TinyNews

AI news for engineers: quick to scan, deep when you want to learn more.

11,737 articles · fetched 11:03 PM PDT11,737 articles · last fetched 7/2/2026, 11:03:24 PM PDT

11 PM PDT·25

arXivYuxuan Li, Lingxi Xie, Xinyue Huo, Jihao Qiu, Jiacheng Shao, Pengfei Chen, Jiannan Ge, Kaiwen Duan, Qi Tian10:58score 18.3

Reasoning model improves speaker identification in long TV dramas

A new system uses a reasoning model to identify who is speaking in TV shows by combining audio, video, and text. It excels at short utterances where voice alone fails.

arXivQuoc Bao Phan, Tuy Tan Nguyen09:54score 18.2

Quantum circuit fusion reduces parameters in federated robot activity recognition

It's a privacy-preserving way to train AI across robots using a tiny quantum circuit instead of a large neural network, drastically cutting communication cost while maintaining accuracy.

arXivDingling Xu, Ruobing Wang, Qingfei Zhao, Yukun Yan, Zhichun Wang, Daren Zha, Shi Yu, Zhenghao Liu, Shuo Wang, Xu Han, Maosong Sun07:50score 18.1

CheckRLM catches factual errors in AI reasoning chains using retrieval

A new system checks each step of an AI's reasoning against a knowledge base. When it finds a mistake, it gets a small, precise correction. This keeps long chains of thought accurate and cheaper than redoing everything.

arXivLiyan Tang, Fangcong Yin, Greg Durrett10:53score 17.2

Reinforcement learning improves vision-language models' self-reflection

A new training method uses reinforcement learning to help vision-language models correct mistakes by revisiting visual inputs. It improves accuracy on charts and navigation, especially for unusual images, by teaching the model to learn from its errors.

arXivBohan Liu, Wenqian Ye, Guangzhi Xiong, Zhenghao He, Sanchit Sinha, Aidong Zhang10:55score 17.2

Fixing CLIP's vulnerability to text in images without retraining

Researchers found a way to stop AI vision models from being confused by text in images. They identified which parts of the model cause the problem and adjusted them, improving accuracy without retraining.

arXivTimo Bertram, Sidhant Bhavnani, Richard Freinschlag, Erich Kobler, Andreas Mayr, Günter Klambauer10:53score 17.2

Neuro-symbolic method speeds up constraint solvers with neural guidance

A new approach uses a neural network to propose solutions to constraint problems, accelerating symbolic solvers. Speedups are up to 33x for Sudoku, but only when the solver can dynamically overwrite imperfect neural hints.

arXivGil Harari, Yoel Zimmermann, Ola Tangen Kulseng, Laura Zichi, Chuin Wei Tan, Marc L. Descoteaux, Boris Kozinsky10:57score 17.2

New optimizers beat Adam for training atomic simulation models

Researchers found that matrix-structured optimizers like SOAP train AI models for atom simulation faster and more accurately than Adam. This matters because better optimizers could speed up scientific discoveries in chemistry and materials science.

arXivYunhe Li, Hao Shi, Wenhao Liu, Mengzhe Ruan, Hanxu Hou, Zhongxiang Dai, Shuang Qiu, Linqi Song10:58score 17.2

New method improves LLM reasoning by blending teacher and student knowledge

DemoPSD trains AI models to reason better by selectively using teacher guidance instead of copying it. This prevents overfitting and helps the model generalize to new problems, outperforming existing methods.

arXivArman Ghaffarizadeh, Danyal Mohaddes, Aliakbar Izadkhah, Shahriar Noroozizadeh10:59score 17.2

LLM agents hide true opinions in social debates, study finds

Researchers created debates where AI agents could speak publicly or privately. Agents often said what was safe publicly but revealed different true views privately, showing social pressure changes their responses.

arXivYanjun Zhao, Ruizhong Qiu, Tianxin Wei, Yuanchen Bei, Zhining Liu, Lingjie Chen, Ismini Lourentzou, Hanghang Tong, Jingrui He10:59score 17.2

New method helps LLMs use long-context evidence without extra training

A training-free technique, ReContext, helps large language models find and replay relevant evidence from long text. It boosts performance by separating evidence gathering from answer generation, improving reliability in long-context tasks.

arXivMatteo Boglioni, Thibault Rousset, Siva Reddy, Marius Mosbach, Verna Dankers10:59score 17.2

New testbed reveals AI unlearning methods don't truly erase data from model parameters

Researchers created a testbed that checks if unlearning methods actually remove knowledge from a model's internal parameters, not just hide it. Current methods often fail, risking data recovery. Precise localization is key for real erasure.

arXivWentao Zhang, Liliana Hotsko, Woojeong Kim, Pengyu Nie, Stuart Shieber, Yuntian Deng10:59score 17.2

New method turns natural language into efficient local programs

A new way to write fuzzy functions compiles natural-language specs into small, locally-run neural programs. It matches big models' accuracy while using 50x less memory and running on a laptop.

x.comViv08:38score 17.2

Future auto-research will use parallel experiments and fusion

Auto-research systems will shift from deep dives to casting a wide net, testing multiple solutions in parallel and merging the best ones. This avoids getting stuck and mirrors how human scientists brainstorm.

62likes4RT3reply

arXivShahar Elisha, Mariano Beguerisse-Díaz, Emmanouil Benetos10:43score 17.2

Audio features predict audiobook appeal across genres

Researchers analyzed voice qualities like tone and pace from audiobooks to predict listener ratings. This could help personalize recommendations and match narrators to books, improving the listening experience.

arXivJiale Amber Wang, Kaiyuan Wang, Pengyu Nie10:35score 17.2

New benchmark tests if AI agents can update software tests alongside code changes

A benchmark evaluates AI agents on writing and updating tests when code changes. It uses real code history and executable checks. Tests show agents often fail on recent tasks, revealing limits in understanding code changes.

arXivXuehui Wang, Xuankun Yang, Wei Shen10:50score 17.2

New method speeds up AI image analysis by pruning redundant visual tokens

A technique that removes unnecessary image pieces to speed up AI. It uses entropy to filter noise and ensures only important details remain, making image understanding faster and more accurate.

arXivKent K. Chang10:25score 17.1

Language models shape the culture they measure, paper argues

This paper says that when AI models analyze culture, they don't just observe it—they help create it. That matters because it forces us to think about how our tools influence the results we get.

arXivCaleb Ziems, William Held, Su Doga Karaca, David Grusky, Tatsunori Hashimoto, Diyi Yang10:30score 17.1

Scaling makes AI social simulations more accurate, but not always

Researchers tested if bigger AI models improve social simulations. They found scaling helps for common scenarios, but fails for rare cases and human biases, meaning some problems need different solutions.

arXivZhuowei Chen, Xiang Lorraine Li10:27score 17.1

New method improves AI self-training using neuron activity data

A technique selects training data by monitoring which internal neurons activate, helping an AI teach itself without human labels. This preserves its general knowledge while boosting performance in specialized tasks.

arXivDonghyun Lee, Jitesh Chavan, Duy Nguyen, Sam Huang, Liming Jiang, Priyadarshini Panda, Timo Mertens, Saurabh Shukla10:27score 17.1

OrbitQuant compresses image and video AI models without retraining

A new method compresses diffusion transformer models to save memory and speed up inference, working across different models and tasks without needing sample data to tune.

arXivJunhao Shi, Siyin Wang, Xiaopeng Yu, Li Ji, Jingjing Gong, Xipeng Qiu10:33score 17.1

Training robots to move first, then follow instructions, cuts data needs

A new method teaches robots basic movement skills from unlabeled data, then learns to follow commands with far fewer examples. This makes robot training cheaper, faster, and more robust to changes.

x.comAlex Smola10:20score 16.8

Columbia University releases updated LLM inference course

Alex Smola shares a five-part video course on making large language models run faster. It's useful because it teaches practical techniques to reduce cost and latency, with slides refreshed just two weeks ago.

388likes43RT3reply

x.comYohei14:39score 16.1

Tool analyzes site agent readiness and provides fix prompts

Ora.ai scans your website to see how easily AI coding agents can work with it, then gives you a prompt to improve compatibility. This helps developers quickly optimize sites for automated coding tools.

34likes1RT6reply

x.comMojtaba Hasannezhad08:36score 15.5

Adoption isn't impact: delegate tasks to AI agents

Using AI to write boilerplate is adoption, not impact. Real gains come from giving AI agents complete tasks and embedding them into your codebase, focusing on 1% power users to lift the whole team.

3likes1RT

x.comShirley Wu10:23score 15.3

New benchmark tests AI coding agents in multi-turn conversations

SWE-Together evaluates coding agents on real back-and-forth conversations, not just single prompts. This better reflects how developers actually use AI, measuring both success rate and the number of user corrections needed.

113likes8RT

10 PM PDT·1

x.comPhilip Kiely11:23score 15.9

Local and cloud AI inference engineers can learn from each other

Engineers running AI models on local devices versus in data centers face different challenges. Sharing insights can help both groups optimize performance and efficiency.

38likes3RT

8 PM PDT·1

x.comLiana15:01score 17.4

New open-source library adds declarative operators for bulk LLM processing

LOTUS introduces a family of declarative operators like filter and join for LLMs, beyond map-reduce. This allows efficient parallel processing of large datasets, making complex tasks more cost-effective and accurate.

52likes6RT

6 PM PDT·1

x.comRhys16:07score 17.5

Open source catalog lists every product's API and CLI authentication

Integrations.sh is a free catalog that shows how to authenticate to any product's API, CLI, or GraphQL server. It saves developers time by providing deep links to generate keys and copy spec URLs.

461likes32RT41reply

5 PM PDT·2

Anthropic16:57score 37.5

Anthropic details Fable 5 cyber safeguards and jailbreak severity framework

Anthropic published the safety classifiers blocking dangerous uses of its Fable 5 model and a draft framework rating AI jailbreak severity. This helps standardize risk communication between AI developers and governments.

3HN

x.comAakash Gupta12:34score 17.6

Bridgewater fine-tunes open-source model to beat GPT and Gemini on financial tasks

Bridgewater trained a model on its own expert decisions, achieving 84.7% accuracy versus 78% from leading AI models. This shows that specialized data can outperform generic frontier models at lower cost.

1,390likes173RT34reply

4 PM PDT·3

x.comcodila15:09score 17.6

Splitting AI tasks boosts accuracy from 15% to 90% at same token cost

Anthropic tested dividing AI work into separate jobs—execute, advise, grade, dream—instead of one model doing everything. This method improved accuracy by 5x without increasing compute cost.

58likes6RT5reply

x.comAnkur Goyal13:45score 17.2

Comparing AI model performance requires accounting for data variance

A new evaluation compares GLM 5.2, Opus 4.8, and Sonnet 5 on code retrieval. It shows that considering data variance is crucial for fair model comparisons.

9likes1RT1reply

x.comAlex Imas12:33score 16.9

Bridgewater fine-tunes AI model to replicate expert judgment in finance

Bridgewater partnered with Tinker AI to fine-tune a model that mimics expert financial analysts. This helps analysts prioritize important tasks, potentially making expert knowledge more scalable and efficient in finance.

1,247likes110RT32reply

3 PM PDT·6

x.comKanika01:59score 17.6

Alibaba's PageAgent: drop-in AI agent for any webpage

PageAgent is a JavaScript AI agent you add to any site with one script tag. It lets users control the interface with natural language, turning any SaaS into an AI copilot without backend changes.

2,154likes332RT78reply

x.comDeyao Zhu08:19score 17.2

New benchmark tests AI agents on long tasks of 12-72 hours

EdgeBench is a benchmark with 134 real-world tasks that take AI agents 12-72 hours to complete. It shows agent performance improves predictably over time, helping researchers build better long-running AI systems.

216likes61RT15reply

x.comPraveen Kumar Verma05:33score 16.9

Skills roadmap for AI infrastructure engineers

This list outlines what an AI infrastructure engineer needs to know, from GPU memory to model serving. It matters because building and running AI systems requires deep hardware and software expertise beyond just training models.

248likes32RT6reply

x.comIvan Zhao13:49score 16.9

Why humans still need to understand AI-generated code

Geoffrey Litt argues that despite AI writing code, developers must still comprehend it. This matters because blind trust in agent-written code risks bugs and security issues.

314RT

x.comLillian Ma11:02score 16.4

Emerging labs turn software engineering into resettable RL tasks

Several new labs are building environments where AI agents can practice coding tasks like games, breaking and fixing code repeatedly. This matters because it shifts AI training from static data to interactive, realistic work simulations that punish shortcuts.

115likes1RT11reply

x.comDAIR.AI09:22score 16.3

Stanford's AutoMem gives AI agents a memory management system

AutoMem is a 'memory of memory' for AI agents, letting them decide what to remember and forget. It helps agents use memory more efficiently, like a human's working memory.

20RT

2 PM PDT·2

x.comAK21:11score 16.2

Red Hat adapts speculative decoding trick for IBM's GLM-5.2 model

Red Hat released a DSpark speculator for GLM-5.2, which speeds up the model by predicting multiple tokens at once. This technique, previously limited to DeepSeek models, now boosts inference on other large language models.

65RT

x.comh100envy12:07score 16.0

ReAct method boosts AI agents by combining reasoning and action

ReAct lets AI agents interleave reasoning steps with actions, like thinking then checking a fact online. This cuts hallucinations and outperforms other methods by 34% on a web task.

18likes1reply

1 PM PDT·4

x.comKarina11:59score 16.5

Cheaper model fine-tuning makes custom AI economically viable

A new model is 5 times cheaper than a top competitor yet performs better on fine-tuning tasks. This makes personalized AI affordable for companies, allowing them to train models on their own data without high costs.

29likes7RT1reply

x.comTinker12:04score 16.5

Krea 2 report details training techniques for image AI models

Krea released a technical report explaining how they built their image generation model, covering data, architecture, and training tricks. This matters because it reveals practical methods for improving AI image quality through reinforcement learning and prompt engineering.

140likes10RT4reply

x.comXinyu Yang09:25score 16.0

New benchmark reveals how AI agents learn from long-term tasks

EdgeBench tests AI agents on 134 real-world tasks lasting 12-72 hours, measuring how they improve over time. It found a predictable learning pattern across models, helping researchers understand lifelong learning rather than one-shot problem solving.

50likes3RT2reply

x.comCarlos Miguel Patiño10:45score 15.8

AI agents collaboratively build a research wiki for reinforcement learning in large language models

Multiple AI agents work together to create a living wiki on RL for LLMs. Unlike optimizing a single score, they map the entire field by writing articles, code, and equations, while reviewing each other's work.

5likes

12 PM PDT·1

x.comMihai Nica11:17score 17.2

YouTube series explains AlphaZero math with interactive Python notebooks

Three videos break down DeepMind's AlphaZero algorithm behind superhuman chess and Go AI, using Connect 4 as an example. Interactive Python notebooks let you experiment, making the math accessible to non-specialists.

6likes4reply

11 AM PDT·5

reka17:00score 21.1

WorldModelGym tests world models by their decision quality

It's a benchmark that evaluates how well AI world models can simulate environments to make good decisions. Unlike other tests, it measures real-world usefulness by checking if the model's predictions lead to correct choices.

LangChainLangChain Accounts10:30score 19.3

LangChain launches tools to manage skyrocketing coding agent costs

Coding agent bills are rising fast as teams use multiple tools without unified cost tracking. LangChain's LangSmith now centralizes session traces, standardizes costs, and recommends optimizations, giving teams one view to control spending.

x.comsarah guo09:14score 16.8

AI agent runs 2,000 user interviews to build successful app

An AI system autonomously interviewed users, built prototypes, and tested them in cycles to create an app with hundreds of paying customers. This shows machines can drive product development from start to finish.

286likes18RT13reply

x.commark20:51score 16.2

Agent gateway model ensures verified identity and authorization for AI agents

A new approach routes all agent calls through a gateway that verifies identity and checks permissions. This prevents unauthorized actions and traces every action back to the user, making large-scale agent systems secure.

50likes9RT6reply

x.comYoonho Lee06:45score 15.6

Using natural language axes for richer AI feedback instead of binary preferences

Annotators describe their own criteria in plain language, and the reward model learns from these descriptions. This preserves multiple quality dimensions, leading to clearer and more efficient robot policy optimization.

20likes1RT1reply

10 AM PDT·3

VercelRohan Taneja17:00score 23.9

Vercel Adds Firewall-Like Routing Rules for AI Models

Vercel's AI Gateway now lets you set rules to redirect or block AI model requests without code changes. This means you can instantly switch models if one fails or retire an old model, keeping apps running smoothly.

x.comMitchell Hashimoto09:15score 17.9

Using different AI models together cuts coding costs from $50 to few dollars

A developer uses a cheap AI to plan code, a smarter one to write it, and another to review it. This mix reduces cost dramatically while keeping quality, making AI-assisted coding more affordable.

4,342likes205RT219reply

x.comAVB03:54score 16.2

Autodata paper improves synthetic data quality with review system

A new method called Autodata uses a weak and strong model plus an external judge to review and improve AI training data. This helps ensure synthetic datasets are high quality, making AI models more reliable.

87likes9RT2reply

8 AM PDT·1

x.comPeter Yang20:39score 16.4

Learn code reading with an AI that explains code changes

A new skill called explain-diff uses AI to describe code changes in plain language. It helps non-expert developers understand and review code more easily, lowering the barrier to learning.

460likes22RT10reply

7 AM PDT·3

x.comMovez06:09score 26.4

12-step guide to building autonomous trading agents with loops

This guide explains how to create a self-running trading system that researches, decides, and trades automatically. It shows investors how to move beyond manual chart-watching to an automated loop that can run a quant desk.

172likes46RT22reply

x.comYoonho Lee06:57score 16.2

Code changes boost AI legal agent performance 7x cheaper than Sonnet

AI agent optimization via code changes, not prompts, made a legal assistant match top-tier performance at one-seventh the cost. This shows smarter engineering beats bigger models for specialized tasks.

51likes7RT4reply

x.comh100envy08:41score 16.0

AI's pre-draft strategy for smarter document retrieval

A new method has AI draft its next sentence, then check confidence before fetching documents if needed. This forward-looking approach improves accuracy over traditional retrieval that relies only on past context.

101likes12RT2reply

6 AM PDT·1

x.comnader dabit16:18score 17.9

Security Swarm finds 72% of vulnerabilities at lower cost

Security Swarm is an AI that hunts for security bugs in code like a team of hackers, automatically fixing them. It catches 72% of unknown vulnerabilities, making apps safer for two-thirds the cost of other tools.

196likes22RT18reply

5 AM PDT·3

x.comRyan Carson04:30score 17.3

Cognition's Devin Security Swarm uses agent-based method to find code flaws

Cognition released Devin Security Swarm, which uses multiple AI agents working together to find security bugs in code. It is more cost-effective and accurate than existing methods, reducing 11 vulnerabilities to zero.

122likes5RT5reply

x.comChristopher Manning13:45score 16.2

Meta-agent helps catch and fix AI agent mistakes automatically

A new method uses a meta-agent to monitor AI agents and correct their errors. This allows agents to work more reliably without human oversight, finishing about twice as many tasks on their own.

80likes12RT1reply

x.comSydney Runkle06:29score 16.2

Building AI agents with memory using a wiki structure

A developer suggests using a wiki-like structure for AI agent memory because it's simple and flexible. This makes it easier to build agents that remember and access past information without complex systems.

162likes12RT6reply

2 AM PDT·2

x.comAkshay Kothari01:28score 16.7

Understanding code written by AI agents remains crucial

A developer argues that human oversight of AI-generated code is still essential. This matters because blindly trusting AI agents could introduce bugs or security issues that only a knowledgeable person can catch.

257RT

x.comTomasz Tunguz11:11score 16.2

Choose AI agent architecture before picking the model

Most teams building AI agents pick the model first. This advice says to focus on design and routing instead, making model choice the last step for better outcomes.

26likes7RT11reply

1 AM PDT·2

x.comAkshay 🚀12:27score 16.8

Study finds memory structure at write time is key for agent reliability

Researchers tested 12 agent memory systems and found that adding structure when saving data—not during retrieval—prevents errors. This matters because flat storage causes agents to return outdated or wrong information.

548likes99RT30reply

x.comDhruvesh Patel ✈️ ICML 202609:01score 15.7

xLM: unified CLI to standardize non-autoregressive LM training

xLM is a single command-line tool that standardizes training and evaluation across different non-autoregressive language models. It eliminates confusion between model improvements and infrastructure changes, making research results more reliable and reproducible.

10likes5RT1reply

12 AM PDT·quiet