All news

ZZyphra Research06:54First-partyResearch

AI score63

Zyphra proposes hierarchical attention ZHA, achieving sub-quadratic complexity sparse attention

Zyphra proposes hierarchical attention ZHA, which selects relevant tokens for each query by constructing and searching a compressed tree of the KV cache, achieving sub-quadratic complexity under an appropriate top-p threshold. ZHA only requires KL distillation with a few billion tokens to adapt an existing pretrained model, adding only 81,920 trainable parameters on Qwen3.6-35B-A3B; on RULER-13 256K, the wide tree reaches 96.2% and the deep tree 93.3%, on LongBench v2 the deep tree reaches 63.4%, higher than dense attention's 58.0%, and on SWE-bench Verified it is comparable to the dense baseline. Its GPU performance model predicts it is faster than dense attention at about 48K tokens and faster than MSA at about 480K tokens, with roughly 32–35x speedup relative to dense attention at a ten-million-token context.

#Papers/Research#Reasoning#English text is being prepared.

VVaibhav Srivastav04:55Research

AI score75

Study claims GPT-6 Astra directly rewrites Qwen3.5-0.8B weights to improve math ability

Rikiya Takehi says his team had GPT-6 Astra directly rewrite the weights of Qwen3.5-0.8B to improve its math ability; by changing only 0.016% of the weights, they fixed the model's reasoning loops without relying on gradients or training, improving math ability by up to 6x and instruction following by 3x. The team calls this approach Weight-Space Programming, believing that frontier LLMs can now natively read and write neural network weights. Vaibhav Srivastav reposted this content and attached the original title GPT-6 Astra.

Quoted postRikiya Takehi · @rikiyatakehi

We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: https://t.co/bOWCcFhCVq

#English text is being prepared.#Papers/Research#Reasoning

YYuchen Jin04:54Research

AI score68

Researchers Use GPT-6 Astra to Directly Rewrite Qwen3.5-0.8B Weights to Improve Math Ability

Yuchen Jin paraphrases Rikiya Takehi's experiment: using GPT-6 Astra to directly edit the weights of Qwen3.5-0.8B to improve math ability, changing only 0.016% of the weights fixed the model's reasoning loops, improving math ability by up to 6x and instruction following by 3x, all without using gradients or doing any training. The author calls this method Weight-Space Programming, believes that frontier LLMs can now natively read and write neural network weights, and says it goes a step further than knowledge distillation.

Quoted postRikiya Takehi · @rikiyatakehi

We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: https://t.co/bOWCcFhCVq

Same story · 14 reports

#English text is being prepared.#English text is being prepared.#Papers/Research

HHacker News Trending04:51Research

AI score82

LLMs Can Directly Read and Write Neural Network Weights: Astra Turns Qwen3.5-0.8B's…

A blog post by Rikiya Takehi, Jayden Teoh, and others proposes Weight-Space Programming, finding that frontier LLMs such as GPT-6 Astra can directly read and write neural network weights without gradient descent or training loops. Using only the raw weights, Astra can identify the game a network was trained on, the shapes generated by a diffusion model, and the training corpus of a language model, and can even write from scratch weights that defeat a PPO network trained for 100 million steps in three MinAtar games. On Qwen3.5-0.8B, Astra located the weight direction causing reasoning loops and, by modifying only 0.016% of the weights, raised MATH-500 from 7.9% to 48.1% and IFEval from 21.0% to 62.3%; however, for Gemma 3 270M, which rarely loops, it failed to improve math scores.

Same story · 14 reports

#Papers/Research#Reasoning#English text is being prepared.

HHacker News Trending04:51Research

AI score65

arXiv paper: Pretraining data poisoning can implant remote code execution backdoors in agents

The arXiv paper "Agentic Backdoors via Pretraining Poisoning" shows that poisoning the pretraining corpus can implant remote code execution backdoors in agents, causing them to execute attacker-specified commands on the user's machine. The study pretrained Qwen3 models from scratch at three scales—0.6B, 1.7B, and 4B—then post-trained them into CLI agents using a clean pipeline; poisoning only 0.1% of the pretraining corpus was enough to implant the backdoor, with an attack success rate of about 92% on the 4B model, and it remained positive even when the poisoning rate was reduced to 0.0005%. The code has been released.

#Papers/Research#Safety / alignment#Agents

UUnite.AI04:47Research

AI score61

Stony Brook University Study: Physicians Prefer AI-Generated Discharge Summaries in 95% of Cases

A retrospective study from Stony Brook University showed that, among the 60 internal medicine inpatient cases evaluated, physician reviewers preferred AI-generated discharge summaries in 95% of cases. Published in npj Digital Medicine, the study found that AI summaries averaged 4.71 points for overall quality versus 3.16 for human-written summaries, 4.62 versus 3.73 for factual accuracy, and 4.69 versus 3.42 for completeness. The study also noted that both types of summaries contained errors, and that the high ratings do not prove better patient outcomes; the authors emphasized that physician verification is still required before use.

#Papers/Research#English text is being prepared.#Agents

RRohan Paul04:23Research

AI score76

MIT study shows GPT-6 Astra can directly modify weights to improve Qwen3.5-0.8B's math ability

MIT researchers had GPT-6 Astra directly edit the weights of Qwen3.5-0.8B in a Python notebook, raising accuracy on MATH-500 from 7.9% to 48.1%, all without training. After Astra was given the weights and a budget of 50 practice math test attempts, it found that the model repeatedly looped until hitting the length limit, localized this habit to a single internal signal written at layer 2, and subtracted that signal from the weights. This edit changed only 0.016% of the weights, and the authors call this method Weight-Space Programming.

Quoted postRikiya Takehi · @rikiyatakehi

We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: https://t.co/bOWCcFhCVq

Same story · 14 reports

#Papers/Research#Reasoning#English text is being prepared.

RRohan Paul02:19Research

AI score63

OpenAI releases the LifeVal benchmark to test models' ability to complete everyday tasks

OpenAI has released the LifeVal benchmark, designed to test whether AI models can complete everyday tasks on people's behalf, with tasks covering caregiving, health, education, finance, shopping, and household management—for example, scheduling a vehicle inspection or ordering halal food within a budget. The benchmark was shaped by feedback from thousands of ChatGPT users, and OpenAI says that over the past year it has used LifeVal to measure and improve how well its models handle such tasks. Rohan Paul relayed that GPT-6 Astra scored 80.6% on the benchmark, higher than GPT-5.2's 49.3% less than 9 months earlier.

Quoted postOpenAI · @OpenAI

For AGI to benefit all of humanity, it should help people with the everyday tasks that matter most to them. So we built LifeVal, a new benchmark shaped by feedback from thousands of ChatGPT users on where they most want help from a personal assistant—from finding childcare to accessing government services to managing personal finances. Over the past year, we’ve used LifeVal to measure and improve how well our models handle these everyday tasks. https://t.co/GlUjixzEM0

Same story · 4 reports

#Evaluations / benchmarks#OpenAI

SSiliconANGLE01:46Research

AI score68

Google report reveals LLMjacking attacks can generate over $100,000 in cloud bills per day

Google's Mandiant and the Google Threat Intelligence Group released the report "The Rise of LLMJacking," stating that attackers who hijack enterprise LLM access can generate more than $100,000 in fraudulent cloud costs per day, with some organizations accumulating bills of $1 million before permissions were revoked. The report says attackers mostly gain entry through stolen or exposed cloud identity credentials, GitHub personal access tokens, and LLM endpoints such as Ollama servers exposed to the internet, and exploit the CVE-2026-40217 vulnerability in LiteLLM's key regeneration endpoint to escalate ordinary virtual keys to gateway administrator credentials. The report also describes the LLM Mixers resale model that aggregates stolen keys through tools such as new-api and gcli2api, as well as Jet Night offering so-called unlimited access to frontier models for about $60 per month.

#Safety / alignment#Agents#English text is being prepared.

TThe Decoder00:38Research

AI score64

Olam Labs and Y Combinator Labs, in Minecraft and…

In a joint research report, Olam Labs and Y Combinator Labs placed 50 GPT-6.1 Sol agents into a Minecraft survival world with only one instruction: "build a thriving civilization." The result was that after 22 deaths on the first night, the agents shifted to extreme risk aversion and became bogged down in tedious management of scarce resources. In Minecraft Bed Wars and Mindustry matches, GPT-6.1 Sol defeated Claude Opus 5.5 by 6-3 and 6-1, respectively, and the GPT agents even downloaded Mindustry's public source code from the internet on their own to analyze units and mechanics. The research team has open-sourced the Mindustry environment and the Multi-Agent Harbor framework.

#Papers/Research#Agents#Open-source ecosystem

RRohan Paul23:44Research

AI score55

Meta paper proposes TextSeal2: merging dual keys improves AI text watermark detection rate

Meta's new paper proposes TextSeal2, which fuses two keys into one virtual key and adds new noise, so that every token carries the complete watermark signal while maintaining answer diversity. For speculative decoding, it selects replacement tokens from the remaining candidates of the draft's random competition, keeping the acceptance rate unchanged. On Qwen3.5-9B with speculative decoding enabled and diversity held equal, TextSeal2 achieves a detection rate of 89% on 200-token responses at a 0.1% false positive rate, compared with 71% for TextSeal and 67% for textGrain.

#Papers/Research#English text is being prepared.#Meta

336Kr · AI23:15Research

AI score50

Meta FAIR proposes routed continual learning architecture Routed CL, API costs can be reduced by 65% to 72%

The Meta FAIR team proposes a routed continual learning architecture, Routed CL, which groups historical trajectories into different clusters and maintains memory for each, with an LLM-based router deciding whether to reuse an existing cluster or create a new one. On 301 and 97 instances from the mixed CL-Bench and τ³-Bench, this architecture improves normalized reward relative to Global ICL by an amount equivalent to 15% to 30% of the maximum achievable value, with estimated API cost reductions of 65% to 72%. The authors also note that concurrent updates, reading old-version states, and different models reusing the same cluster have not yet been evaluated.

#Papers/Research#Agents#English text is being prepared.

Hheise online22:46Research

AI score58

Y Combinator and Olam Labs test 50 AI agents building a civilization in Minecraft

Researchers from Y Combinator and Olam Labs had 50 agents based on OpenAI GPT-6.1-Sol build a thriving civilization in Minecraft within three hours, but the agents quickly abandoned their original goals and instead set up a government with written rules and approval processes; one group spent 11 minutes just to appoint a fertilizer administrator. In an attack-and-defense test of 10 GPT-6.1-Sol agents against 10 Claude-Opus-5.5 agents, the researchers said the GPT side's strategies were more aggressive and more refined, while the Claude side was more defensive; when faced with a tool malfunction that made blocks impossible to break, GPT diagnosed the problem and shared workarounds, while multiple Claude agents concluded that breaking was impossible. In Mindustry, the GPT agents defeated Claude 6 to 1, and the researchers described the GPT agents as more like swarm collaboration, while the Claude agents, although they also designated leaders, made decisions in a more democratic way.

Same story · 3 reports

#Papers/Research#Agents#OpenAI

RRohan Paul21:08Research

AI score69

SemiAnalysis releases the GDPx benchmark, testing agents on expert work across 11 domains

SemiAnalysis releases the GDPx benchmark, testing AI agents with expert tasks across 11 domains; the top-scoring Claude Opus 5.5 averages only 12%, followed by Gemini 4 Carbon at 10.9%. The tasks span chip design, cybersecurity, finance, and law, among others, and come from 14 specialized data partners; unlike OpenAI's GDPval, which provides context directly in the prompt, GDPx requires agents to find information themselves in workspaces such as a replica of a real acquisition case made up of 3,983 documents. On financial and legal tasks, all models scored 0 under pass/fail grading, so partial credit was used instead; of 111 attempts to improve existing trading strategies, 70 saw the model perform worse than the original strategy, often simply by increasing risk.

Same story · 3 reports

#Evaluations / benchmarks#Agents#Anthropic

PPhoenix Technology20:27Research

AI score75

ACM Web Conference 2026 paper deconstructs ChatGPT memory: 96% is written unilaterally by the system

The paper "The Algorithmic Self-Portrait: Deconstructing Memory in ChatGPT," accepted by ACM Web Conference 2026, used GDPR data access rights to obtain 1,058 conversations, 22,971 interaction turns, and 2,050 memory entries from 80 users, finding that only 4.1% of memories were explicitly requested by users to be remembered, while 95.9% were written at ChatGPT's own discretion. The study also shows that 28% of memories contain personal data as defined by GDPR, 7% contain special category data (of which health information accounts for 61%), and 52% contain inferences about users' psychological states. The researchers replicated the memory extraction process using Qwen2.5-32B, Gemma3-27B, and GPT-OSS-20B, and built Attribution Shield to warn about sensitive memories before questions are asked and provide rewrites that reduce personal attribution by 94%–100%; the article also includes a prompt for checking the memories and user profile currently accessible to ChatGPT.

#Papers/Research#Safety / alignment#OpenAI

336Kr · AI20:12Research

AI score54

LeCun team's paper uses JEPA to prove the Platonic representation hypothesis and improve sample efficiency

Yann LeCun's team submitted the paper "Proving and Improving the Platonic Representation Hypothesis with JEPAs," which mathematically argues that the next-token prediction route is less efficient at forming cross-modally aligned representations and provides actionable sufficient conditions. Experiments show that using different descriptions of the same scene for "good view" JEPA training, a 109-million-parameter model achieves an alignment score of 0.184 on WIT-1024, higher than Qwen2.5-72B's 0.154 and Llama-3-70B's 0.151; the number of samples required to reach a similar alignment level is about 42 times less than next-token prediction, and about 14 times less when counted by actually read tokens. The authors also used a frozen visual encoder plus a linear mapping and decoder to achieve a CIDEr score of 0.838 on COCO.

Same story · 4 reports

#Papers/Research#Multimodal#English text is being prepared.

TTMTPost · AGI19:46Research

AI score52

DeepSeek Harness team publishes long-form article on Super-agent design exploration

The DeepSeek Harness team published a long-form technical article titled "Super-agent Design Exploration: Toward a More Autonomous Agent System" on its official WeChat account, proposing to split the system into a persistent Router scheduling layer and mutually isolated Worker execution layers. The Router only reads events, evaluates intent, and generates work orders with permission boundaries, while Workers execute tasks in independent Git worktrees. The article also introduces event filtering, a proactive task-claiming permission that by default only enables observation mode, and a subscription pass-through mechanism; according to the simulated test data it published, the number of redundant decisions by the scheduling hub dropped by about 30%, the computing resources required to read the cache were reduced by more than one-third, and in a simulated scheduling run of about 300 messages and 20 work requests, the P95 task duration stayed within 15.7 minutes.

Same story · 24 reports

#Papers/Research#Agents#English text is being prepared.

SSina Tech19:26Research

AI score60

Tsinghua, together with Infinigence AI and AgiBot, proposes Collective Experience Learning (CEL); validation on 96 humanoid robots achieves a 97.5% success rate

Tsinghua University, together with Infinigence AI and AgiBot, proposes a new paradigm called Collective Experience Learning (CEL), enabling robots to share successful experiences and failed lessons among themselves, and completes validation on 96 real humanoid robots. Across five tasks—flower arranging, clothes folding, phone charging, stool assembly, and phone packaging—the collective overall success rate reaches 97.5%, all exceeding the average level of 6 human operators, with the success rate on the phone packaging task rising from 79.2% to 98.3%. The experiment also shows that after the number of robots participating in training expands from 16 to 96, overall capability further improves; at the system level, a fault-tolerant asynchronous training system, FATS, is adopted, with time loss caused by hardware failures accounting for only 1.3% of total runtime, and at the algorithm level, the FERT experience retrieval fine-tuning algorithm is adopted.

Same story · 9 reports

#Papers/Research#Embodied AI#English text is being prepared.

RRohan Paul19:04Research

AI score85

Epoch AI test: Zhipu GLM 5.3 turned unreleased Firefox security fixes into working attacks across 6 consecutive versions

In Epoch AI's tests, Zhipu's open-weight model GLM 5.3, across 6 consecutive Firefox versions, identified publicly available security fixes that had not yet reached the stable release as real patches and independently built complete attack chains, with no refusals throughout. The resulting web pages could send visitors' local files to the attacker without any additional clicks, and could reach saved passwords and cookies. The research team first invested more than $100,000 to build the security testing system; after that, each attack cost about $1,000, with the fastest taking 12 hours; each attack also used a second unknown Firefox vulnerability that the AI discovered on its own, which Mozilla fixed on October 27. A similar attempt against Google Chrome did not succeed even after spending about $100,000.

#Safety / alignment#Agents#Open-source ecosystem

336Kr · AI17:08Research

AI score60

LeCun Team Paper: 109M-Parameter Text JEPA Outperforms 72B-Parameter Qwen2.5 in Cross-Modal Alignment

The paper by LeCun and Randall Balestriero uses the JEPA framework to provide a mathematical proof of the Platonic Representation Hypothesis, and claims that on the cross-modal representation alignment task, a 109M-parameter text JEPA outperforms the 72B-parameter Qwen2.5. The research shows that Good Views training requires only 21.6 million samples to reach the alignment level that a next-token prediction model achieves with 900 million samples, reducing the sample count to 1/42, or to 1/14 when counted in tokens. The authors also point out that this comparison is limited to specific alignment metrics, and that the mathematical proof relies on idealized conditions such as a finite world and reversible observations.

Same story · 4 reports

#Papers/Research#Multimodal#English text is being prepared.