π PapersΓ1 mentions
Decoupled Contrastive Learning
by Yann LeCun
βDecoupled Contrastive Learning _Blog post by Yann LeCun_β
What the smart people are recommending. 8874 books, tools, and products endorsed by the thinkers absorb.md tracks. Ranked by how many times each has been recommended across compiled podcasts, papers, posts, and tweets.
βDecoupled Contrastive Learning _Blog post by Yann LeCun_β
βWe also introduce a pixel-aligned RGB-thermal instruction-tuning dataset and Thermo-VL-Bench, a manually screened RGB-thermal VQA benchmark for low-light and cross-spectrum reasoning.β
βmckinsey put out a study showing investing in clean energy clean economy creates you know twice as many jobs as if you're investing in some of the heavy carbon oil and gas futuresβ
β# Overcoming leakage in scalable quantum error correctionβ
βwe have a new report out today actually that tries to tackle exactly that question and it takes the data really seriously and it finds that there are really serious connections between economic growthβ¦β
βA number of different architectures and loss functions have been applied to the problem of self-supervised learning (SSL), with the goal of developing embeddings that provide the best possible pre-traβ¦β
βYou will need to use the environment variables defined in .env.example to run Next.js AI Chatbot.β
βA closed-loop meta-control layer (Extension A) reduces forgetting by an additional 81% at ~398M, mapping onto the System A and System M roles of Dupoux et al. (arXiv:2603.15381).β
βwe evaluate Claude Code in an agentic proving framework on CLEVER, a Lean 4 benchmark for verifiable code generationβ
βGAIA: a benchmark for general AI assistantsβ
βIn this paper, we propose Rate-In, an algorithm that dynamically adjusts dropout rates during inference by quantifying the information loss induced by dropout in each layer's feature maps.β
βHere, we present a comprehensive demonstration of the color code on a superconducting processor, achieving logical error suppression and performing logical operations. ... This work establishes the coβ¦β
βAbstentionBench is used as the evaluation benchmark, as this work aims to contribute to the field of abstention learning. All datasets on the benchmark were tested against this method and various baseβ¦β
βThe question this paper addresses is: does this mathematical claim hold at the engineering level? To make the answer as general as possible, we deliberately choose the strictest engineering conditionsβ
βso with that uh I'm want to advertise the journal digital Discovery very briefly okay I really want everybody to submit their best work there impact factor of 6.2 Open Access Data reviewerβ
βTraining data comes from approximately 100 expert chain-of-thought (CoT) annotations produced by the BC Protocol (Zou & Xu, 2026b).β
βThis is the structural reason why ~100 CoT examples are sufficient -- not a purely empirical observation like LIMA (Zhou et al., 2023).β
βWe release Llamion, a family of 14B-parameter open-weight language models...We release three checkpoints (Base, Chat, LongChat) that load with trust_remote_code=False in the Hugging Face Transformers β¦β
βthis stock will not be obviously the same without talking about language language models of course they're extremely important we're all using them for one Fe or another so I just want to show you thiβ¦β
βOver Three Decades, Tech Obliterated Media _Blog post by Kara Swisher_ https://nymag.com/intelligencer/article/kara-swisher-burn-book-excerpt-silicon-valley-media.htmlβ
βWe present a quantum simulator comprising 69 superconducting qubits which supports both universal quantum gates and high-fidelity analog evolution, with performance beyond the reach of classical simulβ¦β
βQuantum error correction below the surface code thresholdβ
β# Exploring Prompt Optimization -- Blog post by LangChain https://blog.langchain.dev/exploring-prompt-optimizationβ
β# Demonstrating dynamic surface codesβ
βNoam Brown, who wrote the Diplomacy paper, said this about Sholto: βhe's only been in the field for 1.5 years, but people in AI know that he was one of the most important people behind Gemini's succesβ¦β
βFor example, Anthropic has the constitutional RL paper where they take another language model and they point it and say, βhow helpful or harmless was that response?ββ
βThrough extensive experiments on 32 text-embedding tasks across various architectures (transformers, state-space models) and domains (language, vision), we demonstrate that intermediate layers consistβ¦β
βHere, we present a novel framework for measuring AI usage patterns across the economy. We leverage a recent privacy-preserving system to analyze over four million Claude.ai conversations through the lβ¦β
βactually John Schulman's theme had already published this thing called Web GPT at the time so I was pretty aware of itβ
βMy thoughts on China, export controls and two possible futures https://darioamodei.com/post/on-deepseek-and-export-controlsβ
βwe promote 2PI as a robust framework for connecting microscopic correlations to macroscopic nonequilibrium phenomena.β
βRecent work RAT+, introduces a recurrence-augmented attention backbone that enables flexible dilated attention at inference time.β
βUsing representative methods including Quest, MoBA, and SnapKV, we show that RAT+ consistently improves accuracy over standard attention across sparse budgets on eight needle-in-a-haystack tasks.β
βUsing representative methods including Quest, MoBA, and SnapKV, we show that RAT+ consistently improves accuracy over standard attention across sparse budgets on eight needle-in-a-haystack tasks.β
βAgentic AI: Whatβs Actually Happening (and What Comes Next) https://t.co/XXfAo3BCQdβ
βUsing representative methods including Quest, MoBA, and SnapKV, we show that RAT+ consistently improves accuracy over standard attention across sparse budgets on eight needle-in-a-haystack tasks.β
βWe release all our models to the community.β
βOur teams new work using AI for quantum circuit transpilation Is even better. The team has extended their earlier work on RL-based synthesis of Clifford circuits to general Pauli Networks.β
βthereby closing a substantial gap left by prior work (Erez et al., 2024, 2025), which incurred an additional Ξ(|A|^9) dependence.β
βthe induced model class admits a sharp DEC bound that scales with s and directly yields the optimal rate.β
βDario wrote this blog post called Machines of Loving Grace... We should put a link to the poem. It's quite short.β
βI would just encourage people to read, you know, as much as they can about anthropic. Chris Ola, work on the! I wrote on this thing called the constitution. It's about 80 pages.β
βthe FBI and CISA just published a Medusa Ransomware joint cybersecurity advisory. We'll post it in the comment section below the YouTube stream here for anybody who wants to review it.β
βwe introduce AMA-Bench (Agent Memory with Any length), a benchmark for evaluating long-horizon memory in realistic agentic settings.β
βSo probably like many people here I I read with great interest uh Steve Jay's blog post a couple months ago uh revenge of the junior developer.β
βAnd so here weβre going to move to the second Amadeus, which came out recently called the adolescence of technology. That is about what you see as the most serious A.I. risks.β
βThe construction is based on the quantum intractability of the Dense-Sparse Learning Parity with Noise problem introduced in [DJ25]β
βThis complements the result of [BZZ24] that showed only factoring-based hardness for the 1D case, though achieving a volume versus area entanglement difference.β
βincluding those using probabilistic mixtures [Quantum 7, 1208 (2023)], are independent of the rotation angle ΞΈ, requiring O(log 1/Ξ΄) T gates.β
βthe security of seedless extractors was proven in Quantum 9, 1654 (2025). The core idea was to use the Bell violation of the raw data, rather than its min-entropy, as the extractor promise.β
βWe also generalize a previous construction from [Bunth, MarΓ³ti, Mosonyi, ZimborΓ‘s, Lett. Math. Phys. 113:(7), 2023] to the case of multiple modes per lattice siteβ
βThe data set is based on previously published numerical homogenization results for ensembles of two-phase stochastic microstructures generated by planar Boolean models...see BrΓ€ndel, Brands, Maike, Rhβ¦β
βWe introduce SpecReason, a system that automatically accelerates LRM inference... SpecReason achieves 1.4-3.0Γ speedup over vanilla LRM inference while improving accuracy by 0.4-9.0%.β
βWe introduce Egostream, a diagnostic benchmark for streaming episodic memory evaluation in egocentric vision... Project website, news and updates at: https://saroo25.github.io/Egostream/β
βaround 2016 there was a lot of research along these lines 2015 2016 like nuring machines for instance and the likeβ
βwe introduce AutoLab, a new benchmark for ultra long-horizon closed-loop optimization...We open-source the full benchmark, evaluation harness, and task artifactsβ
βStefan Volk is a professor at the University of Sydney Business School and author of the HBR article, Tapping Into Your Team's Circadian Rhythms. Here's our conversation.β
βWe study how to use such feedback through a distributional variant of the classic imitation learning algorithm DAgger, where the learner has local access to an expert distribution on states visited byβ¦β
βEmpirically, our approach, DistIL, improves over RLVR and RL with self-distillation baselines across a variety of domains: scientific reasoning, coding, and solving hard mathematical problems.β
βNeural Ordinary Differential Equations (NODEs) for continuous time dynamics approximationβ
βwe propose MPVI: Motion Planner / VLA Interleaving, a framework that integrates model-based motion planning with VLAs to improve robustness without further training.β
βwe introduce \ours, a closed-loop framework for memory augmentation in SE agents... Our project page: https://xhguo7.github.io/MemOp/β
βAperiodic controls should become standard for interpretable physiological time-series deep learning.β
βThis paper introduces a categorical framework for deep learning models that formalizes broadcasting through the novel axis-stride and array-broadcasted categories.β
βthey published an algorithm called r10 about how to bootstrap a model like R1 and like right now like a bunch of people a bunch of like solo hackers are like reproducing the R1 resultsβ
βhow to get rich remember that one how to get rich that like one theads and super Yeah Bang Yeah Yeah I think that is still the most viral threat ever on Twitterβ
βrecent work [PRX QUANTUM 4, 040337 (2023)] extended classical shadows from quantum state to quantum process, enabling efficient single-channel E property predictionβ
βIn this paper, we propose MultiToP, a multimodal-context-aware visual token patching framework that mitigates hallucinations by refining unreliable visual tokens before language generation.β
βyou're probably referring to a paper that Jeff Hinton and I published in nature in 2015... it was basically a bit of a Manifesto if you want or a review paper to tell the wide community of Science andβ¦β
βthere is this architecture which really should be called a macro architecture called jepa... it is not an alternative to Transformers you can have Transformers Inside of japas... jepa is an answer to β¦β
βInspired by these mechanisms, we propose MemoryVLA++, a full temporal modeling framework that equips VLA models with memory and imagination for robotic manipulation.β
βWe investigate the emergence of intuitive physics understanding in general-purpose deep neural network models trained to predict masked regions in natural videos.β
βWe consider the following quasilinear reaction-diffusion equation $$ \partial_tu=Ξu^m+(1+|x|)^Οu^p, \quad (x,t)\in\mathbb{R}^N\times(0,\infty), $$ in dimension $N\geq3$ and in the range of exponents $β¦β
βThis work demonstrates that Transformers without normalization can achieve the same or better performance using a remarkably simple technique.β
βSo when is the Twitter Files, right, which Elon put out with the set of journalists when he took over... you can just read it and you should if you haven't.β
βFacebook produced a lot of documents into that investigation and many of those have now been made public and you can download those reports. And there's like 2,000 pages worth of material.β
βIntegrated modeling of ARC H-modes has been conducted to assess the feasibility of high-performance scenarios compatible with divertor detachment.β
βThe well-known Stein--Tomas restriction theorem gives the sharp range of $p$ for which $L^p\to L^2$ restriction estimates hold for the surface measure on the sphere.β
βUsing the CESM2 Large Ensemble over the western United States, we compare three training-year selection strategies under fixed data budgetsβ
βWe first introduce two datasets, Ex-HateMM and Ex-ImpliHateVid, for explainable hateful video detection.β
βWe first introduce two datasets, Ex-HateMM and Ex-ImpliHateVid, for explainable hateful video detection.β
βwe introduce \textbf{MAVIS}, a novel multi-agent framework that rethinks retrieval as cooperative reasoning rather than brute-force search.β
βwith the first with with the rag paper I believe was released in 2020 since then we've been trying to build some kind of rag systemβ
βwe utilize called something called fusion and decoder we're big fan fusion decoder that actually the paper came from meta our implementation that help us retrieve informationβ
βthere's this paper about actually transformers in 2019 the decoder encoder model T5 looks amazing let's start buildingβ
βAutomatic Question Difficulty Estimation (AQDE) holds growing promise for educational assessment because it has the potential to yield difficulty estimates that are competitive with expert judgmentβ
βThe first International AI Safety Report comprehensively synthesizes the current evidence on the capabilities, risks, and safety of advanced AI systems.β
βIn this paper, we propose a method called Capability-Aligned Hierarchical Learning (CAHL), which leverages RLVR to jointly optimize both policies, enabling better alignment between the high-level planβ¦β
βwe introduce PhysTool-Bench, the first physical tool-use benchmark designed to evaluate MLLMs' ability to comprehend real-world scenarios, identify physical tools, and plan their useβ
βthere's a really good paper uh or there was a really good paper two years ago out of Stanford uh called uh I think generative agents or something like that... That's a really really good one to go andβ¦β
βExperiments on general and multi-hop QA benchmarks show that DAC, implemented via parameter-efficient LoRA modules over a shared backbone, achieves strong performance against prior baselines that relyβ¦β
βWe introduce JANUS, a benchmark for measuring goal-conditioned pragmatic distortion in fact-grounded LLM outputs.β
βThe paper is the theory of that off-diagonal.β
βA multi-similarity Siamese neural network unifies zero-day anomaly detection and one-shot classification in optical networks, achieving over 99% accuracy and instant adaptability across lightpaths andβ¦β
βWe introduce EngVQA, a multimodal benchmark for evaluating engineering reasoning across 5 engineering subjects containing 696 problems.β
βWe propose the Customized Concept Embedding Diffusion (CCE-Diffusion) framework.β
βthis paper introduces Role-Agent, \textcolor{black}{a framework} that harnesses a single LLM to function concurrently as both the agent and the environment, enabling a bootstrapped co-evolution.β
βThis report extends the v1 technical report (arXiv:2605.28062).β
βWe propose SDBN (Small Data Big Noise), a unified framework that brings adversarial training to PEFTβ
βWe also introduce two variants of the method that use discrete uncertainty sets: SDBN-h, which enumerates character-level edits and selects worst-case variants using gradientsβ