arXiv:2604.27169v1 Announce Type: new Abstract: We show that the geometric relations between semantic features in large language models' hidden states closely mirror human psychological associations. We construct feature vectors corresponding to 360 words and project them on 32 semantic axes (e.g. beautiful-ugly, soft-hard), and find that these projections correlate highly with human ratings of those words on the respective semantic scales. Second, we find that the cosine similarities between the semantic axes themselves are highly predictive of the correlations between these scales in the survey. Third, we show that substantial variance across the 32 semantic axes lies on a low-dimensional subspace, reproducing patterns typical of human semantic associations. Finally, we demonstrate that steering a word on one semantic axis causes spillover effects on the model's rating of that word on other semantic scales proportionate to the cosine similarity between those semantic axes. These findings suggest that features should be understood not only in isolation but through their geometric relations and the meaningful subspaces they form.

arXiv:2604.27039v1 Announce Type: new Abstract: Token serves as the fundamental unit of computation in modern autoregressive models, and generation length directly influences both inference cost and reasoning performance. Despite its importance, existing approaches lack fine-grained length modeling, operating primarily at the coarse-grained sequence level. We introduce the Length Value Model (LenVM), a token-level framework that models the remaining generation length. By formulating length modeling as a value estimation problem and assigning a constant negative reward to each generated token, LenVM predicts a bounded, discounted return that serves as a monotone proxy for the remaining generation horizon. This formulation yields supervision that is annotation-free, dense, unbiased, and scalable. Experiments on LLMs and VLMs demonstrate LenVM provides a highly effective signal at inference time. On the LIFEBench exact length matching task, applying LenVM to a 7B model improves the length score from 30.9 to 64.8, significantly outperforming frontier closed-source models. Furthermore, LenVM enables continuous control over the trade off between performance and efficiency. On GSM8K at a budget of 200 tokens, LenVM maintains 63% accuracy compared to 6 percent for token budget baseline. It also accurately predicts total generation length from the prompt boundary. Finally, LenVM's token-level values offer an interpretable view of generation dynamics, revealing how specific tokens shift reasoning toward shorter or longer regimes. Results demonstrate that LenVM supports a broad range of applications and token length can be effectively modeled as a token-level value signal, highlighting the potential of LenVM as a general framework for length modeling and as a length-specific value signal that could support future RL training. Code is available at https://github.com/eric-ai-lab/Length-Value-Model.

arXiv:2604.27019v1 Announce Type: new Abstract: Safety-aligned language models must refuse harmful requests without collapsing into broad over-refusal, but the training-time mechanisms behind this tradeoff remain unclear. Prior work characterizes refusal directions and jailbreak robustness, yet does not explain how dynamic adversarial fine-tuning changes refusal carriers across training. We present a measurement-driven mechanism study, not a new defense, on one 7B backbone under supervised fine-tuning (SFT) and R2D2-style dynamic adversarial fine-tuning. Our protocol aligns fixed-source HarmBench, StrongREJECT, and XSTest with a five-anchor refusal-geometry suite and causal interventions. R2D2 drives fixed-source HarmBench ASR to 0.000 at steps 50 and 100, then partially reopens to 0.035 at step 250 and 0.250 at step 500; SFT remains less robust, with ASR between 0.505 and 0.588 at the same anchors. On XSTest, R2D2 any-refusal is 1.000 early, then falls to 0.664 and 0.228. Geometrically, R2D2 preserves a late-layer admissible carrier through step 100 before relocating to an early-layer carrier, while effective rank remains near 1.23--1.27. Causal interventions indicate low-dimensional but utility-coupled control. These results support a reorganization account rather than a drift-only account, with evidence limited to one backbone and fixed-source attacks.

arXiv:2604.26985v1 Announce Type: new Abstract: Masked diffusion models (MDMs) generate discrete sequences by iterative denoising under an absorbing masking process. In standard masked diffusion, if a token remains masked after a reverse update, the model discards its clean-state prediction for that position. Thus, still-masked positions must be repeatedly inferred from the mask token alone. This design choice limits cross-step refinement. To address this limitation, this paper proposes a simple, yet effective, post-training adaptation for MDMs that conditions each denoising step on the model's own previous clean-state predictions. The resulting method, called Self-Conditioned Masked Diffusion Models (SCMDM), requires minimal architectural change, does not introduce a recurrent latent-state pathway, does not rely on an auxiliary reference model, and adds no extra denoiser evaluations during sampling. This is an important departure from partial self-conditioning approaches which requires expensive model training from scratch. In particular, the paper shows that partial self-conditioning, including the commonly used 50% dropout strategy for training self-conditioned models from scratch, is suboptimal in the post-training regime. Instead, once the model's self-generated clean-state estimates become informative, the specialization to refinement is preferable to mixing conditional and unconditional objectives. SCMDM is evaluated across multiple domains, demonstrating consistent improvement over vanilla MDM baselines, achieving nearly a 50% reduction in generative perplexity on OWT-trained models (42.89 to 23.72), alongside strong improvements in discretized image synthesis quality, small molecular generation, and enhanced fidelity in genomic distribution modeling.

arXiv:2604.27014v1 Announce Type: new Abstract: The scarcity of high-quality annotated medical data, particularly in mental health, poses a significant bottleneck for training robust machine learning models. Privacy regulations restrict data sharing, making synthetic data generation a promising alternative. The use of Large Language Models (LLMs) in a data augmentation pipeline could be leveraged as an alternative in this field. In the proposed methodology, DeepSeek-R1, OpenBioLLM-Llama3 and Qwen 3.5 are used to generate synthetic mental health evaluation reports conditioned on specific International Classification of Diseases, Tenth Revision (ICD-10) codes. Because naive text generation can lead to mode collapse or privacy breaches (memorization), a comprehensive evaluation framework is introduced. The generated diagnostic texts are assessed across three dimensions: semantic fidelity, lexical diversity, and privacy/plagiarism. The results demonstrate that all models can generate clinically coherent, diverse, and privacy-safe synthetic reports, significantly expanding the available training data for clinical natural language processing tasks without compromising patient confidentiality.

Wired: Musk v. Altman: OpenAI lawyers claim Shivon Zilis, a longtime Musk employee and mother to four of his children, acted as a covert liaison between him and OpenAI  —  Messages presented at trial reveal how Zilis, the mother of four of Musk's children, acted as an intermediary between him and OpenAI.

Jasmine Sun / New York Times: The persistent notion that AI disruption could create a permanent underclass signals how much collateral damage AI companies might tolerate in pursuit of AGI  —  Most people I know in the A.I. industry think the median person is screwed, and they have no idea what to do about it.

Jason Beeferman / Politico: Sens. Kirsten Gillibrand and Dave McCormick introduce a bill banning members of the legislative and executive branches from trading on prediction markets  —  The bill comes two days after a U.S. soldier was arraigned in federal court on charges he used insider knowledge of the military's plans …

Users in India are embracing ChatGPT Images 2.0 for creative, personal visuals — from avatars to cinematic portraits.

Steve Jobs once told Tim Cook, "Just do what's right." Cook said he gave John Ternus advice on how to spend his time wisely. Getty Images Tim Cook has said that Steve Jobs told him never to ask what he would do but "just do what's right." Cook now leaves a piece of advice for John Ternus, the incoming CEO of Apple. "One of the most important decisions he’ll make is where to spend his time," Cook said. Steve Jobs once left salient advice for Tim Cook as the Apple cofounder prepared to hand over the reins 15 years ago. Now, the outgoing CEO has some wisdom to impart to his new successor, John Ternus. "What I've told him is that one of the most important decisions he’ll make is where to spend his time, and I would spend it where the greatest benefit to the company and the users are," Cook said during Apple's second-quarter earnings call on Thursday. "And never forget the North Star for the company. We're about making the best products in the world that really enrich other people's lives," Cook added. "And if you keep focusing on that, and make your decisions around that, it will produce a great business, and we'll be able to build more products and do it all over again." Cook, who will step down as CEO in September, has been at the helm of the Cupertino company since August 24, 2011, replacing Jobs due to the Apple cofounder's debilitating health. Jobs died on October 5, 2011, after battling a rare form of pancreatic cancer. During a memorial service, Cook revealed that one of Jobs' last pieces of advice to him was "to never ask what he would do. 'Just do what's right,'" Bloomberg reported in 2011. Cook said at the time that Jobs wanted to avoid the same pitfalls Disney experienced after its founder died and "everyone spent all their time thinking and talking about what Walt would do." That piece of advice guided Cook throughout his 15-year tenure, the CEO said during the Thursday earnings call. "Steve's advice to me lifted a huge burden," Cook said. "So that advice did well for me over the 15 years." Ternus is Apple's senior vice president of hardware engineering and a 25-year veteran at the company who has worked under Cook and Jobs. "I am filled with optimism about what we can achieve in the years to come, and I am so happy to know that the most talented people on earth are here at Apple, determined to be part of something bigger than any one of us," Ternus said in a statement posted on Apple's blog. "I am humbled to step into this role, and I promise to lead with the values and vision that have come to define this special place for half a century." Cook will remain at Apple as executive chairman of the board of directors. Read the original article on Business Insider

Mark Gurman / Bloomberg: Apple forecasts Q3 revenue above estimates and reports Q2 results above projections in China but below expectations in the Americas and Europe regions  —  Apple Inc. delivered a surprisingly strong revenue forecast for the third quarter, even as it warned that memory-chip costs will increase …

Apple CEO Tim Cook has been a regular sight in President Donald Trump's Washington. Brendan Smialowski/AFP/Getty Images Tim Cook said Apple plans to apply for a tariff refund — but will reinvest the funds in the US. President Donald Trump famously said he'd "remember" companies that don't seek tariff refunds. Cook has developed a reputation as a Trump whisperer, as his measured approach to the refunds shows. Tim Cook is positioning Apple to take advantage of tariff refunds as Donald Trump watches closely. The outgoing Apple CEO said on Thursday during the company's Q2 earnings call that Apple is "following the established processes" to apply for tariff refunds and plans to reinvest any money it recoups into US innovation and manufacturing. Those investments, he added, would be "in addition to our prior commitments in the US." Cook's comments show how Apple is navigating tariffs that have become both a cost issue and a political one — and underscore why he's earned a reputation as the Trump whisperer. Trump raised the stakes around those refunds in an April 21 appearance on CNBC's "Squawk Box." Asked about companies like Apple and Amazon, which had not yet announced plans to pursue refunds, Trump said the companies were "brilliant" if they didn't seek reimbursement, and their decision not to pursue tariff refunds served as evidence that they "got to know me very well." "I'll remember them," Trump added, signaling that the decision could carry political weight. Cook has maintained a notably close working relationship with Trump across both of his administrations, previously securing tariff exemptions for Apple products while pairing those efforts with high-profile US investment commitments. When Donald Trump said an Apple manufacturing plant had been built in Texas because of him, Apple CEO Tim Cook didn't correct him. MANDEL NGAN/AFP via Getty Images That relationship has helped Cook stand out among tech executives for his ability to navigate Trump's orbit without the same level of public friction, maintaining his direct line of communication with the White House while other tech leaders clashed more openly with the administration, Business Insider has reported. Apple's strategy — securing refunds while emphasizing domestic reinvestment — reflects Cook's careful balancing act between cost management and political optics. Other major companies are already cashing in on their refunds. Ford said a $1.3 billion tariff refund helped boost its quarterly profit, while General Motors expects a roughly $500 million windfall of its own following the Supreme Court ruling that struck down the levies, Business Insider previously reported. FedEx, UPS, and DHL have said they're also seeking tariff refunds and plan to pass that money back to customers who originally paid the fees. The companies said refunds will flow through once they receive funds from US Customs, a process expected to take roughly 60 to 90 days. Read the original article on Business Insider

Messages presented at trial reveal how Zilis, the mother of four of Musk’s children, acted as an intermediary between him and OpenAI.

Katie Tarasov / CNBC: Intel shares jumped 114% in April, hitting a record on April 24 and lifting its market cap past $470B, closing out the chipmaker's best month on record  —  Intel is on a winning streak unlike any since it became one of the first companies to go public on the Nasdaq nearly 55 years ago.

Sam Altman and Elon Musk Getty Images Elon Musk was back on the witness stand in his civil trial against OpenAI and its CEO, Sam Altman. Musk's money manager, Jared Birchall, testified about Musk's charitable donations. See who will be questioned next week when the trial continues. Day Four of the blockbuster civil trial in Elon Musk's lawsuit against Sam Altman began Thursday with Musk returning to the witness stand. Musk faced grilling by OpenAI and Altman's lead lawyer, William Savitt, and a brief round of questioning by an attorney representing OpenAI backer Microsoft, also a defendant. Jared Birchall, a former Morgan Stanley banker and Musk's money manager, was the second witness called by Musk's legal team. Here are the biggest takeaways from the trial, drawn from witness testimony and evidence presented in the California federal court. Musk said xAI is no threat Musk owns xAI, the maker of Grok, and the defense argues that his lawsuit over OpenAI's transition from a nonprofit to a for-profit entity was filed to take down a rival. Musk on Thursday sought to deflect those allegations by describing xAI as "the smallest" of the companies. He ranked his AI rivals in order of size, starting with Anthropic first, OpenAI second, Google third, and then "some Chinese open-source models." "xAI is very small," he said of his own company. "It is about a tenth of the size of that of OpenAI," Musk added. Musk has taken aggressive steps to shore up xAI, including merging it with his rocket company, SpaceX, earlier this year. He more recently announced a deal with AI coding startup Cursor, giving SpaceX the right to acquire it for $60 billion. Sam Altman and Greg Brockman arrive at the courthouse. Bloomberg/Getty Images Judge shut down talk of a robot apocalypse Musk, who is also the CEO of Tesla and SpaceX, has made multiple references to what he called an AI "Terminator" scenario during his time on the witness stand. On Thursday, US District Judge Yvonne Gonzalez Rogers put a stop to it. "We're not going to talk about extinction in this case," Gonzalez Rogers told Musk before Thursday's morning recess. The reprimand came after Musk answered questions about Tesla's robot-making business and his past comments about "building a robot army." "No, we do not make weapons," said Musk of Tesla, "We don't want a 'Terminator' situation." When asked about the situation he was referring to, Musk said, "In the movie it's not a good situation," adding that the "worst situation would be that AI kills us all, I suppose." William Savitt, Open AI's lawyer, cross-examined Elon Musk again. JOSH EDELSON / AFP via Getty Images Birchall questioned over OpenAI security Birchall, Musk's money manager, testified that Musk made $38 million in donations to OpenAI, which Musk alleges were improperly used to enrich some of the defendants. Birchall said Musk made "approximately 60 donations" to the ChatGPT maker as well as monthly rent contributions. On cross-examination, Savitt asked Birchall whether Musk ever gave him specific instructions or restrictions on how OpenAI should use the funds. "There was a greater understanding, as one would expect there to be, among Elon and other founding members of OpenAI," said Birchall. "I don't believe that 'for whatever use, at any time' was the intended use of those funds." Savitt also questioned Birchall about OpenAI's 2018 office-sharing agreement with Musk's Neuralink, which seeks to connect human brains to computers. The lawyer pointed to emails between Birchall and another OpenAI employee in which Birchall said that the building should have armed security at all times, citing OpenAI's growing profile. "They were sharing that cost with Neuralink," Birchall said of the security arrangement. Birchall testified he was xAI's corporate secretary "up until recent weeks." He has also held positions at Neuralink and The Boring Company, Musk's tunneling company. An Elon Musk sign sits in a bush outside the federal courthouse. JOSH EDELSON / AFP via Getty Images Looking ahead The first week of trial concluded with Birchall because the jury will not sit on Friday. Next week, the nine-person jury will hear from Greg Brockman, OpenAI's president and a cofounder. Brockman is considered a key witness in the case due to journal entries he wrote that shed light on the period before Musk left the company in 2018. "His story will correctly be that we weren't honest with him in the end about still wanting to do the for profit just without him," one Brockman diary entry reads, referring to Musk. "it'd be wrong to steal the non-profit from him. to convert to a b-corp without him," Brockman also wrote. "that'd be pretty morally bankrupt. and he's really not an idiot." OpenAI has said Brockman's diary entries were "cherry-picked" by Musk's legal team. Brockman's journal entries are among the evidence cited by the judge when she ordered a jury trial. Musk has asked that Brockman and Altman be removed as officers of the company and forfeit their equity if found liable. Read the original article on Business Insider

Cook warned that Apple is facing supply-chain headwinds from RAMageddon that could impact its business.

Subscription billing fintech Skio sold to its competitor Recharge in what was a healthy exit, according to its founder and former CEO.

Anhata Rooprai / Reuters: Western Digital reports Q3 revenue up 45% YoY to $3.34B, vs. $3.25B est., and forecasts Q4 revenue above estimates; WDC drops 7%+ after hours  —  Western Digital (WDC.O) forecast quarterly revenue above Wall Street estimates on Thursday, expecting strong demand for data storage …

<p><strong><a href="https://github.com/openai/codex/releases/tag/rust-v0.128.0">Codex CLI 0.128.0 adds /goal</a></strong></p> The latest version of OpenAI's Codex CLI coding agent adds their own version of the <a href="https://ghuntley.com/ralph/">Ralph loop</a>: you can now set a <code>/goal</code> and Codex will keep on looping until it evaluates that the goal has been completed... or the configured token budget has been exhausted.</p> <p>It looks like the feature is mainly implemented though the <a href="https://github.com/openai/codex/blob/6014b6679ffbd92eeddffa3ad7b4402be6a7fefe/codex-rs/core/templates/goals/continuation.md">goals/continuation.md</a> and <a href="https://github.com/openai/codex/blob/6014b6679ffbd92eeddffa3ad7b4402be6a7fefe/codex-rs/core/templates/goals/budget_limit.md">goals/budget_limit.md</a> prompts, which are automatically injected at the end of a turn. <p><small></small>Via <a href="https://twitter.com/fcoury/status/2049917871799636201">@fcoury</a></small></p> <p>Tags: <a href="https://simonwillison.net/tags/ai">ai</a>, <a href="https://simonwillison.net/tags/openai">openai</a>, <a href="https://simonwillison.net/tags/prompt-engineering">prompt-engineering</a>, <a href="https://simonwillison.net/tags/generative-ai">generative-ai</a>, <a href="https://simonwillison.net/tags/llms">llms</a>, <a href="https://simonwillison.net/tags/coding-agents">coding-agents</a>, <a href="https://simonwillison.net/tags/system-prompts">system-prompts</a>, <a href="https://simonwillison.net/tags/codex-cli">codex-cli</a>, <a href="https://simonwillison.net/tags/agentic-engineering">agentic-engineering</a></p>

Discussion | Link