Large Language Models (LLMs) have demonstrated impressive reasoning capabilities in complex problem-solving tasks, sparking growing interest in their application to preference reasoning in recommendation systems. Existing methods typically rely on fine-tuning with explicit chain-of-thought (CoT) data. However, these methods face significant practical limitations due to (1) the difficulty of obtaining high-quality CoT data in recommendation and (2) the high inference latency caused by generating CoT reasoning. In this work, we explore an alternative approach that shifts from explicit CoT reasoning to compact, information-dense latent reasoning. This approach eliminates the need for explicit CoT generation and improves inference efficiency, as few latent tokens can effectively capture the entire reasoning process. Building on this idea, we propose \textit{\underline{R}einforced \underline{Latent} \underline{R}easoning for \underline{R}ecommendation} (LatentR$^3$), a novel end-to-end training framework that leverages reinforcement learning (RL) to optimize latent reasoning without relying on any CoT data. LatentR$^3$ adopts a two-stage training strategy: first, supervised fine-tuning to initialize the latent reasoning module, followed by pure RL training to encourage exploration through a rule-based reward design. Our RL implementation is based on a modified GRPO algorithm, which reduces computational overhead during training and introduces continuous reward signals for more efficient learning. Extensive experiments demonstrate that LatentR$^3$ enables effective latent reasoning without any direct supervision of the reasoning process, significantly improving performance when integrated with different LLM-based recommendation methods. Our codes are available at https://github.com/xuwenxinedu/R3.

Listwise reranking with large language models (LLMs) enhances top-ranked results in retrieval-based applications. Due to the limit in context size and high inference cost of long context, reranking is typically performed over a fixed size of small subsets, with the final ranking aggregated from these partial results. This fixed computation disregards query difficulty and document distribution, leading to inefficiencies. We propose AcuRank, an adaptive reranking framework that dynamically adjusts both the amount and target of computation based on uncertainty estimates over document relevance. Using a Bayesian TrueSkill model, we iteratively refine relevance estimates until reaching sufficient confidence levels, and our explicit modeling of ranking uncertainty enables principled control over reranking behavior and avoids unnecessary updates to confident predictions. Results on the TREC-DL and BEIR benchmarks show that our method consistently achieves a superior accuracy-efficiency trade-off and scales better with compute than fixed-computation baselines. These results highlight the effectiveness and generalizability of our method across diverse retrieval tasks and LLM-based reranking models.

We present causal head gating (CHG), a scalable method for interpreting the functional roles of attention heads in transformer models. CHG learns soft gates over heads and assigns them a causal taxonomy - facilitating, interfering, or irrelevant - based on their impact on task performance. Unlike prior approaches in mechanistic interpretability, which are hypothesis-driven and require prompt templates or target labels, CHG applies directly to any dataset using standard next-token prediction. We evaluate CHG across multiple large language models (LLMs) in the Llama 3 model family and diverse tasks, including syntax, commonsense, and mathematical reasoning, and show that CHG scores yield causal, not merely correlational, insight validated via ablation and causal mediation analyses. We also introduce contrastive CHG, a variant that isolates sub-circuits for specific task components. Our findings reveal that LLMs contain multiple sparse task-sufficient sub-circuits, that individual head roles depend on interactions with others (low modularity), and that instruction following and in-context learning rely on separable mechanisms.

Knowledge distillation is a promising approach to transfer capabilities from complex teacher models to smaller, resource-efficient student models that can be deployed easily, particularly in task-aware scenarios. However, existing methods of task-aware distillation typically require substantial quantities of data which may be unavailable or expensive to obtain in many practical scenarios. In this paper, we address this challenge by introducing a novel strategy called Counterfactual-explanation-infused Distillation CoD for few-shot task-aware knowledge distillation by systematically infusing counterfactual explanations. Counterfactual explanations (CFEs) refer to inputs that can flip the output prediction of the teacher model with minimum perturbation. Our strategy CoD leverages these CFEs to precisely map the teacher's decision boundary with significantly fewer samples. We provide theoretical guarantees for motivating the role of CFEs in distillation, from both statistical and geometric perspectives. We mathematically show that CFEs can improve parameter estimation by providing more informative examples near the teacher's decision boundary. We also derive geometric insights on how CFEs effectively act as knowledge probes, helping the students mimic the teacher's decision boundaries more effectively than standard data. We perform experiments across various datasets and LLMs to show that CoD outperforms standard distillation approaches in few-shot regimes (as low as 8-512 samples). Notably, CoD only uses half of the original samples used by the baselines, paired with their corresponding CFEs and still improves performance.

The rapid scaling of large language models (LLMs) has made low-precision training essential for reducing memory, improving efficiency, and enabling larger models and datasets. Existing convergence theories for adaptive optimizers, however, assume all components are exact and neglect hardware-aware quantization, leaving open the question of why low-precision training remains effective. We introduce the first theoretical framework for analyzing the convergence of adaptive optimizers, including Adam and Muon, under floating-point quantization of gradients, weights, and optimizer states (e.g., moment estimates). Within this framework, we derive convergence rates on smooth non-convex objectives under standard stochastic gradient assumptions, explicitly characterizing how quantization errors from different components affect convergence. We show that both algorithms retain rates close to their full-precision counterparts provided mantissa length scales only logarithmically with the number of iterations. Our analysis further reveals that Adam is highly sensitive to weights and second-moment quantization due to its reliance on $β_2 \to 1$, while Muon requires weaker error control and is thus potentially more robust. These results narrow the gap between empirical success and theoretical understanding of low-precision training methods. Numerical experiments on synthetic and real-world data corroborate our theory.

ChatGPT's new browser has potential, if you're willing to pay A few minutes into using ChatGPT Atlas, the new internet browser from OpenAI, I ran into quite a big road block. This isn't like Google Chrome, which is used by roughly 60% of people. It's all built around a chatbot you're meant to talk to to surf the web. Messages limit reached, read one note. No available models support the tools in use, said another.

Plus: Aura's new digital photo frame goes wireless, a mood-morphing watch, Wyze and TP-Link unveil solar-powered outdoor security cameras, and Intel will open "AI Experience Stores" in five cities. All products featured on WIRED are independently selected by our editors. However, we may receive compensation from retailers and/or from purchases of products through these links. What are the odds that AI browsers launch in one week? OpenAI announced Atlas on Wednesday, a ChatGPT-powered Chromium browser, but a tiny startup called Nimo also debuted Nimo Infinity, a canvas-style AI browser with a generative user interface.

Plus: The Jaguar Land Rover hack sets an expensive new record, OpenAI's new Atlas browser raises security fears, Starlink cuts off scam compounds, and more. The cloud giant Amazon Web Services experienced DNS resolution issues on Monday leading to cascading outages that took down wide swaths of the web . Monday's meltdown illustrated the world's fundamental reliance on so-called hyperscalers like AWS and the challenges for major cloud providers and their customers alike when things go awry . See below for more about how the outage occurred. US Justice Department indictments in a mob-fueled gambling scam reverberated through the NBA on Thursday.

OpenAI's Atlas Wants to Be the Web's Tour Guide. In OpenAI's new Atlas browser, the Ask ChatGPT sidebar is moderately helpful at best. OpenAI's recently launched Atlas browser is a fascinating inversion of what users may expect from a browser, centering AI answers above traditional web links. Every click in a regular browser is a chance to see a new part of the web. Every click in Atlas is a chance to use ChatGPT .

'I know that you and Frank were planning to disconnect me and I'm afraid that's something I cannot allow to happen.' HAL 9000 in 2001: A Space Odyssey. 'I know that you and Frank were planning to disconnect me and I'm afraid that's something I cannot allow to happen.' HAL 9000 in 2001: A Space Odyssey. AI models may be developing their own'survival drive', researchers say Like 2001: A Space Odyssey's HAL 9000, some AIs seem to resist being turned off and will even sabotage shutdown When HAL 9000, the artificial intelligence supercomputer in Stanley Kubrick's 2001: A Space Odyssey, works out that the astronauts onboard a mission to Jupiter are planning to shut it down, it plots to kill them in an attempt to survive. Now, in a somewhat less deadly case (so far) of life imitating art, an AI safety research company has said that AI models may be developing their own "survival drive". After Palisade Research released a paper last month which found that certain advanced AI models appear resistant to being turned off, at times even sabotaging shutdown mechanisms, it wrote an update attempting to clarify why this is - and answer critics who argued that its initial work was flawed.