Post-training updates change what large language models (LLMs) can answer, but may leave their answerability judgments misaligned. We study this knowledge-judgment desynchronization through functional feeling of knowing (FoK), which supports selective generation, retrieval triggering, and model routing without implying subjective experience. We propose two trainable channels: a knowledge channel that changes answerability and a FoK channel that judges the current knowledge state. Three stages train them: on-policy supervised fine-tuning (OP-SFT) updates knowledge from corrected on-policy answers; knowledge-delta FoK training (K$Δ$-FT) adapts judgments across intermediate knowledge states; and FoK-guided policy optimization (FGPO) trains the subsequent answer or explanation. Factual-QA experiments, cross-dataset evaluation, and three sequential updates show that explicit FoK alignment improves judgment quality on updated knowledge. MMLU and ARC-Easy separately assess capability preservation.
At a fixed privacy budget, curvature is exactly the object the budget makes unavailable. A recent answer estimates it from synthetic probes, at neither privacy cost nor distribution shift, and validates a probe by the eigenvalue decay it reproduces. That test reads the spectrum, not the basis, and what a probe reconstructs is unknown. This paper gives a random-matrix account of data-free preconditioning, organized by one dichotomy. Of the two Kronecker factors, the error-signal factor $ G $ and the architecture-determined bulk of the activation factor $ A $ are reconstructible without data up to an explicit error-energy bound; only the data-induced spikes of $ A $ are not. We verify that bound with zero violations along a full training trajectory, though it stays one to nearly three orders of magnitude above the gap it bounds. We locate the failure boundary at a Baik-Ben Arous-Péché threshold: a bulk-only probe suffers an exact damage $s^2/(1-s^2)$ from a spike of strength $s^2$, invisible once its amplitude exceeds the scaled-bulk threshold $( \sqrt {c}/(1+ \sqrt {c}))^{1/2}$, $c = d/n$. The test in use is basis-blind: rotating only the probe eigenbasis at a near-exact matched spectrum inflates the residual by $2.3 \times $ to $89.8 \times $, and a closed-form volume law fixes the within-spectrum ordering of the critical volume while a two-oracle construction proves that no functional of the probe spectrum fixes its magnitude. Over forty seeds on three modalities, per-layer tuning does not transfer to accuracy, no data-free arm beats the flat probe, public-data curvature does on both modalities tested, and the ordering inverts on PathMNIST under real DP noise. The account yields a probe-design protocol and falsifiable predictions for the regimes where data-free curvature can be trusted.
Auto-bidding plays an essential role in online advertising, automatically adjusting bids for advertisers to optimize their commercial goals. The emerging AI-Generated Bidding (AIGB) paradigm widely adopts generative modeling to optimize bidding strategies, yet suffers from the limited mode coverage of offline datasets and inadequate task-state understanding, hindering effective exploration of optimal strategies. Large Language Models (LLMs), with prior world knowledge and reasoning capabilities, offer a promising approach to overcome these limitations. However, directly applying LLMs to auto-bidding tasks faces inherent challenges in limited numerical precision, hallucinations, and inference latency. To address these limitations, we propose AIGB-R1, a hierarchical self-evolving auto-bidding framework aiming to enhance AI-Generated Bidding via LLMs' Reasoning capabilities, comprising a high-level Planner module for macro-level strategy planning and a low-level Executor module for fine-grained decision-making. Building upon this, we design an experience-driven self-evolving loop, enabling autonomous strategy exploration and optimization from accumulated experience. We adopt a two-stage pipeline of offline pre-training and post-training alignment, and build an interactive bidding simulation environment for strategy rollout. Furthermore, we propose Decoupled Group Relative Policy Optimization (D-GRPO) to achieve end-to-end optimization via advantage decoupling. Experimental results on a large-scale public dataset demonstrate the effectiveness of AIGB-R1.
Zejian Wang, Yifan Zhao, Kaihui Ren, Xiaoling Zhou, Yubin Wang, Chenhao Guan, Lin Zhang, Yufeng Li 等 14 位作者
Fully Sharded Data Parallelism (FSDP2) maintains parameters and gradients as per-parameter tensors, whereas its collective path operates on flat, rank-major buffers. Converting between these layouts copies all-gather outputs into parameter storage and packs gradients into full-group fp32 buffers. Layout conversion alone accounts for 6.4% of step time in Llama-3.1 70B training at 8K tokens per GPU. Communication kernels impose a second cost: a default NCCL reduce-scatter can slow a concurrent GEMM that would otherwise complete in a single wave by 69%. We present TIDE, a set of zero-copy collectives built on public NCCL 2.31+ APIs. TIDE preserves FSDP2's per-parameter storage and compute kernels while folding layout placement into collective addressing. TIDE-AG delivers shards directly to parameter storage using NVSwitch multicast and paced recursive-doubling RDMA, with no SMs used for transport or placement. TIDE-RS consumes per-parameter bf16 gradients directly and reduces them into fp32 shards (a mixed-precision mode that NCCL does not provide) using copy engines and short-lived thread blocks (CTAs) for node-local accumulation and NVSwitch for cross-node reduction. A two-phase pipeline separates gradient reads from network progress; Slot-Reuse optionally stages gradients in vacated all-gather slots. On 64 GPUs, TIDE reduces step time by up to 27.1% relative to Stock FSDP2 and 6.0% relative to FSDP2+SymmMem across Llama-3.1 70B, Wan2.2, and Qwen3-30B-A3B. It reduces reduce-scatter SM time by factors of 12.5-14.5 and cuts interference with concurrent compute by up to an order of magnitude. Its reduce-scatter buffer remains nearly constant from 16 to 64 GPUs at a fixed per-GPU output size, and its peak memory is up to 4.7 GiB below FSDP2+SymmMem's and, with Slot-Reuse on Wan2.2, below Stock FSDP2's.
Sohini Roychowdhury, Long Sha, Sunand Agarwal, Payal Kamboj, Shuyang Chen, Junlin Zhang, Scott Chen, Hui Zhang
Engagement-optimized content recommendation systems exhibit a well-documented rich-get-richer dynamic that structurally disadvantages emerging creators on large-scale platforms. Evaluating debiasing interventions at scale demands faithful simulation of creator-viewer ecosystems, yet the fidelity, diversity, and validity of such simulators remain underexplored. We present a grounded user-simulation framework that models a heterogeneous population of ${\sim}42$K creators and their coupled viewer interactions over 12~weeks, and use it to evaluate equitable view-allocation policies formulated as a budgeted multi-armed bandit (MAB). The simulator captures three behavioral dimensions: (i)~\textbf{creator diversity}, modeled via a clustering approach grounded in real platform creator-level distributions; (ii)~\textbf{behavioral fidelity}, captured through coupled creator-content state machines that represent growth trajectories, engagement accumulation, and graduation dynamics; and (iii)~\textbf{population-level validity}, measured by calibration against observed platform metrics. We evaluate simulative-configurations that vary content-viewership target definitions, reward functions, and a small-creator reserve lever to maximize creator-content recommendation equity. A floor-protected per content-view-target with a view-parity-gap reward guarantees 90\% of view targets being met by creators every week, grows small-creator organic reach ${\sim}{+}33\%$ per week, and achieves high distribution stability (set of creators with funded-views Jaccard $= 0.819$). We analyze the Sim2Real gap and propose validation protocols, contributing evidence for when simulation-derived fairness conclusions can be trusted.
Zhuoming Liu, Jinhong Lin, Kwan Man Cheng, Lin Zhang, Shayok Bagchi, Yin Li
Many modern vision-language models (VLMs) build on autoregressive decoding of discrete tokens. While text-based output interfaces enable scalable pretraining and strong zero-shot generalization across diverse tasks, they are poorly suited for problems that require precise continuous outputs, such as localizing temporal boundaries of events or generating robotic control actions. To address this challenge, we propose DRIFT, a general framework for adapting pretrained VLMs to continuous decoding tasks. DRIFT combines a base predictor, which provides a coarse estimate of the target output, with a generative refinement module based on flow matching that iteratively improves the prediction. This residual formulation transforms the generative modeling problem from learning a global output distribution to modeling a localized residual distribution around a strong prior, substantially simplifying optimization. We evaluate DRIFT on both perception and planning tasks, including visual grounding and robotic control. Across multiple tasks and architectures spanning MLLMs, VLAs, and WAMs, DRIFT consistently outperforms a strong set of regression- and generative-based solutions.
Jialin Zhang, Lakshya A Agrawal, Shangyin Tan, Donghyun Lee, Lutfi Eren Erdogan, Dan Klein, Koushik Sen, Alex Dimakis 等 9 位作者
Reflective text optimization makes agentic systems better, but the optimization loop is itself a resource-hungry workload whose sequential structure leaves parallel evaluation capacity idle. We extend the open-source GEPA optimizer to propose and evaluate a batch of candidates on each optimization step instead of one candidate at a time. In our sweep on two tasks, most batched runs finished in half the wall-clock time or less, and the fastest in about a quarter to a third. Batched settings also achieved higher held-out test scores, from 68.9% to 72.1% on LiveBench-Math (with $2\times2$) and from 49.0% to 60.0% on HoVer (with $8\times1$). A simple runtime model explains the measured speedups under a finite worker pool, and batched settings also spend their metric-call and dollar budgets more efficiently. We frame the axes along which a reflective optimization step can scale in parallel, a first look at allocating a fixed optimization budget across them.
Existing low-rank KV-cache indexes preserve model weights or key variance, not the attention scores that determine retrieval. We derive the expected score distortion of key-side rank-r compression, a two-sided covariance-weighted low-rank problem that uses both query and key statistics, and solve it exactly: the optimum is an optimal rank-r linear score map, computed in closed form from two d x d covariances per head, and its predicted per-head score-MSE reduction matches measurement (Pearson r = 0.997). The resulting training-free index, SAKI, beats key-PCA at every rank on LLaMA-3.1-8B, Qwen2.5-7B, Mistral-7B and Llama-3.2-3B, at every context length from 4K to 64K, and in every calibration-domain pair (0.748 -> 0.799 recall@64 at r = 32 on LLaMA-3.1-8B). Because RoPE rotates keys by their distance from the query, we make the closed form RoPE-aware by averaging the query covariance over rotations; this SAKI-RoPE, our main method, raises recall to 0.861, beats PCA on every head of all four models, and needs only 28-39% of PCA's rank to match it at r = 32 across the four models. Random-matrix theory of the covariance estimation noise shows that 512 calibration tokens suffice and that clipping the Marchenko-Pastur noise bulk helps SAKI when calibration is short. We make no perplexity claim: with the attention sink pinned, perplexity differences between indexes are small and depend on model and text, and we show that any low-rank index calibrated on a long prefix drops the attention-sink token, which collapses 32K perplexity unless the sink is pinned.