Search “Search”

Imperial College London Study Reveals Search-Augmented LLMs Cite Sources Without Reading Them

On October 16, 2026, three authors from the Department of Computing at Imperial College London published the paper "Retrieved but Not Read: Understanding Why Search-Augmented LLMs Fail to Use Available Evidence," investigating why search-augmented LLMs frequently cite web pages they never actually opened. The study tested six open-weight configurations—GPT-5.6 Luna, GLM-5.3-Flash, gpt-oss-120b, gpt-oss-20b, gemma-4-31b, and qwen3-8b—using 225 facts that changed after February 16, 2026. The results showed that models often answer based solely on search snippets without opening the source pages; under the same set of search results, the proportion of turns in which no page was opened ranged from 2.6% to 76%, and when Luna used its built-in search, 53.6% of turns relied only on snippets. When answering from snippets alone, outdated answers accounted for 6.4% and misleading citations for 6.4%; when pages were opened and cited, these two figures dropped to 1.3% and 2.3%, respectively. The study also found that placing the correct page directly into the prompt significantly reduced errors, but some models still favored old information even when shown new facts. Currently, this research has been reported by a single source, with no subsequent developments yet.

Convovo study: When forced to choose, Grok 4.7 and GPT-6.1 Sol show clear partisan leanings

A Convovo study of nearly 500 elections found that when forced to pick between two options via API, SpaceXAI's Grok 4.7 recommended Republican candidates in 92% of responses, while OpenAI's GPT-6.1 Sol recommended Democratic candidates in 97% of responses, with the two agreeing less than 10% of the time. The study also found that Grok frequently cited conservative sources, and that each API request carried roughly 1,240 extra tokens that researchers could neither view nor remove; after disabling web search, GPT's Democratic lean persisted, while Grok's Republican lean disappeared, with Grok declining to choose in 27 of 30 cases. OpenAI responded that the test targeted the developer API and that the prompt explicitly required recommending a candidate, so it does not reflect the ChatGPT consumer experience, where the app would refuse to answer such questions.

W3C releases discussion document on the Web and AI agents in the AI era

The W3C Strategy Team has released a discussion document, "W3C and the State of the Web in the AI Era," proposing the use of Web standards to make AI agents safer, with a focus on prompt injection protection, controllable permissions, and accessibility. The report states that the document was released on October 21 to build momentum for the TPAC meeting; it is a proposal rather than an approved standard, with an official presentation scheduled for October 29. Using travel booking as an example, the text points out that having an agent search for flights does not amount to authorizing it to pay for tickets, and recommends setting spending limits, revocable authorization, and traceable credentials for payments. The document is currently still in the discussion stage, with no further progress yet.

Imperial College study reveals that search-augmented LLMs cite sources without reading them

On October 16, 2026, three authors from the Department of Computing at Imperial College London published the paper "Retrieved but Not Read: Understanding Why Search-Augmented LLMs Fail to Use Available Evidence," investigating why search-augmented LLMs frequently cite web pages they never actually opened. The study tested six open-weight configurations—GPT-5.6 Luna, GLM-5.3-Flash, gpt-oss-120b, gpt-oss-20b, gemma-4-31b, and qwen3-8b—using 225 facts that changed after February 16, 2026. The results showed that models often answer based solely on search snippets without opening the source pages; under the same set of search results, the proportion of turns in which no page was opened ranged from 2.6% to 76%, and when Luna used its built-in search, 53.6% of turns relied only on snippets. When answering from snippets alone, outdated answers accounted for 6.4% and misleading citations for 6.4%; when pages were opened and cited, these two figures dropped to 1.3% and 2.3%, respectively. The study also found that placing the correct page directly into the prompt significantly reduced errors, but some models still favored old information even when shown new facts. Currently, this research has been reported by a single source, with no subsequent developments yet.

View story:Imperial College London Study Reveals Search-Augmented LLMs Cite Sources Without Reading Them

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

View story:English translation is being prepared.

UC Santa Barbara paper proposes TTME to replace test-time training with memory

The UC Santa Barbara paper proposes the TTME method, which replaces test-time training with evolutionary memory in mathematical discovery tasks: keeping model weights unchanged, it stores the rules and score-improvement ideas from previous attempts in memory to guide subsequent attempts. The report states that across 5 mathematical tasks, 2 frameworks, and 3 models, TTME exceeded the reported test-time training scores every time; on 3 tasks with Qwen3-8B, its FLOPs were reduced by 82.8% to 84.6%, and runtime was reduced by 47.7% to 58.4%. The authors suggest that if an LLM search loop comes with an automatic checker, one can first try using experiential memory rather than paying the cost of weight updates in the middle of the search. There are currently no further developments.

Chelsea Finn proposes optimizing information organization to extend LLM context

On October 22, 2026, Chelsea Finn reposted the new paper Librarian Agents by Anikait Singh et al., and proposed optimizing information organization to improve downstream utility. She pointed out that agentic search can scale to information of arbitrary size but tends to miss relevant content; whereas optimizing information organization for downstream use can bring 30+% improvements even in OOD domains. The tweet includes links to the paper and code. There are currently no subsequent development facts.

Microsoft researchers prove the KLS conjecture

Microsoft Research researcher Xinzhi Zhang and co-authors proved the KLS conjecture proposed in 1995, giving a dimension-independent fixed bound; the previous best result only allowed the bound to grow slowly with dimension. According to Microsoft Research, the proof came from a human-guided agent search process, in which the team explored more than 100 candidate paths, which researchers evaluated, pruned, and redirected. An independent team also gave a complete proof around the same time, followed by Lean 4 formal verification. The authors state that previous unguided end-to-end attempts with multiple frontier models failed to achieve the goal. Another report states that AI agents explored about 100 research paths, the draft was about 140 pages, AI was responsible for exploring candidate methods, executing derivations, assisting in writing the manuscript, and performing internal proof checks, while the recursive method and error bounds that completed the proof are credited to the two researchers.

English translation is being prepared.

Anthropic announces research on Mythos 5 autonomously designing quantum circuits

On October 17, 2026, Anthropic published a study: when given only a pure natural-language specification and an automated evaluator, Mythos 5 essentially without human intervention repeatedly submitted and improved quantum circuit designs in a closed loop, finding circuits that use fewer qubits and gates than the best previously published solutions, with some improvements reaching an order of magnitude. The elliptic-curve modular inversion design came from a single 53-hour continuous run, in which the model derived a reversible divsteps implementation that requires no storage of branch information and formally verified it with Lean; another effort reduced the gate count of the Reed-Solomon decoding step, the most expensive step in DQI, to about 1/12 of a construction previously made by Google. Anthropic also noted that the model was initially conservative and tended to follow known methods, which needed to be mitigated by requiring parallel exploration of multiple approaches and occasionally prompting it to broaden its search. Subsequent reports said the research came from the STEM Fellows program and gave differing formulations, such as a 17-fold reduction in gate count for the divsteps task.

English translation is being prepared.