Franziska Bell, PhD

New work on Rethinking Thinking Tokens: LLMs as Improvement Operators: https://lnkd.in/errhNuuz Reasoning training encourages LLMs to produce long chains of thought (CoT), improving accuracy via self-checking but increasing context length, compute cost, and latency. This work studies whether frontier models can achieve better trade-offs, higher accuracy with lower cost. The paper develops a simple yet effective Parallel-Distill-Refine (PDR) procedure: Generate diverse drafts in parallel, Distill them into a compact textual workspace, and Refine conditioned on this workspace. This decouples context length from total token count, allowing control over compute via parallelism. PDR yields higher accuracy than long CoT at lower latency. Training an 8B model with RL to align with PDR further shifts the Pareto frontier. On math benchmarks, PDR achieves +11% (AIME 2024) and +9% (AIME 2025) over single-pass baselines. With Lovish Madaan, Aniket Didolkar, Suchin Gururangan, John Quan, Ruan Silva, Manzil Zaheer, Sanjeev Arora, and Anirudh Goyal.

80

1 Comment

I've been following the LLM field on synthetic personas for some time now. It's truly interesting, both as an efficiency play, but also to unlock deep, rapid insights for early product testing and prototyping. In this recent example, PyMC Labs and Colgate-Palmolive developed semantic similarity rating (SSR), achieving 90% of human test-retest reliability across 57 personal care product surveys (9,300 responses). The method maps LLM-generated text to Likert scales, enabling scalable concept testing without traditional panel biases or billion-dollar survey costs. Amazing. Source [arXiv:2510.08338]

33

6 Comments

Evidence-Anchored Reasoning is rapidly becoming critical for trustworthy AI. This approach moves beyond simply *getting the right answer* to demanding that AI systems explicitly demonstrate *how* they arrived at that conclusion, linking outputs to verifiable source material. This isn’t just about accuracy; it’s about auditability, reducing hallucination, and building confidence in automated decision-making. Imagine risk management systems that flag not just potential fraud, but *why* they suspect it, or AI copilots that justify their recommendations with clear evidence. Scaling AI responsibly requires this level of transparency. 🔍💡📈

1

📑 Sharing the most discussed AI paper right now: Chinese researchers claim they have identified a physical mechanism behind LLM hallucinations. If validated, this has direct implications for industrial AI reliability. A recent arXiv paper (2512.01797) describes a small subset of so-called “H-neurons” that activate when a model is uncertain but still generates a confident answer. In other words: 𝗛𝗮𝗹𝗹𝘂𝗰𝗶𝗻𝗮𝘁𝗶𝗼𝗻 𝗺𝗮𝘆 𝗻𝗼𝘁 𝗯𝗲 𝗿𝗮𝗻𝗱𝗼𝗺𝗻𝗲𝘀𝘀. 𝗜𝘁 𝗺𝗮𝘆 𝗯𝗲 𝗮 𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗮𝗹 𝗯𝘆𝗽𝗿𝗼𝗱𝘂𝗰𝘁 𝗼𝗳 𝘁𝗿𝗮𝗶𝗻𝗶𝗻𝗴 𝗼𝗯𝗷𝗲𝗰𝘁𝗶𝘃𝗲𝘀. Sounds like a fix... but not so fast: You cannot simply remove these neurons. Large language models are highly interconnected systems. Neurons are polyfunctional and participate in multiple reasoning pathways. Removing the 'wrong' neurons may reduce hallucinations, but it can also lead to a high level of unpredictable cascading side effects: 🔹 Also degrade reasoning flexibility 🔹 Increase excessive “I don’t know” responses 🔹 Distort unrelated capabilities The exciting part is not that we found “bad neurons.” 𝗜𝘁’𝘀 𝘁𝗵𝗮𝘁 𝘄𝗲’𝗿𝗲 𝗳𝗶𝗻𝗮𝗹𝗹𝘆 𝘀𝗲𝗲𝗶𝗻𝗴 𝘁𝗵𝗮𝘁 𝗵𝗮𝗹𝗹𝘂𝗰𝗶𝗻𝗮𝘁𝗶𝗼𝗻 𝗺𝗶𝗴𝗵𝘁 𝗯𝗲 𝗮 𝗰𝗶𝗿𝗰𝘂𝗶𝘁-𝗹𝗲𝘃𝗲𝗹 𝗽𝗵𝗲𝗻𝗼𝗺𝗲𝗻𝗼𝗻 - 𝗻𝗼𝘁 𝗺𝗮𝗴𝗶𝗰, 𝗻𝗼𝘁 𝗿𝗮𝗻𝗱𝗼𝗺𝗻𝗲𝘀𝘀. This reframes hallucinations from a “black box mystery” to a reward design consequence. The idea is that these neuron-level insights will be useful in the next generation of mitigation strategies and tools, to detect hallucinations on LLM level, and those strategies still need to be developed and tested. Another step forward trustworthy and reliable industrial LLMs 👏 Boris Scharinger Ramon Wartala Debmalya Biswas Bora Ger Maria Sukhareva

45

12 Comments

Smart people update on AI and work: Stanford and Anthropic Economic Index highlights: 1️⃣ 'Codified' vs 'Tacit' skills diverge - basically - knowledge that is easy to consume by LLMS is seeing slow hiring. Knowledge that is experience backed is not seeing this. 2️⃣ 'Augmentation' vs 'Task Automation' Careers that are being augmented vs automated - guess which one is seeing slowest growth? 3️⃣ Early Career Stagnation - GenX was told this too when getting in the workforce (and look at us now). This time, the disruption in the labor market is purely early career knowledge work. https://lnkd.in/evv7CBFH

6

𝐖𝐞𝐞𝐤𝐞𝐧𝐝 𝐥𝐞𝐚𝐫𝐧𝐢𝐧𝐠 𝐭𝐡𝐚𝐭 𝐠𝐞𝐧𝐮𝐢𝐧𝐞𝐥𝐲 𝐬𝐡𝐢𝐟𝐭𝐞𝐝 𝐦𝐲 𝐭𝐡𝐢𝐧𝐤𝐢𝐧𝐠. I spent time diving into Meta’s new 𝐕𝐋‑𝐉𝐄𝐏𝐀 research paper — and it left me with a very different perspective on where AI might be heading beyond LLMs. We’ve all seen how far LLMs have taken us. They generate beautifully, they converse fluently, and they’ve become the default interface for AI. But reading this paper made me pause and ask:   𝐖𝐡𝐚𝐭 𝐢𝐟 𝐭𝐡𝐞 𝐟𝐮𝐭𝐮𝐫𝐞 𝐨𝐟 𝐢𝐧𝐭𝐞𝐥𝐥𝐢𝐠𝐞𝐧𝐜𝐞 𝐢𝐬𝐧’𝐭 𝐚𝐛𝐨𝐮𝐭 𝐠𝐞𝐧𝐞𝐫𝐚𝐭𝐢𝐧𝐠 𝐭𝐡𝐞 𝐧𝐞𝐱𝐭 𝐭𝐨𝐤𝐞𝐧… 𝐛𝐮𝐭 𝐩𝐫𝐞𝐝𝐢𝐜𝐭𝐢𝐧𝐠 𝐭𝐡𝐞 𝐧𝐞𝐱𝐭 𝐬𝐭𝐚𝐭𝐞 𝐨𝐟 𝐭𝐡𝐞 𝐰𝐨𝐫𝐥𝐝? That’s the core idea behind VL‑JEPA — a model that doesn’t try to produce text, but instead tries to understand meaning directly in embedding space. It’s a quiet but profound shift: from generation → understanding, from tokens → representations, from language → world models. And honestly, it made me rethink what “intelligence” should look like in AI.   𝐋𝐋𝐌 𝐯𝐬 𝐕𝐋‑𝐉𝐄𝐏𝐀 — 𝐌𝐲 𝐭𝐚𝐤𝐞𝐚𝐰𝐚𝐲 𝐋𝐋𝐌𝐬 • Core behavior: Predict the next token • Strength: Fluent language generation • Data needs: Heavy supervision and curated datasets • Architecture: Autoregressive models • Primary goal: Produce text 𝐕𝐋‑𝐉𝐄𝐏𝐀 • Core behavior: Predict the next semantic state • Strength: Deep contextual understanding • Data needs: Mostly self‑supervised learning • Architecture: Joint embedding predictive model • Primary goal: Build a world model LLMs are incredible at expression. VL‑JEPA feels like a step toward perception + reasoning. Both matter — but they’re solving different problems.   𝐖𝐡𝐲 𝐭𝐡𝐢𝐬 𝐦𝐚𝐭𝐭𝐞𝐫𝐞𝐝 𝐭𝐨 𝐦𝐞 This paper reminded me that AI is still in its early chapters. And sometimes, the biggest breakthroughs come not from scaling what we already have… but from 𝐫𝐞𝐭𝐡𝐢𝐧𝐤𝐢𝐧𝐠 𝐭𝐡𝐞 𝐟𝐮𝐧𝐝𝐚𝐦𝐞𝐧𝐭𝐚𝐥𝐬. It’s exciting to see architectures that move closer to how humans infer, predict, and understand the world — not just how we speak about it. 𝐈𝐟 𝐲𝐨𝐮’𝐫𝐞 𝐞𝐱𝐩𝐥𝐨𝐫𝐢𝐧𝐠 𝐭𝐡𝐞 𝐟𝐮𝐭𝐮𝐫𝐞 𝐨𝐟 𝐀𝐈, 𝐭𝐡𝐢𝐬 𝐢𝐬 𝐚 𝐩𝐚𝐩𝐞𝐫 𝐰𝐨𝐫𝐭𝐡 𝐫𝐞𝐚𝐝𝐢𝐧𝐠. 𝐈𝐭 𝐦𝐢𝐠𝐡𝐭 𝐣𝐮𝐬𝐭 𝐬𝐡𝐢𝐟𝐭 𝐲𝐨𝐮𝐫 𝐦𝐞𝐧𝐭𝐚𝐥 𝐦𝐨𝐝𝐞𝐥 𝐭𝐡𝐞 𝐰𝐚𝐲 𝐢𝐭 𝐬𝐡𝐢𝐟𝐭𝐞𝐝 𝐦𝐢𝐧𝐞. What do you think — will the next leap in AI come from models that understand rather than generate? For those interested, here’s the paper I read this weekend: https://lnkd.in/dEe3fsxe #WeekendLearning #AIResearch #VLJEPA #LLM #RepresentationLearning #WorldModels #AIThinking

6

The future of AI scaling will prioritize efficient, low-cost inference. As of late 2025, the industry has shifted decisively into an inference dominant era, where inference revenue has surpassed training revenue for the first time, and ongoing operational costs now dwarf one-time training expenses. This transition is driven by exploding real world usage, billions of daily queries, real time agents, and edge deployments making inference the primary bottleneck for scalability, energy efficiency, and profitability. Training remains essential for frontier models but is episodic and increasingly commoditized, with costs dropping dramatically (e.g., GPT-4-level models trained for under $6 million). Specialized inference hardware is emerging as the key enabler. Google’s TPUs (now in their sixth-generation Trillium, with strong efficiency gains for large-scale cloud inference) continue to power hyperscalers’ internal workloads. Meanwhile, Groq’s purpose built LPUs excelled in low latency, high throughput performance at lower cost until Nvidia’s $20 billion licensing deal and talent acquisition in December 2025 effectively integrated Groq’s technology into Nvidia’s ecosystem, consolidating leadership in hybrid GPU/LPU architectures. Ultimately, winners in the next phase will optimize for tokens-per-dollar and milliseconds of latency, enabling widespread adoption of advanced AI agents and real-time applications. Nvidia now positioned to dominate both training and inference, while custom ASICs like TPUs remain vital for hyperscale efficiency. https://lnkd.in/gV6KrTRy

30

1 Comment

High-performance AI is not built in a vacuum but on strong foundational engineering. I spent the holiday break completing my 7th Databricks certification: Machine Learning Associate. While it is tempting to stay at a strategic 30k foot view, I believe an AI executive must stay grounded in the technical reality their teams face every day. In large-scale corporate environments, architecture is where the time dimension becomes critical. A shortcut taken today might deliver a quick win, but the right architectural choices ensure that you are building a scalable asset for the next five years rather than technical debt that paralyzes the organization in eighteen months. Why I believe technical rigor is a business imperative: 🔹 Precision in Execution: I was reminded this week how a simple shift from classic Python UDFs to Pandas UDFs on Spark can transform performance. It sounds like a detail, but it can be the difference between a sluggish pipeline and a cost-optimized engine that potentially saves 100k Euro in annual cloud spend. 🔹 Trust and Reliability: In a recent production use case, we achieved >92% prediction accuracy and cut processing time from hours to seconds. You do not reach that scale by accident. You reach it by validating production workflows against the latest platform capabilities. 🔹Empowering the Team: As a leader, staying current is not about doing the work for your team. It is about setting the standard. Practical experience with frameworks like ISO 27001 and 42001 provides the safety rails for high-speed delivery instead of bureaucratic roadblocks. I have found that the most effective technical leadership ensures engineering precision translates eventually into documented business impact. The objective is rarely just to build a model but to architect the industrial-grade platforms that makes those solutions sustainable and reliable for the long term. I am curious: What foundations are you laying today to ensure your Data & AI strategy scales effectively over the next years?

78

16 Comments

I'm excited to be joining as a panelist at Ground Truth, hosted by Sphere x Andreessen Horowitz. A half-day conference for senior engineers working at the intersection of AI x regulation. Two technical panels: reliability guarantees in production AI and human-in-the-loop design, with compliance as a moat. I'll be sharing specific insights on where the deterministic line sits when the model proposes and infrastructure enforces, whether you can trust your evals, what agent reliability actually takes in production, and where human-in-the-loop is a moat. SF, 17 Sept, 1-9pm Details and application via the link: https://lnkd.in/gbVSAtRG

99

4 Comments