel of experts examines how artificial intelligence is evolving and shaping society.
AI100 | One Hundred Year Study on Artificial Intelligence
ai100.stanford.edu
November 10, 2026 at 3:15 PM
Report sadly claiming non-deterministic string prediction like in recent LLMs/LRMs as AI; NOT mentioning at least batch (in)variance or floating point problems or other arguments against use in serious contexts: work, health or even security. But at least they put more focus on nongenerative stuff
Xingyu Liu, Xiangyang Yin, Tianhua Xia, Haiyu Wang, Sai Qian Zhang: CodeQuant2: Enhanced Outlier Mitigation and Adaptive Clustering for Low-Precision LRMs https://arxiv.org/abs/2611.09868 https://arxiv.org/pdf/2611.09868 https://arxiv.org/html/2611.09868
November 9, 2026 at 6:40 AM
Xingyu Liu, Xiangyang Yin, Tianhua Xia, Haiyu Wang, Sai Qian Zhang: CodeQuant2: Enhanced Outlier Mitigation and Adaptive Clustering for Low-Precision LRMs https://arxiv.org/abs/2611.09868 https://arxiv.org/pdf/2611.09868 https://arxiv.org/html/2611.09868
sitting in the cuck corner slinging LRMs while the tbolt point blank decapitates a corsair with an AC/20
November 3, 2026 at 2:28 AM
sitting in the cuck corner slinging LRMs while the tbolt point blank decapitates a corsair with an AC/20
ConceptARC at Scale: Studying Conceptual Understanding in Reasoning Models
Philipp Mondorf et al.
#arXiv #cs.AI #cs.LG
Philipp Mondorf et al.
#arXiv #cs.AI #cs.LG
ConceptARC at Scale: Studying Conceptual Understanding in Reasoning Models
Large reasoning models (LRMs) perform well on challenging benchmarks, yet it remains unclear whether they form generalizable abstractions across conceptually related reasoning tasks. The Abstraction and Reasoning Corpus (ARC) provides a controlled testbed for studying abstraction, requiring solvers…
arxiv.org
November 5, 2026 at 10:38 AM
SA VeR2 improves Large Reasoning Models (LRMs) safety with fine-grained verifiable rewards, reducing risks while enhancing reasoning capabilities. https://arxiv.org/abs/2611.07482
SAVeR$^2$: Reasoning-based Safety Alignment for Large Reasoning Models via Verifiable Rewards
ArXiv link for SAVeR$^2$: Reasoning-based Safety Alignment for Large Reasoning Models via Verifiable Rewards
arxiv.org
November 7, 2026 at 2:00 AM
SA VeR2 improves Large Reasoning Models (LRMs) safety with fine-grained verifiable rewards, reducing risks while enhancing reasoning capabilities. https://arxiv.org/abs/2611.07482
ConceptARC at Scale: Studying Conceptual Understanding in Reasoning Models
Philipp Mondorf et al.
#arXiv #cs.AI #cs.LG
Philipp Mondorf et al.
#arXiv #cs.AI #cs.LG
ConceptARC at Scale: Studying Conceptual Understanding in Reasoning Models
Large reasoning models (LRMs) perform well on challenging benchmarks, yet it remains unclear whether they form generalizable abstractions across conceptually related reasoning tasks. The Abstraction and Reasoning Corpus (ARC) provides a controlled testbed for studying abstraction, requiring solvers…
arxiv.org
November 5, 2026 at 8:16 AM
I think it's a valkyrie. Medium lasers in the right arm and LRMs in the left torso
November 2, 2026 at 2:01 AM
I think it's a valkyrie. Medium lasers in the right arm and LRMs in the left torso
rip out the LRMs and you could fit so much cosy in the launcher boxes
October 28, 2026 at 6:51 AM
rip out the LRMs and you could fit so much cosy in the launcher boxes
word to the wise
new bind for firegroup 6 on Shift/nearby (LRMs, or 'sustained' weapons IE MG/flamer)
ERPPC on 1 and 2 (left/right click)
MPL on 3, chainfire (mouse forward)
SRM on 4 chainfire (mouse back)
peeking & heat ceiling expands.
in the opinion of this mind it is written.
new bind for firegroup 6 on Shift/nearby (LRMs, or 'sustained' weapons IE MG/flamer)
ERPPC on 1 and 2 (left/right click)
MPL on 3, chainfire (mouse forward)
SRM on 4 chainfire (mouse back)
peeking & heat ceiling expands.
in the opinion of this mind it is written.
October 29, 2026 at 6:18 PM
word to the wise
new bind for firegroup 6 on Shift/nearby (LRMs, or 'sustained' weapons IE MG/flamer)
ERPPC on 1 and 2 (left/right click)
MPL on 3, chainfire (mouse forward)
SRM on 4 chainfire (mouse back)
peeking & heat ceiling expands.
in the opinion of this mind it is written.
new bind for firegroup 6 on Shift/nearby (LRMs, or 'sustained' weapons IE MG/flamer)
ERPPC on 1 and 2 (left/right click)
MPL on 3, chainfire (mouse forward)
SRM on 4 chainfire (mouse back)
peeking & heat ceiling expands.
in the opinion of this mind it is written.
The Illusion of Thinking: Models Don't Actually Reason
Discussion | hackernews | Author: CGMthrowaway
#MachineLearning
Discussion | hackernews | Author: CGMthrowaway
#MachineLearning
The Illusion of Thinking: Models Don't Actually Reason
Recent generations of language models have introduced Large Reasoning Models (LRMs) that generate detailed thinking processes before providing answers. While these models demonstrate improved performance on reasoning benchmarks, their fundamental capabilities, scaling properties, and limitations rem
arxiv.org
October 30, 2026 at 12:30 AM
The Illusion of Thinking: Models Don't Actually Reason
Discussion | hackernews | Author: CGMthrowaway
#MachineLearning
Discussion | hackernews | Author: CGMthrowaway
#MachineLearning
I'm waiting till Swe and Fin fighter jets wipe that scumbag off the face of the earth with LRMs.
October 27, 2026 at 8:29 PM
I'm waiting till Swe and Fin fighter jets wipe that scumbag off the face of the earth with LRMs.
Valkyries offense is based on the 3 pt shot anyhow, so LRMs are kinda fitting.
October 25, 2026 at 12:22 AM
Valkyries offense is based on the 3 pt shot anyhow, so LRMs are kinda fitting.
The secret bonus, less lrms builds means less AMS, less ECM, making my bullshit more effective
October 24, 2026 at 7:56 PM
The secret bonus, less lrms builds means less AMS, less ECM, making my bullshit more effective
Pgi be like
"We nerfed lrms"
Oooohh you nerfed ghost heat? Chain fire turbo clicking active.
You opened the spread? Her name is Madness and she's got NINE LRM5S. Can't spread that enough to matter.
"We nerfed lrms"
Oooohh you nerfed ghost heat? Chain fire turbo clicking active.
You opened the spread? Her name is Madness and she's got NINE LRM5S. Can't spread that enough to matter.
Keegan Michael Key: I Took That Personally
ALT: Keegan Michael Key: I Took That Personally
static.klipy.com
October 24, 2026 at 7:55 PM
Pgi be like
"We nerfed lrms"
Oooohh you nerfed ghost heat? Chain fire turbo clicking active.
You opened the spread? Her name is Madness and she's got NINE LRM5S. Can't spread that enough to matter.
"We nerfed lrms"
Oooohh you nerfed ghost heat? Chain fire turbo clicking active.
You opened the spread? Her name is Madness and she's got NINE LRM5S. Can't spread that enough to matter.
The Vulture will target that soft spot with LRMs from behind a hill somewhere
October 21, 2026 at 9:41 PM
The Vulture will target that soft spot with LRMs from behind a hill somewhere