arXiv:2605.21573v1 Announce Type: new Abstract: We introduce Lens, a 3.8B-parameter T2I model that achieves performance competitive with, and in several cases surpassing, state-of-the-art models with more than 6B parameters across various benchmarks, while requiring significantly less training compute. For example, Lens requires only about 19.3% of the training compute used by Z-Image. The training efficiency of Lens stems from two key strategies beyond its compact model size. First, we maximize data information density per training batch by (i) training on Lens-800M, a dataset of 800M densely captioned image-text pairs whose captions are generated by GPT-4.1 and contain approximately 109 words on average, providing richer semantic supervision than conventional short captions, and (ii) constructing each batch from images with multiple resolutions and diverse aspect ratios, thereby enlarging the effective visual coverage of each optimization step. Second, we improve convergence speed through careful architectural choices, including adopting a semantic VAE that provides better latent representations and employing a strong language encoder that accelerates optimization while enabling multilingual generalization from English-only training data. After pre-training, we apply RL with taxonomy-driven prompts (Lens-RL-8K) and structured reward rubrics to suppress artifacts and improve visual quality, a reasoner module with training-free system prompt search to better align user requests with the model, and distillation-based acceleration for 4-step inference. Through efficient training and systematic optimization, Lens generalizes to arbitrary aspect ratios from 1:2 to 2:1 and resolutions up to 1440^2, and supports prompts in several commonly used languages. Thanks to its compact size, Lens generates a 1024^2 image in 3.15 seconds on a single NVIDIA H100 GPU, while its distilled turbo version performs 4-step generation in 0.84 seconds.

arXiv:2605.21726v1 Announce Type: new Abstract: The generative nature of Large Language Models (LLMs) is reflected in the conditional probabilities they compute to sample each response token given the previous tokens. These probabilities encode the distributional structure that the model learns in training and exploits in inference. In this work, we use these probabilities to situate LLMs within the mathematical theory of stochastic processes. We use this framework to design a model-agnostic probabilistic token attribution measure, using Bayes rule to invert the next-token log-probabilities so as to capture the models internal representation of the distribution over token sequences. The representation is independent of the models computational structure. This representation yields the conditional probability of the response given the prompt, and of the response given the prompt with a token marginalized away. Our attribution score is the log of the ratio of these probabilities. We further compute the entropies of a single prompts token distributions, conditioned on the remaining context. The interplay between entropy and attribution score sheds light on LLM behavior. We evaluate 8 models across 7 prompts and investigate anomalies, token sensitivity, response stability, model stability, and training convergence, thereby improving interpretability and guiding users to focus on uncertain or unstable parts of the generation.

arXiv:2605.21713v1 Announce Type: new Abstract: How can we distinguish whether a peer review was written by a human or generated by an AI model? We argue that, in this setting, authorship should not be attributed solely from the textual features of a review, but also from the ideas, judgments, and claims it expresses. To this end, we propose Sem-Detect, an authorship detection method for peer reviews that operationalizes this principle by combining textual features with claim-level semantic analysis. Sem-Detect compares a target review against multiple AI-generated reviews of the same paper, leveraging the observation that different AI models tend to converge on similar points, while human reviewers introduce more unique and diverse ones. As a result, Sem-Detect is able to distinguish fully AI reviews from authentic human-written ones, including those that have been refined using an LLM but still reflect human judgment. Across a dataset of over 20,000 peer reviews from ICLR and NeurIPS conferences, Sem-Detect improves over the strongest baseline by 25.5% in TPR@0.1% FPR in the binary setting. Moreover, in the three-class scenario, we empirically show that LLM refinement preserves the semantic signals of human reviews, which remain distinct from the patterns exhibited by fully AI-generated text; as a result, fewer than 3.5% of LLM-refined human reviews are misclassified as AI-generated.

arXiv:2605.21712v1 Announce Type: new Abstract: Transportation safety analysis requires integrating crash records, roadway attributes, and geospatial data through GIS-based workflows, but access remains uneven across agencies and community stakeholders. Technical prerequisites create a gap between analytical tools central to safety planning and the practitioners able to use them. Local agencies, school committees, and residents may have safety concerns but limited capacity to retrieve, filter, map, and analyze relevant data. Generative AI offers a way to narrow this divide, but its public-sector use raises questions about reliability, reproducibility, and governance. This paper presents a schema-grounded natural language interface for transportation safety analysis, using a large language model (LLM) to interpret user intent while preserving deterministic, reviewable execution against an authoritative database. User queries are translated into structured semantic frames, validated by a rule-based layer, compiled into a typed directed acyclic graph of spatial operations, and executed against a PostGIS database. This bounded design separates language interpretation from deterministic execution, keeping results reproducible and schema-grounded while removing access barriers. The framework is evaluated using a statewide Massachusetts transportation safety database integrating crash records, roadway attributes, and geospatial layers including schools, bus stops, crosswalks, and municipal boundaries. All queries executed successfully; the validation layer corrects errors in 29% of evaluation queries, reflecting the gap between flexible natural language and strict schema-grounded requirements. The results suggest that combining natural language accessibility with deterministic execution is a practical direction for broadening access to transportation safety data, with implications for trustworthy AI in public-sector planning.

arXiv:2605.21491v1 Announce Type: new Abstract: As language models accelerate scientific research by automating hypothesis generation and implementation, a new bottleneck emerges: evaluating and filtering hundreds of AI-generated ideas without exhaustive experimentation. We ask whether LMs can learn to forecast the empirical success of research ideas before any experiments are run. We study comparative empirical forecasting: given a benchmark-specific research goal and two candidate ideas, predict which will achieve better benchmark performance. We construct a dataset of 11,488 idea pairs grounded in objective outcomes from PapersWithCode. While off-the-shelf 8B-parameter models struggle (30% acc.), SFT dramatically boosts performance to 77.1%, outperforming GPT-5 (61.1%). By framing evaluation as a reasoning task via Reinforcement Learning with Verifiable Rewards (RLVR), we train models to discover latent reasoning paths, achieving 71.35% acc. with interpretable justifications. Through additional ablations and out-of-distribution tests, we show robustness to surface-level heuristics and transfer to both a cross-domain time-split test set and an independently constructed test set. Our results demonstrate that compute-efficient small language models can serve as effective, objective verifiers, offering a scalable path for autonomous scientific discovery.

arXiv:2605.21516v1 Announce Type: new Abstract: Harness engineering has emerged as an important inference-time technique for large language model (LLM) agents, aiming to improve long-term performance through task decomposition and guided execution. However, more elaborate harnesses are not uniformly better: increasing decomposition or guidance can sometimes improve execution, but can also reduce final task success. We study harness design through the lens of inference-time trajectory alignment. This perspective separates harness into two mechanisms: task decomposition, which structures a task into sub-goals, and guided execution, which reshapes local action distributions during execution. This decomposition allows us to quantify how workflow granularity, retry budgets, and guidance-induced action reweighting shape the performance limits of harness design. It further reveals concrete failure modes, including over-decomposition, over-pruning, and hallucinated execution. We validate these predictions through controlled synthetic experiments and real terminal agent benchmarks. Inspired by the theory, we further show that effective harnesses can be partial: specifying only the initial steps and leaving the remaining execution to agent can achieve higher pass rate than fully structured workflows.

arXiv:2605.21492v1 Announce Type: new Abstract: No feature ranking can be simultaneously faithful, stable, and complete when features are collinear. For collinear pairs, ranking reduces to a coin flip. We prove this impossibility, quantify it for four model classes, resolve it via ensemble averaging (DASH), and machine-verify it with 305 Lean 4 theorems. We characterize the complete attribution design space: exactly two families of methods exist -- faithful-complete methods (unstable, with rankings that flip up to 50% of the time) and ensemble methods like DASH (stable, reporting ties for symmetric features) -- and no method lies outside this dichotomy. The impossibility is quantitative: the attribution ratio diverges as 1/(1-rho^2) for gradient boosting, is infinite for Lasso, and converges for random forests. DASH (Diversified Aggregation of SHAP) is provably Pareto-optimal among unbiased aggregations, achieving the Cramer-Rao variance bound with a tight ensemble size formula. In a survey of 77 public datasets, 68% exhibit attribution instability. Switching to conditional SHAP does not escape the impossibility when features have equal causal effects. The framework includes practical diagnostics -- a Z-test workflow and single-model screening tool -- and has direct consequences for fairness auditing: SHAP-based proxy discrimination audits are provably unreliable under collinearity. The design space theorem, diagnostics, and impossibility are mechanically verified in Lean 4 (305 theorems from 16 axioms, 0 sorry) -- to our knowledge, the first formally verified impossibility in explainable AI.

arXiv:2605.21611v1 Announce Type: new Abstract: We introduce spatially grounded contextual image generation, a controllable image generation task that reframes the conditioning paradigm. Instead of supplying a reference image and a global text prompt through two separate encoders, one for vision and one for language, UniVL is trained to bind semantics to spatial locations directly from a single unified visual input, where the textual instruction is rendered onto the spatial mask. This removes the need for a standalone text encoder at inference time. The resulting model supports contextual image generation by following user-specified instructions about what should appear where, while substantially reducing computation. To address this task, we propose a framework in which the UniVL encoder, adapted from an optical-character-recognition-pretrained backbone, reads the unified condition optically and produces a UniVL embedding, fVIL, that fuses visual and semantic intent with spatial locations in a single token sequence. A two-stage pipeline first aligns UniVL with the VAE embedding space and then conditions a pretrained diffusion backbone entirely on UniVL embeddings, eliminating the standalone text encoder, such as T5. Although this reframing uses a deliberately minimal text interface, it yields strong empirical gains. On UniVL-ImgGen, a benchmark of 477K mask-annotated images that we construct for training and evaluation, UniVL improves image quality over text-prompted baselines, reducing FID from 14 to 11 and increasing PSNR from 16 to 20. It also eliminates the text encoder entirely, reducing inference TFLOPs by up to 52% and runtime by up to 44%. Additional ablation studies validate the contributions of the proposed components, paving the way for efficient, spatially grounded image generation with a unified conditioning paradigm.

Discussion | Link

Nikolaos Dimou/Pexels, Kostiantyn Klymovets/Pexels, The Conversation, CC BY-SA Another day, another literary scandal involving AI. It has been alleged that the judges of the Commonwealth Short Story Prize have been duped by an author using AI in his winning entry. Jamir Nazir’s The Serpent in the Grove, which won for the Caribbean region, was then published in leading literary magazine Granta, along with other winning entries. Almost immediately, it attracted accusations of being AI generated. Users on X posted screenshots of reports from AI detection tool Pangram, which claim 100% of the text was AI authored. Of course, the reliability of such tools can’t be guaranteed. Sigrid Rausing, publisher of Granta, which was not involved in selecting the story, said in a statement that she showed the story to Claude.ai and asked if it was A.I.-generated. “The response was long, concluding that it was ‘almost certainly not produced unaided by a human.’” Again, this may tell us nothing. The New York Times and the Guardian have contacted Nazir for a response to the allegations, but they report he has not responded. Where does this leave literary prizes, literary writing, and the literary short story? Razmi Farook, the Commonwealth Foundation’s director general, has defended the integrity of the prize’s judging process. Using AI checkers on unpublished submissions “would raise significant concerns surrounding consent and artistic ownership”. Competitions like the Commonwealth Prize, she notes, must operate on a “principle of trust”. I agree stories should be received on a principle of trust – but that may be hard to maintain. This year has already seen revelations of significant AI use in a New York Times book review and a debut horror novel. Maybe standards for short fiction will need to shift. ‘Like sunrise over a sink’ The Commonwealth Writers Prize is open to submissions from adult citizens of Commonwealth member states. It typically attracts thousands of submissions each year. Judging panels for each of the award’s regions (Africa, Asia, the Caribbean, Europe and Canada, and the Pacific) decide on regional winners, who are all published in Granta. An overall winner is also chosen. The Serpent in the Grove was praised by the regional judging panel for its “vivid, lush imagery” and “quiet authority”. It tells the story of a young, unhappily married couple, living in poverty in their rural village, alongside their older, gossipy neighbour. A near – though not entirely accidental – brush with death unexpectedly unites the three of them, creating an awareness of the hidden, subterranean stories that shape their lives and locality. Writers, literary critics and commentators on X and other platforms have derided what they see as its “obvious” AI writing syntax. The frequent use of “not x, but y” sentence structures and unusual word repetition have been presented as textual evidence. The story also includes, to put it charitably, a few choice sentences you’d have to hope were written by a robot. As Booker Prize winning author Marlon James put it on Facebook: Forget AI for a minute. A story won an International Competition with a line like this: “The girl smiled like sunrise over a sink.” But AI usage is difficult to prove. Detection tools are not reliable and have been known to hallucinate their results. Just because the story feels AI generated to many readers, with its somewhat overworked, metaphor-laden prose, hazily dramatised action and slippery refusal to commit to a concrete subject or theme doesn’t necessarily make it the case. It eventually concludes with a string of poetic aphorisms: A story is a well Ultimately, Rausing has taken a noncommittal stance, stating “perhaps we will never know” if the story was authored by AI. A new AI milestone? If the Serpent in the Grove is AI generated, perhaps a new milestone has been reached, with AI now managing to produce the kind of “quotidian, plotless, moment-of-truth revelatory story” that United States author Michael Chabon warned was coming to dominate publications and prize lists. Regardless of its provenance, I suspect that if you entered the prompt “write a short story that will get at least a regional shortlisting for an international prize” into the AI platform of your choice, something like Serpent in the Grove would be the likely result. Everything about it, from its frequently solid, if routine, descriptive passages (“Inside, air clung thick as porridge skin: damp earth, woodsmoke, and the sour tang of fermenting cocoa”) to its occasionally sharp observations (“Men who set traps plan for silence, not for the squeal.”) to its outright howlers (“She had the kind of walking that made benches become men”) feels true to the form. And like something that could be created by an authentically human literary mediocrity. It’s every bit of worthy filler you’ve found padding out the middle of a debut collection published straight out of a prestigious MFA (Master of Fine Arts) writing workshop. It’s every short story you’ve ever skipped past in the New Yorker to get to the film reviews and the caption competition. What’s the solution? Certainly, there is still at least a possibility that this small scandal may fall into the time-honoured tradition of the literary hoax. Despite the bio note by Nazir, a 63-year old resident of Trinidad and Tobago, claiming to be “a prolific poet and author”, readers have had difficulty tracking down any other works. His LinkedIn profile reveals he has previously written on the possibility of AI eliminating jobs and the AI arms race. The Serpent in the Grove may yet prove to be a parody or a warning. Advances in AI platforms and composition have often been oddly benchmarked through their engagements with the form of the short story. The launch of ChatGPT saw a major science fiction magazine having to temporarily close submissions due to being bombarded with AI-generated stories. Last year, Sam Altman of Open AI triumphantly claimed to have been “struck” by how his AI model “got the vibe of metafiction so right” in following his prompt to write “a metafictional literary short story about AI and grief”. And AI generation can now clearly produce the kind of technically competent creative writing that The Serpent in the Grove exemplifies (the occasional glories of a smile that is “like sunrise over a sink” aside). Perhaps the solution is to more actively look for new forms and structures for the short story. Or even just make a return to concrete detail, grounded specific action, plot. The measure of a successful story is its capacity to surprise us, however it is created. Julian Novitz had a story shortlisted for the Pacific region of the Commonwealth Short Story Prize in 2014 and subsequently published in an anthology of writing from the Prize.

Discussion | Link

Axios: Sources: Trump delayed signing the AI EO because “he just hates regulation”; he spoke with Mark Zuckerberg, Elon Musk, and David Sacks just prior to the delay  —  Everything seemed set for a photo op of tech and AI CEOs who would surround President Trump Thursday as he signed …

Wired: Socket: TeamPCP, the gang claiming GitHub's repositories breach, also executed 20 “waves” of supply chain attacks recently, compromising 500+ pieces of software  —  GitHub is just the latest victim of TeamPCP, a gang that has carried out a spree of software supply chain attacks that has impacted hundreds of organizations.

Engineers could make another attempt to launch Starship as soon as Friday evening.

ClickUp laid off 22% of its workforce and pledged million-dollar salaries for high-impact employees. Business Wire/AP ClickUp laid off 22% of its workforce and pledged million-dollar salaries for high-impact employees. CEO Zeb Evans emphasizes AI adoption at ClickUp, aiming to reach "100x output." Tech layoffs are surging while top AI talents are landing larger paychecks. ClickUp is taking a winner-takes-all approach to its workforce. Zeb Evans, CEO of the cloud-based productivity platform, wrote on social media on Thursday that the company has cut 22% of its workforce, while simultaneously promising that remaining employees could earn million-dollar salaries. "This wasn't about cutting costs," Evans wrote on X. "Most savings from this change will flow directly back into the people who stay. We'll be introducing million-dollar salary bands." "If you create outsized impact using AI, you'll be paid outside of traditional bands," Evans added. The layoffs come after months of rapid AI adoption inside the company, which centralizes enterprise workflows and communication. Andy Cabasso, a growth operations manager at ClickUp, recently told Business Insider that the company has a mandate to use AI agents more and that it fosters a culture of sharing AI workflows. Cabasso said he oversees 37 AI agents himself. As of 2021, ClickUp was valued at around $4 billion. According to a 2023 company blog post, it has over 1,000 employees. In his post on X, Evans said that the company's new goal is to reach "100x output," and added that, rather than gradually improving existing systems, ClickUp needs to rebuild itself from the ground up around AI. Evans' post also separated the company's future workforce into three types: "builders," "system managers," and "front-liners." According to his description, the first two groups will primarily oversee AI systems and automate parts of their own jobs, while the latter group will focus on customer relationships and represent "the human touch." The layoff reflects a broader trend sweeping the tech industry, where hundreds, if not thousands, of roles are eliminated at a time, while a small pool of top AI talent is rewarded with unprecedented paychecks. Meta has notably waged a multibillion-dollar battle to poach talent from the likes of OpenAI over the past year, while recently laying off over 8,000 employees. Similarly, Amazon, Cloudflare, and Atlassian have all seen layoffs this year, which company leadership has directly tied to AI adoption. Read the original article on Business Insider

Chloe Taylor / CNBC: Shares of quantum computing companies surged Thursday after the US government announced grants with equity stakes: D-Wave closed up 33%, Rigetti 30%, IBM 12%  —  Quantum computing shares popped on Thursday, as the U.S. government said it would award $2 billion in grants to nine firms operating in the space.

Global affairs chief Chris Lehane wants to tone down the debate over AI’s societal impacts—and get states to pass laws that won’t derail OpenAI’s meteoric rise.

OpenAI is named a leader in the 2026 Gartner Magic Quadrant for Enterprise AI Coding Agents, with Codex recognized for innovation and enterprise-scale deployment.

How Virgin Atlantic used Codex to ship its revamped mobile app on a fixed holiday travel deadline, reaching near-total unit test coverage and zero P1 defects.

Yusuf Mehdi, Microsoft executive vice president and consumer chief marketing officer. JASON REDMOND/AFP via Getty Images Microsoft executive Yusuf Mehdi is transitioning to a new role and eventually leaving the company. That's according to an internal memo viewed by Business Insider that Mehdi sent Thursday afternoon. Mehdi wrote that he will "help reimagine Windows for the agentic era" through the next fiscal year. Yusuf Mehdi, a 35-year Microsoft veteran who has been its commercial chief marketing officer since 2023, will leave the company after the next fiscal year, according to an internal memo viewed by Business Insider. "After 35 extraordinary years at Microsoft - years filled with adventure, challenge, reinvention, and innovation - I've decided the time is right to begin planning for my next adventure," Mehdi wrote in the memo. "I will work through the next fiscal year to help reimagine Windows for the agentic era, grow Microsoft 365 services, and bring our One Copilot vision to life." Mehdi leads product marketing for Microsoft AI and Copilot. He is the latest Microsoft executive whose role has changed amid CEO Satya Nadella's AI reboot. For example, Microsoft recently promoted Judson Althoff, its longtime sales chief, to an expanded role as CEO of the company's commercial business, and it moved its AI CEO, Mustafa Suleyman, to its superintelligence team, Business Insider previously reported. Microsoft did not immediately respond to a request for comment. Read the memo below: Team, After 35 extraordinary years at Microsoft - years filled with adventure, challenge, reinvention, and innovation - I've decided the time is right to begin planning for my next adventure! It's an exciting but difficult decision, as Microsoft has been far more than a place to work. It has truly been the canvas for my life's work. I've had the privilege of being a part of some of the most consequential shifts in technology - from the rise of Windows and the early Internet, to search, gaming, devices, and now one of the most profound platform transitions yet: AI. But while the products and platforms have been remarkable, the most meaningful part of this journey has always been the people. The teams I've worked alongside, learned from, built with, and grown with are what have made this experience so special. As I thought about this decision, one thing was crystal clear: I want to ensure I have the time and space to set the team - and our mission - up for continued success. We are in the middle of an incredibly important moment for Microsoft and for our customers. With that in mind, and in alignment with Satya and Takeshi, I will work through the next fiscal year to help reimagine Windows for the agentic era, grow Microsoft 365 services, and bring our One Copilot vision to life. Those of you who know me, know this means I'll be fully engaged, likely more intensely than ever! I've always believed the right way to finish is the same way I've always tried to lead: with urgency, ambition, and a commitment to leave things stronger than I found them. There will be time later to reflect and celebrate, but for now, it's full speed ahead on our mission. I'm genuinely thrilled about the year ahead and grateful for the opportunity to spend it building alongside the teams here at Microsoft in service of our customers. With deep gratitude, Yusuf Have a tip? Contact this reporter via email at astewart@businessinsider.com or Signal at +1-425-344-8242. Use a personal email address and a nonwork device; here's our guide to sharing information securely. Read the original article on Business Insider