Yacine Mahdid

sometime there is an idea that just makes other weird research results make so much more sense and sort of cascade to explain a whole bunch of phenomenon. neural thicket from the great yulu gan is one such paper that I absolutely loved from start to finish. the old mental model I had for pre-trained neural networked was that the pre-training regime made them land on a single good set of weight that was optimal for many tasks. but after going through the paper "Neural Thickets: Diverse Task Experts Are Dense Around Pretrained Weights" my view has been changed. the better idea seems to be that the model after pre-training is in a basin where it is surrounded by more specialized expert at varying distance from that set of weights which the authors calls thicket. a bit like the valley surrounded by mountains. now depending on whether the model has big or small capacity this density of thicket will be varying and these expert weights will also have varying level of difficulty to be reached via post-training. this finding elegantly presented by Yulu & co. has far reaching consequence on how we should think about pre-training and it's connection with post-training models (and even distillation / continual learning) very densely packed interview in there and there is a few open threads that are ripe to be picked up if you are interested in this research direction! ps: my favorite bit is the connection with the baldwin effect in biology which is an evolutionary concept proposing that learned behaviors can shape the course of evolution, very fitting with post-training dynamics! Table of Content: 00:00:00 - pre-trained LLM are surrounded by tasks experts 00:05:35 - introduction to the great yulu gan and his beautiful paper 00:08:21 - what are neural thickets? 00:13:25 - relationship between scale and neural thickets? 00:14:56 - why it’s happening for the pretrained base model? 00:18:00 - solution density for small and large model 00:19:30 - do all perturbation help in the same way? and why? 00:25:00 - why does mixed data pretraining create neural thickets? 00:30:20 - would a distill model exhibit a different kind of distribution landscape? 00:32:50 - RandOpt algorithm 00:36:50 - where are the gains coming from with adding random noise? 00:45:45 - what happens to the distribution of the weight when doing the distillation? 00:48:55 - does the useful perturbation of the layers evenly distributed? 00:51:40 - how should we releasing model with a weights distribution perspective? 00:53:56 - does routing work better than distilling you think? 00:56:10 - what do you think will happens at 400B? 00:58:10 - how does this connect to continual learning? 01:02:20 - connection with the baldwin effect from biology? 01:06:30 - discover novelty by random exploration of the weights? 01:09:00 - what’s next for the great yulu gan enjoy folks hope this stuff is useful🌹 --- ##### Comments > **M4rc0z @dreamworks2050** · [2026-08-12](https://x.com/dreamworks2050/status/2087589665041994211) > > I like my neural networks THICC

@yacinelearning (Yacine Mahdid) — 6h one of the most important element of a good algorithmic research paper is a very colourful and simple to understand diagram of said algorithm [Image: photo of a printed paper page showing "Figure 2: Outline of Vector Policy Optimization" — a diagram with boxes for Sample rollout (x → π_θ → y), Shared Reasoning Prefix (three colored answer sequences y1/y2/y3), Evaluation (r(x,y1), r(x,y2), r(x,y3)), Sample Reward Weightings w_i ~ Dir(1) shown as small bar charts, and a grid of green squares under "max in set", flowing to RL update and Â(y|x) ← Estimate advantage ← R(y|x) ← mean across weightings. Below the figure, printed body text reads: "Figure 2: Outline of Vector Policy Optimization. Given a prompt x, a model π_θ outputs m answers in a single autoregressive chain. Each answer y_i is evaluated on multiple objectives and receives a score vector [r1, r2, ..., rn]. We repeatedly sample weighting vectors w_i ~ Dir(1) in the set of answers. This set level reward is treated as the mean best-of-m under each weighting w in the set of answers, and evaluate the reward for the entire completion π_θ(x). proposed algorithm, Vector Policy Optimization (VPO), which trains a policy to produce such sets. VPO has two key components. First, we train a model to generate multiple candidate completions per prompt within a single autoregressive rollout. Second, we replace a fixed reward weighting with a distribution over weights, so the model is incentivized to span its candidates across different trade-offs. Together, these components define a set-level objective that rewards the model for producing diverse, high-quality solutions. We describe each component in detail below. 3.1 Multi-Answer Chains as In-Context Exploration Following Puri et al. [2026], we train a language model to produce a set of m candidate completions S = {y1, ..., ym} within a single rollout. The completions are emitted sequentially, separated by a delimiter token, so when generating y_i, the prefix already contains y1, ..., y_{i-1}. This fundamentally changes the nature of exploration. Under standard independent sampling, diversity arises only from stochastic decoding applied to a fixed conditional distribution, producing small variations around whichever mode the policy has concentrated on. In multi-answer rollouts, each new candidate can attend to the ones already emitted, giving the model the capacity to recognize which regions of the solution space are covered and steer subsequent candidates towards different ones. Diversity becomes an explicit, in-context mechanism rather than a byproduct of sampling noise. Importantly, this mechanism provides the capacity for diversity, but not a strong incentive¹. Without an appropriate training signal, the model will still collapse to producing near-identical answers. We confirm this empirically in Section 5: Multi-RLVR, which combines multi-answer rollouts with a fixed scalar reward, produced sets whose reward-diversity collapses early in training. The objective defined in the next section supplies this missing incentive by rewarding sets whose elements specialize to different reward trade-offs. 3.2 Set-Level Optimization via Stochastic Scalarization To train a policy to output diverse sets, we replace the fixed scalarization prevalent in RL post-training..." [text cut off at bottom edge of photo]] @yacinelearning (Yacine Mahdid) — Jun 13 on this beautiful saturday afternoon we're going to read about how vector policy optimization inject solutions diversity in these models [quoted image: thumbnail of a hand holding the same printed paper] Engagement: 8 replies, 11 reposts, 266 likes, 11K views @GlennMatlin (Glenn Matlin) — 5h Fig1 is critical for modern publishing [reply continues below, cut off — a partially visible meme image with text "THE BAD METAGAME" is cropped at the bottom of the screenshot]

Yacine Mahdid @yacinelearning · 6h if you have any goblins X codex related questions do let me know I'm preparing an interview on this very important topic > QUOTED THREAD: > roon @tszzl · 3h > I think it becomes annoying when it mentions goblins ever single chat and it's fair shakes to try and reduce that > 💬 53 🔁 11 ❤️ 382 👎 > > Yacine Mahdid @yacinelearning · 2h > hey roon would you be open to hop into an interview to discuss the goblins situation > 💬 1 🔁 ❤️ 10 📊 301 > > roon @tszzl · 1m > Ok > 💬 1 🔁 ❤️ 2 👎