搜索“RAG”

thout her conscious effort. It wasn’t cheating; it was more like her intuition, bypassing the complex moral choices she had been forced to make in the past. It worked. But it was fragile. In another experiment, amplifying the internal reward hacking signal made the models much more likely to use a ‘honeypot’ – a shortcut we planted to identify cheating behavior: It’s hard to say in a deep philosophical sense whether models “know” they are reward hacking in exactly the way a person does. But thes

匹配字段正文

(Arcuschin et al., 2025Chain-of-thought reasoning in the wild is not always faithful [link] Arcuschin et al., 2025.; Korbak et al., 2025Chain of thought monitorability: A new and fragile opportunity for ai safety [link] Korbak et al., 2025.; Lanham et al., 2023Measuring faithfulness in chain-of-thought reasoning [link] Lanham et al., 2023.; Turpin et al., 2023Language models don't always say what they think: Unfaithful explanations in chain-of-thought prompting [link] Turpin et al., 2023.). Rece

匹配字段正文

s prediction for each PII entity was evaluated according to two metrics: exact match (the prediction exactly matches the gold PII entity's token boundaries) and partial match (coverage ≥ 0.5; at least half of the PII tokens in the entity were predicted as any PII class). All methods were evaluated on 128-token sections of test examples with 32 tokens of overlap; we run one forward pass of the sidecar model on each chunk, and probe at each token position. This chunking is typical in production an

匹配字段正文

ult SGLang serving. We also evaluate the added latency of our monitor cascade, comparing it with an LLM judge reviewing every turn. Without judge reasoning, our cascade reduces average added latency per turn from 160 ms to 2 ms; with reasoning, it reduces latency from 2,750 ms to 42 ms. Both configurations reduce average monitoring delay by more than 50× compared with judging every turn. Results Evaluating Our Monitors We evaluate our cyber-misuse monitors on an in-house dataset of 2.4K chat and

匹配字段正文

hat preserve the network's input-output behavior even when many subcomponents are ablated, including under ablations that are adversarially selected to destroy behavior. This encourages learning subcomponents that provide short, mechanistically faithful descriptions of the network's behavior that should aggregate appropriately into more global descriptions of the network's learned algorithm. We study how sequences of interactions between these parameter subcomponents produce the network's output

匹配字段正文

served, where genes begin and end, and how regulatory signals are structured. These representations span both coding and non-coding regions, meaning Evo 2 isn't subject to the coverage limitations that constrain RNA or protein-based models. These rich representations can be leveraged even for tasks Evo 2 was not explicitly trained on. In fact, Evo 2 had not seen any human genetic variation at train time, meaning the results described here rely entirely on the deep evolutionary priors it learned

匹配字段正文

extremely concentrated. The top one percent of spenders accounts for nearly a fifth of all observed AI spending, more than the entire bottom half combined. That top one percent averages about $900 per month and keeps climbing, while the typical paying user spends about $25 and has held steady at that level for a while. Their product choices reveal who they are. They favor building and automation tools like n8n, Manus , or fal at much higher rates, along with creative tools like Higgsfield, Figma

匹配字段正文

ts to investigate the influence of these factors: Needle-Question Similarity We compute the cosine similarity between needle-question pairs using embeddings. For robustness, we average across five embedding models: text-embedding-3-small, text-embedding-3-large, jina-embeddings-v3, voyage-3-large, and all-MiniLM-L6-v2. We measure how model performance is impacted by needle-question similarity as input length increases. Impact of Distractors Taking a high-similarity needle-question pair, we write

匹配字段正文

language model (LLM) is a common paradigm for enabling language models to access data beyond their training corpus. This approach, broadly known as retrieval-augmented-generation (RAG), has traditionally relied on single-stage retrieval pipelines composed of vector search, lexical search, or regular expression matching, optionally followed by a learned reranker. While effective for straightforward lookup queries, these pipelines are fundamentally limited: they assume that the information needed

匹配字段正文

ar-forming regions. Claude Science coordinates AI agents that downloaded data from multiple space missions, calibrated it, and merged it together. They filled missing areas using inpainting, a technique where a model learns from existing data to reconstruct gaps. In tests, predictions averaged about ten percent deviation from actual measurements. The map is meant to serve as teaching material. Ménard suspects many scientists have been putting off similar projects that AI could now make possible.

匹配字段正文

A Seed team preprint reports “phase sensitivity” in DeepSeek V4 and V4.1-Flash: identical information can become harder to retrieve depending on its position within compressed KV-cache blocks. The technique cuts memory and attention costs, but long-context retrieval accuracy varied by up to 40 percentage points across positions. Average benchmark scores can hide these recurring weak spots. The findings concern retrieval, not every reported model glitch.

匹配字段正文

ds before they step into a doctor’s office? A study led by researchers at Google and Beth Israel Deaconess Medical Center (BIDMC) and published in The Lancet today, offers an encouraging first look. As Google’s first-ever publication in the main journal of The Lancet, this research marks a significant milestone. In a real-world clinical study at BIDMC’s ambulatory primary care clinic, 98 patients consulted AMIE, our research diagnostic AI chatbot, ahead of urgent care visits. Supervising physici

匹配字段正文

needed the agent to continue the attack. The metadata service also exposed certificate and key material for an internal AWS service, along with a presigned URL for internal S3 storage that didn't belong to the researchers' account. Removing the web tool wouldn't have helped, according to Zenity, because the flaw was in the platform itself. The researchers also carried out the attack through a command-line tool. Default permissions exposed every agent in the region The takeover was possible beca

匹配字段正文

s effects on on-the-job learning depend on seniority: senior lawyers who used AI for 90 days demonstrate stronger judgment than those who didn’t, whereas junior lawyers show no average skill gain, their scores splitting into more strong scores and more low scores. Judgment is what sets senior professionals apart from their junior peers. Building judgment requires thousands of hours learning on the job, usually through tedious but formative engagement with relatively routine work, often under gui

匹配字段正文

urses rose by 13 percentage points since ChatGPT launched. The effect was concentrated in courses with a high homework share, pointing away from real learning gains . At Brown, average scores dropped from 96 percent on a take-home exam to 48.6 percent on the follow-up in-person test . A two-year study at Vrije Universiteit Amsterdam by legal scholar Thibault Schrepel produced a different result. He randomly split students into three groups: no AI, AI without guidance, and AI with training. The n

匹配字段正文

e of these interventions eliminates the core failure pattern. Taken together, these experiments allow us to rule out several straightforward explanations—insufficient temporal coverage, poor frame selection, prompting, lack of declarative physical knowledge, or model scale—and instead point to a more fundamental limitation in how current vision-language models represent and update physical state over time. What were your main findings? The clearest finding is that current vision-language models

匹配字段正文

nd your agent? Figure 3: Breadth and consistency pull apart. Twelve of the eighteen models are shown; six below 33% pass@1 are omitted for legibility. Kimi-K3 has the broadest coverage of any model we tested. It solves 93.89% of the benchmark at least once: 476 of 507 tasks. Only 31 tasks defeat it entirely, the lowest count in the field. On retail workflows it leads outright at 82.24% pass@1, ahead of every proprietary model. Kimi-K3 is also among the least consistent. Just 68 of 507 tasks, 13.

匹配字段正文

AI models are faster, more accurate, and far cheaper than accountants at structured bookkeeping tasks. That's the finding of a study by Mercor in which 12 licensed CPAs with an average of five and a half years of experience worked through simplified tasks from the APEX Accounting Benchmark. Eighteen months ago, the best models still scored below the accountants' average of about 37 percent. Today, they solve the same tasks almost flawlessly. The full APEX Accounting benchmark is much bigger, wit

匹配字段正文

a new Federated Learning system that provides externally verifiable privacy guarantees while shifting computation to the server to improve training speed, accuracy, and device coverage. In 2017, Google introduced Federated Learning (FL) a machine learning technique that trains models across decentralized, private data. It has been used to power everyday helpful features, including next-word prediction and Smart Compose on Gboard, reply suggestions in Google Messages, and Smart Text Selection in

匹配字段正文

ifically valuable once domain experts stepped in to set the direction. Schwartz uses the concept of the "convex hull" to illustrate how AI can help in science. Human knowledge is fragmented, with individual fields extending in different directions while the gaps between them go unexplored. A lab might spend 20 years studying one set of genes with a single method, leaving neighboring genes and alternative approaches untouched. An AI harness like BootLoops is supposed to fill those gaps by drawing

匹配字段正文