Image created with Flux Pro v1.1 Ultra. Image prompt: Assembly instruction diagram for an oscilloscope with CRT display assembly, 1980s electronics style, green phosphor and black colors, electronic schematic background, “TECH” in digital display font, probe calibration points shown, control panel layout detailed
You can make LLMs more creative by training them on human “”creativity signals”” (novelty, diversity, surprise, quality). Result: Even small models score higher on all 4 creativity dimensions simultaneously. Looks like we can optimize AI for creativity just like any other metric https://x.com/emollick/status/1927738753285607557
Opus 4 + Claude Code + Claude Max plan = best ROI of any AI coding stack right now”” / X https://x.com/alexalbert__/status/1927410913453203946
Since Claude 4 launch: SWE friend told me he cleared his backlog for the first time ever, another friend shipped a month’s worth of side project work in the past 5 days, and my DMs are full of similar stories. I think it’s undebatable that devs are moving at a different speed”” / X https://x.com/alexalbert__/status/1927803598936887686
On two of the most common tests of creativity (the DAT and the AUT), recent models scored well above the average human in creativity, but not as high as the most creative humans There was lots of variability, but better prompts seem to improve performance https://x.com/emollick/status/1927436892770840883
BOND | BOND https://www.bondcap.com/reports/tai
DeepSeek is aiming for the king: o3 and Gemini 2.5 Pro https://x.com/i/web/status/1928067335014793526
On GPQA Diamond, a set of PhD-level multiple-choice science questions, DeepSeek-R1-0528 scores 76% (±2%), outperforming the previous R1’s 72% (±3%). This is generally competitive with other frontier models, but below Gemini 2.5 Pro’s 84% (±3%). https://x.com/EpochAIResearch/status/1928489527204589680
One reason I think AI development is likely to continue at a rapid pace is that there a growing number of research papers outlining promising directions to take if current approaches to improving AI start to run into barriers. This is an interesting one.”” / X https://x.com/emollick/status/1927373420737560848
Claude 4 Sonnet beating o3-preview on ARC-AGI 2 while being <1/400th of the price https://x.com/scaling01/status/1927418304718623180
VideoGameBench Can Vision-Language Models complete popular video games? best performing model, Gemini 2.5 Pro, completes only 0.48% of VideoGameBench and 1.6% of VideoGameBench Lite https://x.com/_akhaliq/status/1927722717068869750
Web Bench – A new way to compare AI Browser Agents https://blog.skyvern.com/web-bench-a-new-way-to-compare-ai-browser-agents/
Find out more about our open-source interpretability tools, and how to use them on open-weights models, here: https://x.com/AnthropicAI/status/1928119231213605240
🚀 Introducing HunyuanVideo-Avatar, a model jointly developed by Tencent Hunyuan and Tencent Music, bringing photos to life. ✅ Upload a photo + audio — auto-detect scene context & emotion, then generate lifelike speech/singing with dynamic visuals. ✅ Supports multi-style, https://x.com/TencentHunyuan/status/1927575170710974560
🎵 Dream come true for content creators! TIGER AI can extract voice, effects & music from ANY audio file 🤯 This lightweight model uses frequency band-split technology to separate speech like magic. Kudos to @fffiloni for the amazing demo! https://x.com/fdaudens/status/1927455842653102291
OFFICIAL BENCHMARKS OUT – we have a new open source frontier approaching O3 and Gemini 2.5 Pro 🔥🔥🔥 https://x.com/i/web/status/1928054949247693219
OPUS 4 NEW SOTA ON ARC-AGI-2 IT’S HAPPENING – I WAS RIGHT Claude 4 models are the first models that effectively use test-time-compute for ARC-AGI-2 https://x.com/scaling01/status/1927818210331521044
Breaking: Claude Opus 4 jumps to #1 in WebDev Arena! A strong comeback from @AnthropicAI – Opus 4 and Sonnet 4 now on top of the chart, surpassing previous Claude 3.7 and matching Gemini 2.5 Pro. Massive congrats to @AnthropicAI🔥 https://x.com/lmarena_ai/status/1927756554188566803
🚨 Sakana + UBC release Darwin Gödel Machine: AI rewrites own code, boosts SWE-bench from 20% → 50%, Polyglot from 14.2% → 30.7%. Open-ended self-editing agents now real. ⚙️ The Details → The Darwin Gödel Machine (DGM) is a self-improving AI framework where agents evolve by https://x.com/rohanpaul_ai/status/1928309781628080346
I built a platform to make it easier for anyone to land their first AI internship — and I used @lovable_dev to do it. Here’s how InternshipsAI was born and how Lovable helped me bring it to life. 🧵 https://x.com/InternshipsAI/status/1921115383257247997
just used codex’s “”Ask”” functionality to understand the usage of a setting across our entire codebase. very easy to underappreciate the value of AI-enhanced code reading compared to that of code writing https://x.com/gdb/status/1926367007772901508
Our friends @crypticcracking analyzed the latest AI models on a freshly released Modern Sudoku puzzle that “contains some really clever initial logic that might be obvious to many human solvers.” Can the latest AIs spot it and use it to solve the puzzle? https://x.com/SakanaAILabs/status/1927358732339425646
Claude 4 underperforming on Aider Polyglot”” / X https://x.com/scaling01/status/1926795250556666341
Introducing InsightfulPMV2, built with @lovable_dev All-in-One CxO Dashboard – Track Revenue, Cost, Sales Forecast & Cash Flow InsightfulPM V2 is a unified, executive-level dashboard built for CxOs Please upvote this at Lovable Launched: https://x.com/PrafulPujar/status/1922279121092042755
Detecting subtle source code vulnerabilities remains a challenging task. Single language models often struggle with this. This paper proposes VulTrial, a multi-agent framework inspired by a courtroom debate, to improve detecting these difficult vulnerabilities. Methods 🔧: https://x.com/rohanpaul_ai/status/1926949649279041658
Announcing LMEval: An Open Source Framework for Cross-Model Evaluation | Google Open Source Blog https://opensource.googleblog.com/2025/05/announcing-lmeval-an-open-ource-framework-cross-model-evaluation.html
Finally completed and merged the SWE_RL environment that was described by Meta’s SWE RL paper into Atropos – A really difficult environment that can teach a model to be a much better coding agent! Check out the PR: https://x.com/Teknium1/status/1927089897833140647
GUI-explorer: Autonomous Exploration and Mining of Transition-aware Knowledge for GUI Agent https://xieincz.github.io/GUI-explorer.github.io/
LLMs use fixed scheduling, causing long tasks to block short ones inefficiently. ELIS solves this with an iterative length predictor and a Shortest Remaining Time First scheduler, significantly cutting average job completion time. Methods 🔧: → The frontend scheduler accepts https://x.com/rohanpaul_ai/status/1926626806909853961
LLMs using Chain-of-Thought reason discretely, limiting the exploration of reasoning paths. This paper proposes Soft Thinking, a training-free method that uses probability-weighted token embeddings to create soft concept tokens, allowing reasoning in a continuous concept space. https://x.com/rohanpaul_ai/status/1928341102672891987
Very happy to release Codestral Embed, our first code embedder. It can use up to 3072 dimensions, ordered by relevance. Using embeddings of 256 dimensions with int8 precision is already sufficient to outperform all existing models. https://x.com/GuillaumeLample/status/1927736663419007031
On SWE-bench Verified, a benchmark of real-world software engineering tasks, DeepSeek-R1-0528 scores 33% (±2%), competitive with some other strong models but well short of Claude 4. Performance can vary with scaffold; we use a standard scaffold based on SWE-agent. https://x.com/EpochAIResearch/status/1928489533886058934
The ability of o3 to agentically use tools in sequence in its chain-of-thought in the chat interface remains a huge differentiator among AIs. I am not sure o3 is “”smarter”” than Gemini 2.5 or Claude 4 (both do better websites than o3, for example), but you can see how tool use https://x.com/emollick/status/1925705390731165754
Anthropic has released its Claude 4 family of models: Claude Opus 4 and Claude Sonnet 4. We evaluated both models on a suite of benchmarks. The main highlight is a significant improvement in coding performance for Sonnet 4. Results in thread! https://x.com/EpochAIResearch/status/1927813645343305902
Claude 4 Sonnet might be the VERY FIRST model to significantly benefit from test-time-compute on ARC-AGI 2″” / X https://x.com/scaling01/status/1927425665055302023
It’s happening 👀 Anthropic just announced Claude 4 Opus and Claude 4 Sonnet just now at the Code with Claude event Early testers said it could code autonomously for up to seven hours https://x.com/rowancheung/status/1925591664548356555
Fantastic to see Anthropic, in collaboration with @neuronpedia, creating open source tools for studying circuits with transcoders. There’s a lot of interesting work to be done I’m also very glad someone finally found a use for our Gemma Scope transcoders! Credit to @ArthurConmy”” / X https://x.com/NeelNanda5/status/1928169762263122072
Our interpretability team recently released research that traced the thoughts of a large language model. Now we’re open-sourcing the method. Researchers can generate “attribution graphs” like those in our study, and explore them interactively.”” / X https://x.com/i/web/status/1928119229384970244
🚀 We’re open sourcing Chatterbox – our state-of-the-art Voice Cloning model that includes text-to-speech and voice conversion! In recent testing, 63.75% of listeners preferred Chatterbox over ElevenLabs. Not only is it free and open source (MIT license), it’s demonstrably https://www.resemble.ai/chatterbox/
[2505.17862v1] Daily-Omni: Towards Audio-Visual Reasoning with Temporal Alignment across Modalities https://arxiv.org/abs/2505.17862v1
“Sudoku-Bench,” which was announced in March, is a benchmark to encourage AI to acquire more creative reasoning abilities through a multi-step process. https://x.com/i/web/status/1926905826465161629
ARC Prize – Leaderboard https://arcprize.org/leaderboard
Explore all the benchmarks—both Epoch-evaluated and external—in one place, our AI Benchmarking Hub: https://x.com/EpochAIResearch/status/1928498593725399123
Individual results across our independent intelligence evaluations: https://x.com/i/web/status/1928074992065073567
It’s hard to make scheming evals look realistic for LLMs – LessWrong 2.0 viewer https://www.greaterwrong.com/posts/TBk2dbWkg2F7dB3jb/it-s-really-hard-to-make-scheming-evals-look-realistic
Inference providers aren’t sleeping on the switch. https://x.com/fdaudens/status/1927834963509961041
DeepSeek’s R1 leaps over xAI, Meta and Anthropic to be tied as the world’s #2 AI Lab and the undisputed open-weights leader DeepSeek R1 0528 has jumped from 60 to 68 in the Artificial Analysis Intelligence Index, our index of 7 leading evaluations that we run independently https://x.com/i/web/status/1928071179115581671
Love this approach by @RedHat_AI. We need more trust & validation in AI and this can help! https://x.com/ClementDelangue/status/1928551872027116000
Sigh, it’s a bit of a mess. Let me just give you guys the full nuance in one stream of consciousness since I think we’ll continue to get partial interpretations that confuse everyone. All the little things I post need to always be put together in one place. First, I have long”” / X https://x.com/i/web/status/1928148705145934252
Capgemini and SAP partner with Mistral to deploy AI for sensitive sectors | Reuters https://www.reuters.com/business/capgemini-sap-partner-with-mistral-deploy-ai-sensitive-sectors-2025-05-26/
Welcome to the Image Arena: FLUX.1 Kontext! 🖼️ 🎨 You’ll find that FLUX.1 Kontext Pro can generate AND edit images. Congrats to @bfl_ml on this exciting release. 🌲👏🌲 Check it out in the Arena, and get voting! https://x.com/lmarena_ai/status/1928236637709865217
We’re expanding the Epoch AI Benchmarking Hub with four more external benchmarks: VPCT, Fiction-liveBench, GeoBench, and SimpleBench! These benchmarks test visual physics understanding, Geoguessr ability, long-context comprehension, and reasoning and logic skills. 🧵 https://x.com/EpochAIResearch/status/1928498581305971139
1. Visual Physics Comprehension Test (VPCT) by @ChaseBrowe32432 Can models predict where this ball will fall? VPCT, a set of similar problems with varying ramps and obstacles, shows that models struggle with basic physical intuition that humans find trivial. https://x.com/EpochAIResearch/status/1928498584418398522
Some insights we realized during this project: GQA is now the standard for most models, but you can even cut the KV cache size in half by sharing the K & V, increasing arithmetic intensity. Our GTA leverages decoupled RoPE to retain the quality of GQA at half the KV cache size”” / X https://x.com/tri_dao/status/1928170650838995236
FM-Intent: Predicting User Session Intent with Hierarchical Multi-Task Learning | by Netflix Technology Blog | May, 2025 | Netflix TechBlog https://netflixtechblog.com/fm-intent-predicting-user-session-intent-with-hierarchical-multi-task-learning-94c75e18f4b8
We’ve been thinking about what the “”ideal”” architecture should look like in the era where inference is driving AI progress. GTA & GLA are steps in this direction: attention variants tailored for inference: high arithmetic intensity (make GPUs go brr even during decoding), easy to”” / X https://x.com/tri_dao/status/1928170648863473892
🚀 DeepSeek-R1-0528 is here! 🔹 Improved benchmark performance 🔹 Enhanced front-end capabilities 🔹 Reduced hallucinations 🔹 Supports JSON output & function calling ✅ Try it now: https://x.com/i/web/status/1928061589107900779
DeepSeek dropped DeepSeek R1 v2 this morning! We at Hyperbolic Labs now serve DeepSeek-R1-0528, the first inference provider serving this model on @huggingface. My vibe check: It seems to be the only model that consistently answers “”what is 9.9 – 9.11?”” correctly. 🐋 To whale: https://x.com/Yuchenj_UW/status/1927828675837513793
DeepSeek has maintained its status as amongst AI labs leading in frontier AI intelligence https://x.com/i/web/status/1928071183276159117
DeepSeek has released DeepSeek-R1-0528, an updated version of DeepSeek-R1. How does the new model stack up in benchmarks? We ran our own evaluations on a suite of math, science, and coding benchmarks. Full results in thread! https://x.com/EpochAIResearch/status/1928489524616630483
DeepSeek’s R1 update consolidates the lead of 🇨🇳 Chinese AI Labs in open weights intelligence https://x.com/i/web/status/1928226455424528519
Happy to share 💭 Mixture of Thoughts 💭 A curated, general reasoning dataset that trims down over 1M samples from public datasets to ~350k through an extensive set of ablations 🧑🍳 Models trained on this mix match or exceed the performance of DeepSeek’s distilled models — not https://x.com/_lewtun/status/1927043160275923158
Hey guys! We noticed some of you sharing screenshots and links to our DeepSeek-V3-0526 article on @UnslothAI. The link was hidden and wasn’t meant to be shared publicly or taken as a fact but it seems a few of you were scrapping through the site and uncovered it early! 😅 The”” / X https://x.com/danielhanchen/status/1926966742519091327
Now imagine DeepSeek R2 The 150IQ Tsinghua grads are about to outsmart the 120IQ MIT midwits https://x.com/i/web/status/1928073407943319588
Ollama can now think! 🤔🤔🤔 For thinking models, and especially useful for very thoughtful models like DeepSeek-R1-0528, Ollama can separate the thoughts and the response. Thinking can also be disabled! This is useful for getting a direct response. This works across https://x.com/ollama/status/1928543644090249565
Live in Cline: DeepSeek-R1-0528 It’s showing significant benchmark gains, now matching OpenAI o3 in reasoning tasks (key for Plan mode). We’re excited to observe how these improvements impact real-world coding performance in Cline. https://x.com/i/web/status/1928140455923044636
There will be DeepSeek R1 0528 Qwen 3 8B too matching Qwen 3 235B Thinking in performance too 🤯 Whale COOKED! https://x.com/i/web/status/1928058862923391260
DeepSeek R1 05-28 LiveBench results: – 8th in the Overall ahead of o4-mini, Gemini 2.5 Flash Preview and Qwen3-235B-A22B (biggest competitors) – 1st on Data Analysis !!! – 3rd on Reasoning !! – 4th on Mathematics ! – 11th on Language – 20th on Instruction Following – 23rd on https://x.com/i/web/status/1928173385399308639
Deepseek MLA is afaik the first attn variant that can hit compute-bound regime during inference decode, thanks to high arithmetic intensity (~256). If you’re paying $30k for an H100 and only max out the mem bw and not the FLOPS during inference, you’re leaving $20k on the table.”” / X https://x.com/tri_dao/status/1928170652516725027
We made dynamic 1bit quants for DeepSeek-R1-0528 – 74% smaller 713GB to 185GB. Use the magic incantation -ot “”.ffn_.*_exps.=CPU”” to offload MoE layers to RAM, allowing non MoEs to fit < 24GB VRAM on 16K context! The rest sits in RAM & disk. Quants here: https://x.com/danielhanchen/status/1928278088951157116
My conclusion from the last couple of weeks of AI launches is that none of the AI companies can explain very well what their systems do. This is partially because they don’t always know & partially because there is no established approach for documentation of AI capabilities.”” / X https://x.com/emollick/status/1925715197676691615
🐯 Liger GRPO meets TRL https://huggingface.co/blog/liger-grpo
RenderFormer: Transformer-based Neural Rendering of Triangle Meshes with Global Illumination https://microsoft.github.io/renderformer/
introducing: VLM vibe eval 🪭 vision LMs are saturated over benchmarks, so we built vibe eval 💬 > compare different models with refreshed in-the-wild examples in different categories 🤠 > submit your favorite model for eval no numbers — just vibes! https://x.com/mervenoyann/status/1928105481907687891
Multimodal LLMs struggle with complex physics reasoning across intermediate steps and multimodal data. This paper introduces PHYSICSARENA, the first benchmark evaluating models on multimodal physics problems across variable identification, process formulation, and solution https://x.com/rohanpaul_ai/status/1927265658665947443
[2505.16953v1] ICYM2I: The illusion of multimodal informativeness under missingness https://arxiv.org/abs/2505.16953v1
Existing Vision-Language Models fail to effectively analyze complex urban planning maps. PlanGPT-VL tailors Vision-Language Models for urban planning by synthesizing specialized data and using a novel verification method to reduce errors. Methods 🔧: → PlanAnno-V framework https://x.com/rohanpaul_ai/status/1927300101019480531
Exploring Model Graders for Reinforcement Fine-Tuning https://cookbook.openai.com/examples/reinforcement_fine_tuning
Introduced Blocks sections & pages to Launchmvpfast(.)com – Copy, Paste or Install it with CLI in any @reactjs/@nextjs codebase – Open in @v0 Built with @shadcn UI Free & Open Source ⭐️ https://x.com/AliFarooqDev/status/1922411643826270402
How does an LLM writing out this program (WITHOUT a code interpreter running the output) make things more accurate? Verified on Qwen 3 – a30b (below) Lots of interesting takeaways from the Random Rewards paper. NOT that RL is dead, but honestly far more interesting than that! https://x.com/hrishioa/status/1927974614585725353
random rewards only work for Qwen models but not for other models improvements with random rewards were due to clipping, and disappear once clipping is removed Conjecture by authors: “”Under clipping, random rewards don’t teach task quality – instead, they trigger a”” / X https://x.com/scaling01/status/1927424801938825294
Why are almost all RL experiments done on qwen models? Kind of interesting right…”” / X https://x.com/i/web/status/1927948317931000277
Worth thinking about how this paper reflects on every other RL paper using Qwen. If Qwen works with any random reward, how do we know if any of these papers actually does anything”” / X https://x.com/nrehiew_/status/1927424673702121973
[2505.19590] Learning to Reason without External Rewards https://arxiv.org/abs/2505.19590
🚀 Excited to share the most inspiring work I’ve been part of this year: “”Learning to Reason without External Rewards”” TL;DR: We show that LLMs can learn complex reasoning without access to ground-truth answers, simply by optimizing their own internal sense of confidence. 1/n https://x.com/xuandongzhao/status/1927270931874910259
An Operating System for Memory-Augmented Generation in LLMs Lots of great ideas on how to think about memory and better manage it in LLM-based agents. Must read! Here are my notes: https://x.com/omarsar0/status/1928116365640225222
Code LLMs often overlook small changes in problem details, reducing accuracy. This paper introduces a new benchmark, CTF-Code, and instruction tuning framework, CTF-Instruct, to evaluate and enhance this detail sensitivity in Code LLMs. Methods 🔧: → CTF-Code problems use https://x.com/rohanpaul_ai/status/1928402507811725629
Fast-dLLM Training-free Acceleration of Diffusion LLM by Enabling KV Cache and Parallel Decoding https://x.com/_akhaliq/status/1928507150206181613
Forward-only Diffusion Probabilistic Models https://algolzw.github.io/fod/
Github 👨🔧: 一 prompt-optimizer It helps you refine AI prompts to get better results from LLMs. You can use it as a web application or a Chrome browser extension. It supports iterative prompt improvement and allows direct comparison between original and optimized versions for https://x.com/rohanpaul_ai/status/1927357121256956167
Huh. Looks like Plato was right. A new paper shows all language models converge on the same “”universal geometry”” of meaning. Researchers can translate between ANY model’s embeddings without seeing the original text. Implications for philosophy and vector databases alike. https://x.com/emollick/status/1925744569636856072
I once discovered I’d been training networks without any biases for 3 months, because I forgot y += b in my conv layer implementation 🙃 Turns out it didn’t really matter 🤷 although that wasn’t quite as well-established at the time, so it was a bit of a shock to find out!”” / X https://x.com/i/web/status/1926962144253211093
Infinite tool use https://snimu.github.io/2025/05/23/infinite-tool-use.html
LLMs are chmod a+w artifacts yay”” / X https://x.com/karpathy/status/1926135417625010591
LLMs produce factually incorrect information, hallucinations, especially in long text, and existing detection needs external tools. This paper introduces Rationale and Auxiliary Task Enhanced Fine-Tuning, RATE-FT, boosting detection accuracy by training the model on rationales https://x.com/rohanpaul_ai/status/1926996997275373805
LLMs struggle to unlearn specific data points like copyrighted text while preserving general, shared knowledge found across sources. This makes machine unlearning evaluations that assume disjoint datasets unrealistic. This paper introduces the DUSK benchmark to evaluate https://x.com/rohanpaul_ai/status/1926619260576555365
LLMs struggle with complex multi-hop questions due to sparse global rewards. StepSearch trains LLMs with detailed step-wise reinforcement learning rewards to improve external knowledge search, yielding 11.2 percent and 4.2 percent absolute improvement for 3B and 7B models. https://x.com/rohanpaul_ai/status/1928387156197683296
Look what I found in the original DSPy paper. Interesting quote for sure! For the record, I don’t think any *method* can fix this issue — but I believe the right *abstractions* are the most essential step. That’s because a paradigm *is* by definition a set of abstractions. https://x.com/lateinteraction/status/1928430832324161681
Memory Changes Everything – by Nikunj Kothari https://writing.nikunjk.com/p/memory-changes-everything
Missing nuance in the collective realization today: The non-trivial negative result is not that “”RL just amplifies skills that are already there with low probability””. Duh, that’s obvious and not an issue actually. What got questioned today is that “”dumb pretraining teaches the”” / X https://x.com/lateinteraction/status/1927445094002487554
Reasoning Language Models mix languages during their thinking process, which can impact performance and readability. This paper systematically studies this language mixing phenomenon and shows controlling the reasoning script improves accuracy for certain languages. Methods 🔧: https://x.com/rohanpaul_ai/status/1926981898523328866
Refreshing BERT – a groundbreaking shift in NLP BERT, or Bidirectional Encoder Representations from Transformers, was the first to pre-train a deep Transformer in a bidirectional way: ▪️It processes both left and right context of a word at the same time, to deeper understand https://x.com/i/web/status/1928044527870431562
RL from a single example works”” “”RL with random rewards works”” “”Base model pass@256 can match RL model pass@1″” “”RL updates a small % of params”” Recent papers all point in the same direction: RL is mostly just eliciting latent behavior already learned in pretraining, not”” / X https://x.com/corbtt/status/1927428584257261994
Some nice analysis by Nicolas & Francesco of a clear case of emergence — and how to accelerate its acquisition!”” / X https://x.com/i/web/status/1927066495537799502
The Way of Code | Rick Rubin https://www.thewayofcode.com/
This paper proposes Creative Preference Optimization (CRPO), a new alignment method using a large human preference dataset called MUCE, to improve LLM creativity across multiple dimensions simultaneously. Methods 🔧: → CRPO injects signals from novelty, diversity, surprise, https://x.com/rohanpaul_ai/status/1926576601447215169
Tuning instruction-tuned LLMs often reduces performance instead of improving it. This paper introduces Shadow-FT, a new method that tunes the pretrained base model and transfers the learned updates to the instruction-tuned model. This simple process avoids performance drops and https://x.com/rohanpaul_ai/status/1928325751553806702
What I’m trying to say is this conceptual ablation: If X + just about any technique -> huge improvements, for only a limited set of recent base models X, then the discovery is about the class X, not any of the techniques (yet). https://x.com/lateinteraction/status/1927392900632985694
With AI — everyone should have the ability to run their own experiments, build their own hypothesis and understanding of the world, and run iteration loops on this. That’s how we can create new glorious times. We need to scale the scientific method. An army of deep researchers”” / X https://x.com/i/web/status/1928197430891401476
Zochi Publishes A* Paper https://www.intology.ai/blog/zochi-acl
Can SWE-Agents aid in High-Performance Software development? ⚡️🤔 Introducing GSO: A Challenging Code Optimization Benchmark 🔍 Unlike simple bug fixes, this combines algorithmic reasoning with systems programming 📊 Results: Current agents struggle with <5% success rate! https://x.com/StringChaos/status/1928476388274716707
Following our Sudoku-based reasoning benchmark announcement, we’ve been evaluating the latest models to track improvements in their reasoning capabilities. Today, we’re launching the Sudoku-Bench Leaderboard: https://x.com/SakanaAILabs/status/1926798125060002243
LoRAvatar https://starc52.github.io/publications/2025-05-28-LoRAvatar/
Why We Think | Lil’Log https://lilianweng.github.io/posts/2025-05-01-thinking/
ChatHuman https://chathuman.github.io/
DreamDance https://kebii.github.io/DreamDance/
Introducing LisanBench LisanBench is a simple, scalable, and precise benchmark designed to evaluate large language models on knowledge, forward-planning, constraint adherence, memory and attention, and long context reasoning and “”stamina””. “”I see possible futures, all at once. https://x.com/scaling01/status/1928510435164037342
LLMs optimized to fool detectors still exhibit a distinct, detectable writing style. This paper develops a style-aware paraphraser to mimic human styles and evade detection, introducing a metric to measure the human-machine text distribution gap. Methods 🔧: → Train a https://x.com/rohanpaul_ai/status/1926993600480518475
Nemotron-CORTEXA just reached the top of the SWEBench leaderboard for using LLMs to solve software engineering problems, solving 68.2% of SWEBench GitHub issues! It does so by using a multi-step problem localization and repair process, generating multiple proposal candidates https://x.com/ctnzr/status/1927391895879074047
This paper introduces QA-prompting. It first asks the model questions about the article. Then, it uses the article and answers to generate a summary. This highlights key facts, improving accuracy by up to 29% in ROUGE scores. Methods 🔧: → Sample candidate questions and https://x.com/rohanpaul_ai/status/1927350326249578526
SpatialScore https://haoningwu3639.github.io/SpatialScore/
⚒️ Integrate LangSmith prompts with your SDLC You can already test, version and collaborate on prompts in LangSmith. Now, with webhook triggers on prompt changes, you can automatically sync prompts to GitHub, external DBs, or kick off CI. 📓 Docs: https://x.com/LangChainAI/status/1927401850405257283
Discovering smaller models perform just as well while using 200x less energy: 🤯 Sometimes the best AI strategy is just… not using the biggest hammer for every nail. https://x.com/fdaudens/status/1927735676063088889
On why DSPy turns ChatAdapter on by default and only falls back to JSONAdapter on parse failures:”” / X https://x.com/lateinteraction/status/1928233572676042870
