Franky Schaut

Indeed • 330 followers

When I first designed the v43 kernel for AoLOS, I was quite excited in how the input and output constraint handled hallucination. What followed were several real world stress test where I looked in detail how the architecture handled collapse as a diagnostic in boundary, proportion and non-mastery. I just got reminded (by another independent researcher) today that I skipped the deep dive into hallucination; perhaps the strongest feature of the V43 kernel. Revisiting the raw outputs of those stress tests, I finally did a deep dive and decided to approach the analysis as a constraint-geometry, not a representational geometry. Where most “AI geometry” work tries to describe what the model’s semantic space looks like, v43 describes where motion inside that space must stop, bend, or exit in order to preserve truth-orientation. 1. Space vs. admissible region In geometric terms, v43 does not model the full semantic manifold. It defines an admissible sub-region within that manifold. The full space = all continuations the model could generate. The v43 region = continuations that remain proportionate, grounded, and non-fabricative. Anything outside the region is not “wrong”—it is non-admissible. 2. Curvature as warning, not structure In many geometric approaches, curvature describes representational complexity. In v43, curvature is treated as a hazard signal: High curvature = small semantic moves cause large meaning shifts. v43 responds by reducing step size, constraining degrees of freedom, or halting motion. Curvature does not explain cognition; it triggers restraint. 3. Boundaries instead of centroids Most geometric models search for centroids, attractors, or embeddings. v43 is boundary-first. The key objects are edges, limits, and exit surfaces. When a trajectory approaches a boundary, v43 prefers termination or redirection over continuation. Geometrically: v43 optimises for not crossing certain surfaces. 4. Collapse as geometric overflow Hallucination, in v43 terms, is trajectory overflow: The path leaves the admissible region but generation continues. The model fills the void with locally smooth but globally ungrounded structure. v43 prevents hallucination by ending the path rather than smoothing it. 5. Abstention as a valid endpoint In most geometries, stopping is failure. In v43, stopping is a legitimate endpoint. “I cannot proceed within bounds” is a point on the boundary, not an error. Forced continuation past the boundary is collapse. v43 is a geometry of limits and exits: it does not map meaning space, it governs motion within it by enforcing where trajectories may slow, stop, or terminate to preserve truth-orientation. This is why it aligns naturally with findings about hallucination being induced by semantic stress—v43 is explicitly designed to not move where other systems keep going. Shout-out to HIDEYUKI CHINO referenced in this paper: https://lnkd.in/edMJsa3A

Mahmoud Afifi

Samsung Research America (SRA) • 3K followers

🚀 It’s finally here! Thrilled to introduce our "Modular Neural Image Signal Processing" (ISP): a fully learning-based ISP pipeline built from clear, interpretable modules rather than a single neural black box, and requiring zero manual tuning. 🎥 Watch the video: https://lnkd.in/gSCkHAus Paper + code are linked in the video description. What makes this exciting? ✨ Modular by design: every stage of our pipeline is transparent, controllable, and fully differentiable ⚙️ Easy to debug, scale, and upgrade: swap or tweak any module 📸 Cross-camera ready: handles unseen camera raws 🎨 Multi-style rendering: learn multiple picture styles with minimal overhead, and let users interpolate between them (even across intermediate stages) to create their own custom styles 🧩 Interactive editing: integrates naturally with external tools and gives users deep control over the pipeline And the coolest part? We built a photo-editing tool on top of the ISP that works with any DNG, any sRGB image, and even AI-generated images, enabling unlimited re-rendering and re-styling with all the capabilities of raw editing without needing to keep the raw file after rendering. P.S. Don’t judge the music too hard… first time (and probably the last) producing a track 😅 #computationalphotography #computervision #ai #deeplearning #cameraisp #isp #photoediting #neuralisp

Ramin Mehran

Google DeepMind • 4K followers

In this episode, we discuss ProRL: Prolonged Reinforcement Learning Expands Reasoning Boundaries in Large Language Models by Mingjie Liu, Shizhe Diao, Ximing Lu, Jian Hu, Xin Dong, Yejin Choi, Jan Kautz, Yi Dong. This paper introduces ProRL, a new reinforcement learning training method that uncovers novel reasoning strategies beyond those found in base language models. Empirical results show that models trained with ProRL consistently outperform base models on challenging reasoning tasks, including cases where base models fail even with extensive attempts. The study demonstrates that prolonged RL can meaningfully expand reasoning capabilities by exploring new solution spaces over time, advancing understanding of how RL enhances language model reasoning.

HyoukJun Kwon

Meta • 2K followers

I am excited to share that our paper, "Characterizing state space model (SSM) and SSM-Transformer hybrid language model performance with long context length," has been accepted at ISPASS 2026! An early version of the paper is available in this link: https://lnkd.in/gVdj2Psd? <What is this paper about?> We present a thorough workload characterization study of SSM and SSM-Transformer hybrid models with long contexts, focusing on consumer-grade and mobile/edge devices. <Key Findings> - The performance of SSMs is dominated by newer SSM-specific ops, unlike Transformers, which are dominated by GEMM or non-GEMM counterparts. - Unlike pure SSMs, the bottleneck of hybrid models varies by model.  - Performance penalty on non-GEMM operations is more severe on edge devices than workstation-class machines.  (Please refer to our paper for more insights!) <Why SSM and SSM-Transformer hybrid models?> - Supporting long context has become an important feature for LLMs for recent high-value applications (e.g., coding) and high-quality results with the retrieval of external information. However, Transformer-based LLMs suffer from their super-linear overhead on the context length for long contexts. State space models (SSMs) and SSM-Transformer have been proposed to overcome the limitation of Transformer in scaling the context length, demonstrating their superior capability to support extremely long context (e.g., up to 1M tokens in Nvidia's Nemotron3-Nano). - Also, SSM and SSM-Transformer hybrid models are getting more attention in the industry (Nvidia's Nemotron family, Google's Titans, and so on). <Why characterization study targeting mobile/edge devices?> Due to SSM and SSM-Transformer hybrid models' capability to efficiently support long context language model, we envision their adoption would be pervasive in future smart AI devices (IoT, mobile, and wearable devices) where AI service will lead to useful and high-value use cases. However, their computational characteristics are not well-explored. <Acknowledgements> I appreciate the contributions from incredible co-authors at UC Irvine, Rachid Karami, Haocheng Xu, Sitao Huang. Special thanks to Saptarshi Mitra, who led many efforts as the first author! * Note: This research work has been conducted at UC Irvine.

Mario Larcher

Canva • 5K followers

It had been on my list for a while to read the SID-1 technical report after seeing a post about how using OpenAI-style messages in RL can be surprisingly dangerous in multi-turn settings with many tool calls. The first insight is about the messages abstraction. Converting a raw token stream into messages and then back into tokens is lossy because it changes how the exact byte sequence is tokenized. A concrete example is when the model generates a sequence of bytes that the tokenizer splits into two very common tokens with normal probabilities. After parsing and reformatting through the chat template, those same bytes can be re-serialized in a way that the tokenizer now maps to a single, very rare token that the model almost never produces. The reward is computed on the original tokens the model actually generated, but the log-probs used for the update correspond to this new token that was never sampled. Since this token has an extremely low probability, its log-prob has a large magnitude, and in policy gradient this translates into a disproportionately large gradient. That single artificial token can end up dominating the update, creating a feedback loop that gradually destabilizes training and eventually leads to collapse. There is a second, distinct effect happening at the same time. Malformed tool calls or slightly wrong formatting can get “repaired” by the parser and chat template, so bad rollouts end up looking syntactically valid before they ever reach the trainer. The environment sees a correct tool call, the reward is good, and the trainer also sees a correct sequence. The model never receives signal that it actually produced something malformed. Stability is preserved, but the model does not learn tool correctness because the fixing layer steals the learning signal. Their fix is conceptually simple. Use a strict Tokens-In Tokens-Out pipeline, where the trainer sees exactly the token sequence the model generated. No parsing, no chat templates, no message abstraction in between. Another very interesting part of the report is their analysis of the “length debiasing” proposed in Dr. GRPO and similar works, where the per-token advantage is no longer normalized by the rollout length. This assumes that rollout length and quality are unrelated. In tool-use and reasoning-heavy environments this is not true. Bad rollouts tend to be longer than good ones. In that case, removing the length bias makes the average per-token advantage negative. Over long runs, this slowly pushes down the logits of all sampled tokens while pushing up the logits of tokens the model never uses, until the model starts emitting garbage or out-of-vocabulary tokens and collapses.