Shift is offering limited-time free cleanings, with workers recording their tasks to train AI. Monika Skolimowska/picture alliance via Getty Images A startup is offering free home cleanings in New York City to train AI. Shift started offering the cleanings on Thursday as part of an effort to train robots. Human cleaners wear hat-mounted cameras that capture video for training AI, Shift said. Your dirty windows could help create AI-powered cleaning robots. Shift, an AI training startup, is offering free home cleanings in New York City, it said on Thursday. There's one catch: cleaners wear head-mounted cameras that record first-person footage while they scrub bathrooms, mop floors, and organize kitchens. The company said the footage will help train household robots and AI systems to complete chores autonomously in the future. "Your home. Cleaned for free," the company says on its website, which advertises services ranging from laundry folding and dishwashing to fridge organization and bathroom scrubbing. The offering is the latest example of the booming AI training space. Companies, from startups to bigger players like Uber and LinkedIn, are trying to get in on the action. And while many of the jobs training AI focus on white-collar work, some, like Shift's, attempt to train AI for physical tasks in the real world. Shift said the training data generated from routine household tasks is valuable enough to subsidize the cleaning service entirely. The startup said it operates across more than 15 countries and works with thousands of people to record videos for AI training. The company also has a form on its website for people interested in working for the service. Shift said on its website that it protects customers' privacy, "with any sensitive details blurred" in the video before the footage is used for AI training. And don't be ashamed if your apartment is particularly dirty, Shift says. In fact, grubbier might be better. "More challenging cleaning environments can be especially useful," an FAQ document posted on the company's website reads. "That said, cleaners may decline any specific task they are not comfortable performing." Do you have a story to share? Contact this reporter at abitter@businessinsider.com or via encrypted messaging app Signal at 808-854-4501. Use a personal email address, a nonwork WiFi network, and a nonwork device; here's our guide to sharing information securely. Read the original article on Business Insider

Sunjana Ramana is building her company, Flax, in the UK. Courtesy of Sunjana Ramana The "American dream" didn't last long for Sunjana Ramana. After 4 ½ years in the US and three failed attempts at the H-1B visa, she left for the UK. Ramana is now building a tech company, and she no longer ties her dreams to being in the US. This as-told-to essay is based on an interview with 27-year-old engineer Sunjana Ramana, who moved from New York City to London in February. The following has been edited for length and clarity. As a naive 22-year-old, living and working in America was my dream. I come from a middle-class household in Hyderabad, India. I was accepted to Columbia for a master's in electrical engineering, and I moved to New York in August 2021. I wanted to break into data science and engineering, and my focus was on meeting the right people in the US — those working at big companies with successful careers — so I could build my network. During my 4½ years in the US, I tried to get an H-1B visa, but wasn't successful after three attempts. I moved to London in early 2026, where I've been exploring entrepreneurship. Instead of tying my dreams to being in America, I've taken a more open-ended approach to my career and life. I felt the pressure of student debt I remember paying over $70,000 in fees to go to Columbia, and my living expenses were roughly $2,000 to $2,500 a month, which came with a lot of financial pressure, even though my family was helping to support me. I simultaneously worked three part-time jobs in addition to my coursework. My entire focus was, "I cannot fail." A lot of immigrant students come to the US carrying significant financial pressure, which can feel incredibly overwhelming at times. There's a mindset that failure simply isn't an option because so much is riding on the opportunity. I was disappointed that I never got a H-1B visa I came to the US on an F-1 student visa. After graduating in February 2023, I joined a global insurance company that I'd interned with as a student. As a STEM graduate, I was eligible for three years of work authorization in the US called Optional Practical Training. After I started working, I began the process of applying for the coveted H-1B, a type of employment visa that would allow me to continue working in the US. My employer put in three applications for me in 2023, 2024, and 2025. I received three rejections in a row. Most people are very hopeful about getting picked in the H-1B lottery. I felt the third year would be "my year." When I wasn't picked again, I was really disappointed. As an international student, you come to the US with the idea that this country has the best opportunities. You get to work with the best people and solve the best problems. You want to keep that streak going — to have continuity. Nobody wants to come out of the life that they've built. My STEM OPT ended in early 2026, and since I didn't get an H-1B visa, I had to leave the US. I zeroed in on London and have built my business here When I didn't get the H-B visa, I tried learning about my other options. For example, I attended an AI summit where I met people who had left the US for places like Germany, Australia, and the UK. I understood that if you really want to do something, there are opportunities everywhere. Finally, after researching all available options and zeroing in on London as my next destination, I received a High Potential Individual (HPI) visa, which is a two-year work visa in the UK. It's flexible because it's not tied to a single employer, and you don't need to secure a job before applying. It also enables self-employment with ease. In the US, I was always itching to build something in the AI space, so this felt like a good opportunity to finally pursue my entrepreneurial dream. I got accepted into a founder-in-residence program offered by Antler, a VC firm, and started working on early ideas and prototypes for my own company, Flax. We're a two-person tech company focused on helping to elevate founders and builders in the AI space. Ramana has had to start from scratch in London. Courtesy of Sunjana Ramana. Being in London has been a learning curve. I don't know anybody here, so I'm starting from scratch and rebuilding my network. It's a little uncomfortable, but I've always wanted to do something of my own, and my vision for my company has taken shape more clearly over time. A big milestone was hosting our first in-person AI hackathon in London, in partnership with Founders Factory and sponsored by AWS, which was also a key moment grounding me in the London ecosystem. I hope I'll have the freedom to go back to New York in the future if I want to. I was comfortable there, and it was a great place to begin my career. A week before I left for London, I went to the Brooklyn Bridge and felt very emotional at the thought of never coming back. I no longer tie my dreams to the US When I came to the US aged 22, my American Dream was all about career progression, networking, and global exposure, but today, as a 27-year-old, I see it was a mistake to tie those dreams to the country I was in. I want to go back and tell my younger self not to associate the "dream" with America, because it doesn't matter where you are. The dream should be aligned to you and your goals. Ramana has learned to trust the process. Courtesy of Sunjana Ramana One thing I've learned from my journey is that it's important to talk about your experiences and build your personal brand online. You never know who will reach out to help you. When I posted that I was moving out of the US, I was overwhelmed by the amount of online positivity. I landed coffee meetings with people who commented, welcoming me to London, which led to more opportunities. I've also learned to trust the process. When I didn't get the H-1B visa, I felt like it was the end of everything I'd built over three years. I realized later, though, that I had tunnel vision; there are many other options beyond this specific visa. No matter what, don't lose hope. Do you have a story to share about your US immigration journey? Contact the editors, Jane Zhang and Charissa Cheong, at janezhang@businessinsider.com and ccheong@businessinsider.com Read the original article on Business Insider

Cisco says AI transformation goes beyond bolting the technology onto existing workflows. Elena Podkina/Getty Images For Cisco, a nearly $500 billion company with over 80,000 employees, adopting AI is complicated. Its chief customer experience officer says the process is like "surgery without the drugs." She said adopting AI has forced Cisco to rethink how it does its work. When a corporate leader has to integrate AI into their giant global company with tens of thousands of employees scattered across offices around the world, there's probably going to be some discomfort. Cisco's chief customer experience officer, Liz Centoni, described the process as "surgery without the drugs." "It's painful," she told Business Insider. Cisco makes networking, security, and collaboration technology for companies, including routers, switches, cybersecurity tools, and services that help connect and protect corporate IT systems. So its customer support team is essential. Centoni, who leads Cisco's roughly 20,000-employee customer experience division, said she is now transforming it into an AI-native services organization. She said that adopting AI has required more than simply bolting it onto old workflows — something Cisco tried before realizing it only made flawed processes move faster. One example was customer support. Cisco first used generative AI to create case summaries when one support engineer handed a case to another, whether because of a shift change, availability, or the need for different expertise. The goal was to give engineers more context. However, Centoni said it only "annoyed our customers faster," because it made handoffs more efficient but did not address the bigger issue. "The outcome was never the handoff," she said. It was "how do I get to the right engineer the first time around?" That realization pushed Cisco to redesign the workflow, using "intelligent routing" to send cases to the right expert from the start. Cisco's customer experience division gets about 1.5 million support cases a year, and now, nearly 88% are routed to the right engineer the first time, Centoni said. A spokesperson for Cisco said the company now measures customer service success by the number of calls that require only one — or even zero — handoffs. The best places to integrate AI, Centoni said, are in repeatable workflows that can be performed autonomously with more than 90% accuracy. The company recently launched Cisco IQ, a digital interface for support and professional services that Centoni said is "designed as a single source of truth for customers to address recurring customer pain points." She said it helps detect preventable outages, redirects employees who spend too much time "interpreting data" rather than acting on insights, and reduces frustrating support calls. For Centoni, the test for any AI project is not just whether it improves efficiency. She said each initiative has to show what work it will stop, and whether it can grow revenue, expand margins, deepen customer trust, or help teams build what comes next. Read the original article on Business Insider

Vitalii Dodonov. Courtesy of Vitalii Dodonov Vitalii Dodonov started his career at Deloitte and then eBay before cofounding his business, Stan. A few years into the business, the cofounders vibe-coded Stanley, an AI tool for Stan, in two weeks. Stan hit $50,000 in monthly recurring revenue within six weeks and continues to grow. This as-told-to essay is based on a conversation with Vitalii Dodonov, a 29-year-old founder and CTO based in Toronto. It has been edited for length and clarity. I studied chemical engineering at the University of Alberta in 2014 after someone recommended it as a solid career path. During college, I worked in a mineral processing lab, where I observed people performing repetitive manual tasks. I thought there had to be a better way, so I began teaching myself to code on the side. Deloitte came to my campus to recruit. I was hired because I demonstrated hard work through my history of teaching myself to code. They offered me a job at the beginning of my fourth year in 2018 to work as a data scientist. I started working there three months after graduating. At Deloitte, I realized my passion was really in coding and software, so I kept building side projects on my own. One of them was Vhinny, a platform that aggregated financial information about publicly traded companies. My experience building Vhinny laid the groundwork for my current company, Stan, which brings in tens of millions in annual revenue. One of Stan's AI products, Stanley, was vibe-coded in just two weeks. After almost two years at Deloitte, I began seeking other opportunities I felt underpaid, and I knew the best market for my skill set was in software engineering, not consulting. I struggled because I didn't have formal software engineering credentials, and most Big Tech companies wouldn't consider me. I became aware of an open software engineering role at eBay, and it was a natural pull toward doing what I was most interested in. I got lucky by using LinkedIn's easy-apply feature for the eBay role. I failed the technical interviews with eBay My strength was discussing Vhinny. I built it over two years, working 5 to 9 a.m. and 5 to 9 p.m., teaching myself engineering. When the hiring manager asked me why I'm good at what I do, I pulled up Vhinny's architecture and gave a full demo, explaining in detail how it worked. I used skills generally practiced by senior and staff-level engineers — all self-taught from Google and YouTube, working backward from questions. They hired me as a senior software engineer in May 2020. During my time at eBay, I met my now cofounder, John, and we decided to build Stan, an all-in-one platform that enables creators to sell digital products through a single link in their bio. I quit eBay to work for Stan full-time in June 2021 I'm now the CTO. We spent the next several years building out our core creator-commerce platform. As we looked to better serve our customers, our mission became clear: help creators make a living working for themselves. We realized the biggest challenge isn't selling digital products — it's building an audience willing to buy them. We saw an opportunity to help creators build their audiences rather than treat them as a utility-driven monetization product. My cofounder and I locked in a flat for 14 days to vibe-code a product to do this from scratch I believe that to be creative, you need to step outside the box — a new environment helps you see things differently. We picked London because neither of us had been there. We wanted to target creators and entrepreneurs, so we focused on LinkedIn. I didn't write a single line of code; it was all me and the cursor. We vibe-coded Stanley — our AI "head of content" product — a fully functional tool in 14 days to help creators make high-performing content on LinkedIn. The platform uses AI to generate posts that sound like the user's own voice, then tracks which ones perform best by analyzing engagement data. It evolved from a writing assistant into a performance-driven tool that helps creators refine their content and consistently produce viral posts. We documented the entire process on YouTube and launched it to 250 paying customers on the first day. Stanley was quickly successful Within six weeks of its launch earlier this year, we hit $50,000 in monthly recurring revenue. Right now, Stanley is tracking for $3 million in annual recurring revenue: $1.5 million from Stanley Instagram, which we launched later, and $1.5 million from Stanley LinkedIn. Stan generates roughly $40 million in annual revenue. Our vision is to build the world's AI head of content for everyone. We see so many people doing remarkable work, but with little online presence. Building that presence is a full-time job, and many people don't have the time or skills to do it well. We believe that if we can give people a tool that lets them focus on what they do best, while an AI head of content grows their online footprint, it will create more opportunities and make them more successful. You don't need prior experience to start vibe coding, but you need a certain mindset One of the pitfalls of vibe coding is the belief that coding itself is the solution. The correct way to think about it is that it's just a way to deliver value to customers. Success comes down to understanding the problem you're solving and solving it for as many people as possible. Next, define your goal, whether it's building a $10 million business, quitting your job, or creating financial security for your family. Work backward from that goal and ask, "What do I need to do to get there?" From there, you can achieve anything you want. We live in a world where AI agents give individuals the firepower of an entire company. Getting there becomes a matter of effort. If you're willing to work backward and put in the work, almost nothing is out of reach. My ambition is to build a generational company that the world runs on in 30 to 40 years. I'm working on various aspects that position me to do that. Would I ever go back to corporate? Very unlikely. I don't see myself ever working for anyone else. Read the original article on Business Insider

Bloomberg/Getty Images Google is making its AI models cheaper and faster. More compute is helping it accelerate. It's a playbook that helped Google win in Search. As companies look more closely at their AI bills, Google is tapping a key advantage. While Anthropic hypes its unreleased Mythos AI model as dangerously powerful, Google is changing the conversation — to cost and speed. Google says its latest Gemini 3.5 Flash model rivals frontier offerings, while saving money for companies that are racking up huge bills by churning through billions of tokens, the core unit of AI usage. "Companies are already blowing through their annual token budgets and it's only May," Google CEO Sundar Pichai said recently. "If companies used a mix of Flash and other frontier models they could save a lot of money." The timing of Google's new model is no coincidence. As companies embrace token-hungry AI agents they're also paying closer attention to their bills. Meanwhile, smaller AI companies under pressure to generate revenue are cranking up the cost of their products, pushing customers to reconsider their AI spend. That presents an opportunity to win on value rather than raw capability. It's also where Google has an edge that will be hard for rivals to replicate — one it's been working on for a quarter-century. Flash Sale For the first three or so years, the generative AI wars were largely about who had the biggest and smartest model. Now, as the performance gaps between labs shrink, the advantage is shifting to infrastructure and inference, or how models are run. As OpenAI President Greg Brockman recently declared: "the model alone is no longer the product." A big reason for that shift is agents, which are becoming more useful and expensive to operate. Google knows how high the token burn is getting. Pichai recently noted that monthly usage of its AI products has increased sevenfold to 3.2 quadrillion tokens since last year. He also said that if the top customers of Google Cloud moved 80% of their AI workloads to a combination of Gemini 3.5 Flash and other frontier models, they could save more than $1 billion a year. Companies are noticing how much AI is adding up. Uber's COO recently said it was becoming harder to justify the company's ballooning AI costs. Venture capitalist Chamath Palihapitiya said in March that his company, 8090, was moving away from using Cursor because it was spending too much money on tokens. "As AI agents become more complex, long-running processes have become the norm," Dan Morgan, an analyst at Synovus Trust, told Business Insider. "This has created sticker shock at many organizations." Cost and ROI go hand in hand because it's hard make a profit in this space, said Morgan. For some companies, having access to models on the absolute frontier may no longer be necessary. Good enough may be good enough. This is where Google is well positioned. It has tighter control over the cost and speed of AI than most rivals because it owns the full stack — chips, data centers, cloud, the models, and many of the big applications on top. Google pays around 50% less (and possibly as much as 75% less) for its internal AI compute than rivals because it uses its own TPU chips and sources components directly from manufacturers, analysts at William Blair estimated earlier this month. OpenAI, on the other hand, pays Microsoft, Oracle, and other cloud giants a margin on every ChatGPT and Codex request, and those providers pay Nvidia for the GPUs that run it all. In fact, just about every company that isn't a hyperscaler is having to pay someone else for infrastructure right now. The Search playbook If compute is destiny, as OpenAI CEO Sam Altman likes to say, Google has spent more than 25 years sealing its fate. In 2006, Google Search commanded more than 40% of the market and was pulling away fast — not only because its results were good, but because Google was making its search engine faster and cheaper to serve. Google liked to boast about this by showing you the exact number of miliseconds it took to return answers. Rather than invest in expensive servers, Google built custom systems using cheap, off-the-shelf parts to maximize speed and keep costs down. Meanwhile, data from all those searches — which increased as Google became more popular — improved the engine, creating a flywheel that slowly strangled rivals like Yahoo. Google's results didn't need to be the absolute best, they just needed to be fast enough — and cheap enough to serve — that users would keep coming back for more. Google is creating a similar flywheel with Gemini, except now it also has a hugely successful search advertising business that can subsidize its AI efforts as rivals like OpenAI and Anthropic race for more money and compute. The search race was really an infrastructure race in disguise. Google is betting the AI race will be similar. Have something to share? Contact this reporter via email at hlangley@businessinsider.com or Signal at 628-228-1836. Use a personal email address and a non-work device; here's our guide to sharing information securely. Read the original article on Business Insider

PLUS: Use Codex to build a functional game in one prompt

Discussion | Link

Bloomberg: Sources: SpaceX is currently targeting an IPO valuation of at least $1.8T, down from a previous $2T+ target, after consultations with advisers and investors  —  SpaceX is currently targeting a valuation of at least $1.8 trillion in its initial public offering, according to people familiar with the matter …

Cissy Zhou / Nikkei Asia: Tencent bets on smaller AI models in the race with Chinese rivals, as EVP Dowson Tong says AI now contributes 20%+ of its revenue and 95%+ of new internal code  —  HONG KONG — As competition for AI users intensifies in China, Tencent is taking a different approach than rivals Alibaba …

Discussion | Link

Bloomberg: BYD announces the Xuanji A3 chip, which it calls China's most powerful chip for ADAS and the centerpiece of its new laptop-sized central computing platform  —  BYD Co., the world's largest electric vehicle maker, unveiled a series of technology advances including what it calls China's most powerful chip for self-driving cars.

80 Years. Countless Mathematicians. One AI Prompt.  For nearly 80 years, the best mathematical minds in … OpenAI’s AI Cracked an 80-Year Math Problem, Most Companies Missed the PointRead more The post OpenAI’s AI Cracked an 80-Year Math Problem, Most Companies Missed the Point appeared first on .

Financial Times: UK Chief Secretary to the Treasury Lucy Rigby warns that rejecting AI in public services means choosing “decline”, vowing to prioritize its Whitehall rollout  —  Newly appointed chief secretary to the Treasury Lucy Rigby wants to roll out technology across Whitehall

Yoolim Lee / Bloomberg: Samsung says it has started shipping its first 12-layer HBM4E samples to major clients; SK Hynix said in April that it aimed to ship HBM4E samples in H2 2026  —  Samsung Electronics Co. has begun shipping samples of the industry's most advanced memory to customers, taking an early lead …

arXiv:2605.28850v1 Announce Type: new Abstract: We study behavioral alignment and representation dynamics of large language model (LLM) agents in financial decision environments. Using TradeArena, an auditable trading-agent testbed with risk reports, execution simulation, memory, and replayable trajectories, we analyze how rationales, positions, and interventions evolve under market stress. We find measurable pre-failure signatures: planning embeddings drift from normal-state centroids, fused plan-risk representations separate normal from pre-drawdown states, and manifold diagnostics show effective-rank contraction before failures. To address small-sample and embedding-choice concerns, we use 80 rolling failure anchors across eight LLM trajectories and show that contraction persists across hash, LSA, Transformer, and white-box hidden-state probes. Stress tests with CoT-free target weights, lexical controls, OHLCV noise, and false-audit reports indicate that rationale-level contraction can vanish without rationales, while intent-space contraction may remain; lexical diversity does not collapse; and fused signatures remain informative under noise. We also find that structured risk feedback can act as an external alignment signal without fine-tuning, but not as a universal performance enhancer: true audit feedback improves calibration for some models, return and drawdown for others, and reveals cases where hidden or placebo feedback has higher short-horizon return but weaker alignment diagnostics. Finally, a 51-stock intraday experiment reveals a correlation blind spot: LLM rationales often justify concentrated exposure to coupled assets that the risk layer repeatedly clips, with a rolling Markowitz baseline as a covariance reference. These results support a research claim rather than a profitability claim: auditable risk feedback and representation trajectories reveal when LLM financial reasoning is aligning, drifting, or failing.

arXiv:2605.28867v1 Announce Type: new Abstract: Generating high-quality time-series data is challenging because real-world signals often exhibit multimodal patterns and multiscale dynamics, including oscillations and high-frequency variations. Flow Matching (FM) offers an efficient alternative to diffusion models, but practical implementations typically rely on a single finite-capacity global vector-field estimator. In such heterogeneous temporal distributions, distinct regimes may pass through nearby flow states while requiring incompatible conditional velocities. A monolithic estimator trained with the standard $\ell_2$ velocity-matching objective may therefore learn an overly smoothed approximation of the local transport field. This estimator-level smoothing can attenuate branch-specific dynamics, leading to spectral distortion and poor mode coverage. To address this, we propose PrismFlow, a new FM method with Koopman-inspired dynamical experts. Each expert learns residual corrections in a latent space where local nonlinear temporal evolution can be approximated by linear transitions. We further propose a confidence-aware Winner-Take-All (WTA) objective that updates only the expert best aligned with each sample while masking gradients to the others, encouraging mode-specific specialization. During sampling, the selected expert adds a residual dynamical correction to the global transport field, preserving FM stability while recovering fine-grained and high-frequency temporal structures. Across various benchmarks, PrismFlow effectively mitigates the spectral contraction in standard FM and achieves state-of-the-art performance, with a 15.6% gain in Context-FID and a 38.6% improvement in Discriminative Score, while remaining robust in low-data settings and effective for forecasting and imputation.

arXiv:2605.29012v1 Announce Type: new Abstract: Diffusion-based and iterative methods have become effective tools for solving imaging inverse problems. Their reconstruction process naturally forms a trajectory of intermediate estimates. Although these intermediate estimates define a reconstruction trajectory, most methods do not explicitly regularize the transitions between consecutive states. To address this limitation, we introduce TRACE, a training-free TRAjectory-Constrained rEconstruction framework that stabilizes the reconstruction path by coupling adjacent states along the trajectory. This gives a trajectory-level model that can be interpreted as a sequence of proximal updates. Since the exact proximal update is generally intractable, we approximate it with a neural mapping. This yields a diffusion-like reconstruction process with an explicit coupling between neighboring states. We provide a stability analysis showing that temporal coupling bounds trajectory variation and that this control is preserved under untrained network updates. Experiments on linear and nonlinear image reconstruction tasks show that TRACE improves reconstruction quality. Trajectory-level analyses and ablations confirm that temporal coupling directly affects state transitions along the reconstruction path.

arXiv:2605.28839v1 Announce Type: new Abstract: Knowledge editing methods such as ROME and MEMIT update factual associations in transformer models by modifying MLP weights. While evaluated mainly by output behavior, their internal mechanism remains underexplored. We investigate whether edits rely on a common mechanism, regardless of which fact is modified. Despite fact-specific weight changes, we argue that ROME and MEMIT target the same subset of weights critical for maintaining edits. To isolate this subset, we train a compact binary mask over the edited weights. The mask reverses 80% of edits on the training set and over 70% on the test set, confirming that diverse edits share a common functional structure. Our analysis reveals that the mask reverses edits by eliminating overattention in later layers. Additionally, we show that injecting the mask during editing drops editing success from 98% to 38%, demonstrating that this mechanism is necessary for edits to succeed. Our finding that edits suppress rather than overwrite knowledge explains why ROME and MEMIT fail to propagate changes to related facts. The identified common functional subspace informs detection and defense against unwanted edits.

arXiv:2605.28860v1 Announce Type: new Abstract: Fine-tuning large language models (LLMs) frequently induces catastrophic forgetting of prior capabilities. Recent work has shown that reinforcement learning (RL) retains prior capabilities more effectively than supervised fine-tuning (SFT), attributing this to policy-gradient updates remaining closer to the base policy \cite{shenfeld2025rl}. We extend this behavioral account to the mechanistic level and ask whether RL's advantage is mirrored by stronger preservation of internal computational circuits. We introduce differential circuit vulnerability, a head-level measure of how much a circuit degrades under fine-tuning, and use it to compare RL and SFT on Qwen2.5-3B-Instruct adapted to scientific question-answering. We find a clear mechanistic trade-off: SFT adapts more rapidly to the target task but produces substantially greater circuit disruption and forgetting of prior capabilities, whereas RL preserves a larger fraction of the base circuit at the cost of slower task adaptation. These findings suggest that circuit preservation may help explain why RL is more robust to catastrophic forgetting. We released our code here: https://github.com/rl-sft-circuit-research/differential-circuit-vulnerability.

arXiv:2605.28862v1 Announce Type: new Abstract: Drug discovery is a lengthy and resource-intensive process composed of multiple stages. Among these stages, lead optimization plays a critical role in transforming early hit compounds into viable drug candidates. This stage requires improving ADMET-related properties through subtle structural refinement while preserving key molecular substructures responsible for binding affinity to disease targets. Recent advances in artificial intelligence have shown promise in accelerating various aspects of drug discovery; however, most existing approaches to lead optimization rely on one-step molecular optimization, which fail to account for the long-term consequences of sequential design decisions. To address this limitation, we propose TRACE, a trajectory-aware, LLM-reasoning agent for molecular lead optimization that formulates tool selection as a sequential decision-making problem over action trajectories. Given a lead molecule and an optimization objective, TRACE makes trajectory-aware decisions over molecular optimization tools, enabling forward-looking refinement under structural constraints. Experiments on multiple ADMET optimization tasks show that our agent achieves higher optimization success, larger property improvements, and higher validity, while preserving molecular similarity compared to baseline models.