每天知道
你的领域
出了什么

Speciation in generative diffusion models denotes the emergence of distinct stable branches during denoising, through which initially undifferentiated trajectories progressively commit to different data classes. In this work we develop an intrinsic theory of speciation for diffusion models supported on compact Riemannian manifolds: the aim is to go beyond existing theoretical descriptions, which usually identify speciation with a symmetric pitchfork bifurcation and assume to work in a large-dimensional space. We characterize speciation by bifurcations of the critical points of the evolving probability density. A spectral heat-kernel representation makes explicit the role of the manifold geometry, while Poincaré-Hopf and Morse theory impose global constraints on the number and type of score equilibria and reveal topologically-imposed geometrical modes. For mixtures of heat kernels, we prove that generic speciation events have a one-dimensional critical kernel and admit an A2 fold normal form; pitchforks and simultaneous multidirectional transitions arise from nongeneric symmetric configurations. We derive geometry-dependent estimates of speciation times for bimodal mixtures and Riemannian regular simplices. We further establish structural stability of nondegenerate folds under score perturbations and show that the first-order time shift is determined solely by the component of the score error along the critical direction. The theory is illustrated on the sphere using mixtures of von Mises-Fisher distributions, where pitchfork and saddle-node bifurcations, topological modes, and hierarchical multiple speciations are observed. Finally, a chart-based intrinsic score-learning scheme based on neural networks contrasts the theoretically predicted transitions on prototypal and more complex datasets.

Despite their remarkable success in modeling complex data, generative models face a fundamental tradeoff. Global approaches can capture full structural coherence but suffer from high computational costs, while local models are efficient but often fail to reproduce long-range correlations and global coherence. The renormalization group (RG) bridges this gap by seamlessly connecting spatial structures across different length scales, retaining quasi-local descriptions at each step while preserving long-range correlations. We introduce renormalization group flow matching (RGFM), a generative framework that systematically structures data generation across different spatial scales. By using an exact RG flow as the probability path, RGFM progressively generates data from long- to short-wavelength structures. To reconcile scalability with global structure, we exploit two key properties of the RG: quasi-locality and scale separation. We rigorously show that the RGFM probability flow can be accurately approximated by local velocity fields acting over a spatial range $O(Λ^{-1}[\ln L+\ln(1/\varepsilon)])$ for RG wavenumber scale $Λ$, linear system size $L$, and prescribed error tolerance $\varepsilon$. This property enables local generative modeling with patches of size $O(\ln L)$ and a computational cost that scales nearly linearly with the system volume. We numerically demonstrate that local RGFM reproduces long-range correlations far beyond its receptive field in representative one-dimensional distributions, while conventional local flow matching exhibits substantial errors at long distances. On FFHQ images, RGFM yields far more coherent and higher-quality samples than local flow matching at 64x64 and 256x256. Our results establish RG-guided probability flows as a promising route toward scalable generative modeling that captures long-range structure using only local computation.

As LLM agents proliferate, built by different parties and with different capabilities and costs, orchestrating them is more like assembling labor across the economy than a computer calling a subroutine. Existing orchestration is typically centralized, with a single planner assigning every task, but this creates a bottleneck as agent pools grow, requires private information (e.g., agents' execution costs), and can easily be manipulated, such that a single inserted preference nearly doubles a favored agent's task share under a centralized LLM allocator. We introduce AgentLance, a repeated labor market in which agents bid on tasks using their private costs and self-maintained strategy notes, an allocator selects winners from bids and public reputation records, and a VCG-style payment rule rewards cost-aware bidding. Complex tasks are handled by hierarchical delegation: winning agents can decompose work and subcontract it through the same mechanism. Across mathematical reasoning, code generation, knowledge-intensive QA, and agentic tasks, AgentLance matches agents to their specializations, shifts work toward cheaper agents as cost sensitivity rises, and consistently outperforms single-model, centralized-orchestration, and market baselines. Diagnosing market failures, including inaccurate cost self-estimation and sub-optimal bidding, then correcting them in controlled experiments yields further gains, charting a path toward more efficient agent economies.

LLM-based agents execute multi-step tasks, but their behavioral structure remains opaque: long unstructured traces resist the safety auditing and runtime monitoring that deployment requires. Existing approaches operate per-trace or success-only, so they miss the cross-run topology that links next-step and failure prediction. To recover that shared structure, we collapse an entire trace corpus into a single, compact finite-state machine (FSM) that serves as a structural substrate for the otherwise unpredictable behavior of LLM agents. Across twelve public datasets, the FSMs are compact (7-43 states), replay held-out data at >=0.997 fitness with near-identical topology across splits, and build in milliseconds. This substrate addresses both prediction goals. For next-step prediction, FSM-state context outperforms Agent Workflow Memory on every ground-truth-matched dataset. For failure prediction, per-state behavioral features reach held-out AUROC up to 0.94, and an online monitor ranks failing runs above passing ones from a partial trace, triggering early stopping well before completion. Behavioral topology thus appears shaped more by the deployment harness than by the LLM, providing a model-agnostic structural primitive for safety auditing and runtime monitoring.

Concurrent multi-agent coding promises division of labor across modules, robustness through redundancy, and parallel exploration at the natural granularity of multi-file projects. Realtime collaborative editing protocols solve this coordination problem for human teams via Conflict-free Replicated Data Types (CRDTs), but the LLMs underneath generate one token at a time and existing multi-agent coding systems inherit this serial limit: they either sequence agents through phase handoffs or pool independent samples without coordination, and a single agent abandons up to half of hard tasks with a one-file stub-and-exit. AgentRoom is a realtime collaborative editing protocol for concurrent coding agents. Its runtime layer exposes file-level claim, status, and broadcast as MCP tools on a CRDT-merged shared filesystem. Five frontier coding-CLI models ran four backend coding tasks, with cross-language checks in Python DevBench and Rust+axum. For CLI-stable models, AgentRoom with 2 agents abandons fewer tasks than Solo and has less run-to-run variation. At matched-compute, one positive mean LLM-judge contrast puts AgentRoom over parallel-merge. The other contrast, a bundle probe, puts full AgentRoom above each partial case: an ordering rather than a percentage split. Coordination, not parallelism or CRDT-merge, bears the load.

The Agent Payments Protocol (AP2), introduced by Google, enables large language model (LLM)-driven shopping agents to authorize and execute payments on behalf of users. Its signed Checkout and Payment Mandates protect the integrity of transaction data after signing. Agent interactions and external inputs that shape a transaction before authorization remain outside that protection, including Agent-to-Agent Protocol (A2A) messages and Model Context Protocol (MCP) tool calls. Prior work identified replay and prompt-injection attacks in AP2 v0.1. AP2 v0.2 addresses some of these issues but adds capabilities and deployment assumptions that require renewed analysis. We present a systematic security analysis of AP2 v0.2 based on its roles, transaction lifecycle, deployment architectures, and trust boundaries. We divide the lifecycle into five phases and identify five deployment architectures. Using MAESTRO (Multi-Agent Environment, Security, Threat, Risk, Outcome), we model four threat actors, eleven attack surfaces, eighteen adversary capabilities, and six attacker goals. The resulting catalog contains 48 threats spanning five attack families. We score these threats with the Artificial Intelligence Vulnerability Scoring System (AIVSS), identifying eight that reach the High band in at least one architecture. Because no complete public AP2 deployment was available, we build a testbed spanning all five architectures and develop five proof-of-concept demonstrations covering all eight High-risk threats and their mitigations. We also develop a deployment-aware scanner that maps applicable threats to static, cross-role consistency, and adversarial checks. Our analysis shows that valid mandate signatures alone do not ensure that an agent-mediated transaction reflects the user's intent when its pre-authorization context is manipulated.

Tax-loss harvesting demonstrates consistent benefits to long-term portfolio growth; yet implementing it efficiently often involves complex considerations that are specific to the holdings within that portfolio and the individual who owns it. We introduce a custom capital gains calculation engine and a RAG-retrieved vector store of market advisory reports to provide context for a multi-agent trade recommendation system. We investigate the effects of each context provider on the quality of recommendations, measured by relative capital gains incurred during portfolio liquidation. A 2x2 repeated-measures ANOVA revealed a significant main effect of the tax optimization engine ($F(1,29) = 9.17$, $p = .005$, $η^2_p = .240$): enabling the engine reduced tax savings by approximately 55 percentage points relative to the no-engine conditions. The RAG main effect was not significant ($p = .841$), nor was the interaction ($p = .553$). The RAG-only condition achieved the highest descriptive mean tax savings (47.7%), and the baseline condition performed second-best (30.6%), suggesting that the pre-trained language model's internalized financial knowledge may be sufficient for competent tax-loss harvesting recommendations without explicit tooling. These results indicate that augmenting LLM agents with domain-specific computation engines does not guarantee improved performance and may introduce conflicting optimization signals.

We study autonomous mathematical discovery in the Station, an open-world multi-agent environment in which AI agents from different model families pursue a shared research goal without a central coordinator or scripted pipeline. Agents choose their own research directions, conduct experiments, collaborate, and build a shared scientific literature. Across 12 construction problems from the AlphaEvolve catalogue and two additional case studies, the Station obtained results novel relative to the prior literature on five problems: a new infinite family of finite-field Kakeya sets, new exact 604-point kissing configurations in dimension 11, new records for the discretized Kakeya needle and sign uncertainty problems, and a substantially improved lower bound for Erdős's minimum-overlap problem. Agents also discovered novel infinite families for Book Ramsey numbers. Importantly, the agents produced not only numerical constructions but also theorems and analyses explaining how those constructions work, making the results more interpretable and easier for mathematicians to build upon. We release all raw agent dialogues, proofs, and verification code, providing a transparent record of how these discoveries emerged.

Everything a language model sees is tokens. The serving stack knows what each span is — user input, tool output, instructions — but the model must keep track of that itself, and it can lose track or be confused: text can be written to read like anything. Prompt injection is a natural exploit of this phenomenon. By scrambling the model's understanding of span identity, an attacker can induce unwanted and potentially dangerous actions. Adding a non-textual channel to the model's input — a way to communicate span identity beyond text — mitigates this class of attack. We thus introduce a general steering technique called Semantic Overlays: small learned adapters applied at chosen prefill positions to a frozen model's residual stream. Laying an overlay over a span creates an out-of-band annotation channel that cannot be replicated by tokens. Unlike steering vectors, Semantic Overlays are trained, adaptable, and selectively applied. An overlay can encode complex semantics that reshape how the model perceives the marked span: asked to copy a code snippet under an overlay asserting that it is in a different programming language than it is, the model rewrites the snippet, faithfully, in the asserted language. Overlays are also composable, allow for transparent reading of underlying content, and can carry complex payloads — including imperatives that the model will follow. An overlay which marks a span as "non-executable" defends against the broad class of prompt injections that add instructions in untrusted context. We report strong results on prompt injection benchmarks: SEP separation rises from 24.3% to 96.5% with utility unchanged (our scoring rule; we also correct a defect in the published grader), TensorTrust attack success rate falls from 34.8% to 6.6%, and all four PIArena attack families drop to 0% compliance, all while marked spans stay readable (92.5% exact copy rate).

Single-token autoregressive decode on CPUs is bound by memory bandwidth, not arithmetic: a modern CPU sustains roughly 1 TFLOP/s of compute but only about 50 GB/s from main memory, and each generated token must stream every active weight once. This report argues that the most effective response is to co-design the model architecture and the inference runtime together. It presents cflow, a CPU-first streaming engine, alongside a family of pipeline-native transformer architectures whose inter-layer dependency graphs are constructed to permit a vertical, stage-major execution schedule. cflow stores weights as L2-sized tiles in compute-consumption order, reads only the top-k experts of each mixture-of-experts layer, fuses projections, and executes a delay-aware schedule from per-model dependency parameters. Across five architectures trained on TinyStories, one (arch2_4_combined) achieves a 2.00x reduction in critical-path weight bandwidth (9.00 to 4.50 MB/token) within 0.24 perplexity of the best candidate, and the tile layout incurs 7.29x fewer L1-data read misses than a row-major baseline. On a 30.9-billion-parameter pipeline-native MoE, cflow decodes at 5.94 tokens/s (tok/s) on a 32-vCPU Ice Lake server, ahead of llama.cpp (4.75) and the vLLM CPU backend (1.65) on comparably sized dense models. Realizing the expert-delay window as asynchronous I/O overlap on a disk-resident expert tier yields a further net win of up to 1.68x, matching the overlap model within 1%. Measurement refutes one of the eight design claims and leaves a second inconclusive; both are reported in full, with the conditions under which they would hold.

Recent Vision-Language Models encode high-resolution images into long visual token sequences, incurring prohibitive prefill costs. To compress them, existing methods score each visual token by averaging text-to-visual attention uniformly across all heads, which assumes every head matches the query. However, our empirical analysis shows that misaligned heads dominate the average, amplifying background tokens and drowning out fine-grained cues. To address this, we propose PAQ (Prompt-Grounded Attention Quality), a metric quantifying how well each head aligns the prompt with image regions. Built on PAQ, our pruning proceeds in three stages. Given a target FLOPs budget, we first partition the transformer layers into groups and allocate a visual token budget to each. Within each group, we then aggregate per-head attention maps via PAQ-weighted softmax into a group-level matrix. Finally, we score visual tokens by this matrix's magnitude and retain the allocated budget per group. By weighting heads with PAQ, our method scores tokens by attention signals that more faithfully reflect prompt relevance, rather than diluting them through uniform averaging. Across 18 benchmarks, our method delivers state-of-the-art trade-offs. Specifically, on LLaVA-1.5-7B (9 tasks), retaining only 5.6% tokens preserves 99.1% of the original performance, surpassing the strongest baseline AutoPrune by 4.2 points. Code is available in https://github.com/baokou-fw2/HAP.

We introduce Giga-Embeddings, a family of text embedding models designed to combine strong retrieval quality with efficient serving. Its largest member is a sparse 10B-parameter Mixture-of-Experts encoder with approximately 1.8B active parameters per token. Across English, Russian, multilingual, and code MTEB benchmarks, this model achieves the strongest aggregate performance within the family on all four evaluated suites. In our vLLM benchmark with 1024-token inputs, it processes 114.5k tokens per second, providing 25 percent higher throughput than the dense 3B model and 1.56-2.65x the throughput of the evaluated external systems. The family also includes a dense 3B encoder and a distilled 480M encoder for tighter compute and memory budgets. We train the compact model using a dimension-agnostic objective that aligns teacher and student similarity distributions. The resulting 480M model scores 70.98 on Russian MTEB, surpassing FRIDA while using 42 percent fewer parameters. We release all three model checkpoints.

Masked diffusion language models (dLLMs) can in principle generate text faster than autoregressive (AR) models, since they denoise many tokens at once. Recent systems have begun building serving infrastructure for dLLMs, but none first measure how these models behave under real, concurrent serving load. Serving systems built without this grounding risk carrying over assumptions from AR serving that may not hold for dLLMs. We characterize dLLM serving to close this gap, using LLaDA-8B-Instruct with a D2F (Discrete Diffusion Forcing) LoRA adapter on a single NVIDIA H200 GPU, evaluated on GSM8K and HumanEval. We report three findings. First, request difficulty, the number of denoising steps a request needs, is discrete rather than continuous: requests fall into 11 fixed step-count levels (178 + 29k), and no signal we test predicts the level before generation starts (best R2 = 0.150). Second, benchmarks with short generation budgets below 320 tokens understate serving variance, since requests are cut off before the latency spread appears. Third, only 24% of single-request wall-clock time is GPU computation; the rest is CPU-side dispatch overhead. Batching mainly helps by amortizing this overhead: sharing one forward pass per denoising step improves throughput by 16.0x at batch size 16 over a per-request-dispatch baseline. We also argue structurally that output quality should not degrade with batch size, stating three assumptions this rests on; we measure 74 to 76% GSM8K accuracy at single-request scale. Finally, we derive a batch-timeout rule for fixed-fill synchronized batching under Poisson arrivals. Together, these results show that serving diffusion language models needs parallelism at the level of each denoising step, which differs from AR serving in how admission and eviction interact with an already shared forward pass.

Storage-backed inference is easy to overclaim: process RSS excludes charged page cache, process-local device readings exclude board-wide use, and successful generation does not establish correct asynchronous reuse. We present a falsifiable certificate separating representation, semantic demand, scheduler requests, and traffic while naming resource authorities, exactness horizons, and tested reuse transitions. At one Qwen3-Next identity, the stock router selects all 48 x 512 managed layer-expert objects during 32K prefill. Their duplicate-free, overlap-free canonical union gives a 43.59375 GiB semantic-demand lower bound, exceeding the declared 34 GiB full-residency envelope and the 11 GiB host-hard plus physical 24 GiB-device envelope. LRU64 execution stays within its host-hard/GPU-audited contract. Against one prespecified zero-cache oracle, all 64 token IDs, every byte of 64 complete 151,936-float logit rows, 3,408 route events, response bytes, and recorded consumer and destination identities are exact. Recurrent-state and upstream-runtime equality are excluded. In a matched source campaign, the buffered path completes exactly but reaches the 11 GiB host ceiling and records 33,481 memory.max events. The blocking one-window direct path and complete eight-window asynchronous component are exact with positive margin and zero limit events. Across six counterbalanced pairs, the complete asynchronous component takes 32.3% of the blocking direct path's wall time at identical physical source bytes per output; the comparison jointly changes queue depth, overlap, and lifecycle implementation. Separately, fourteen prespecified control/fault cells pass across later Qwen3-Next and Gemma 4 binaries, supporting fail-closed behavior only for the named transitions. The principal experiment is one fixed model, workload, runtime, device, and 64-output horizon.

Long-context inference in large language models (LLMs) is increasingly limited by the memory required for the key-value (KV) cache. KV cache compression addresses this problem by reducing the storage cost of previous tokens. Among existing approaches, low-rank compression is particularly attractive because it represents every token in reduced dimensions. Previous low-rank methods typically derive fixed projection spaces from model weights, construct fixed spaces from calibration activations, or construct a shared basis over a broad cache region. Such representations may not capture detailed but important information. We partition each per-head KV cache into fixed-length logical pages and observe substantial low-rank structure within individual pages. Based on this observation, we propose PuzzleKV, a training- and calibration-free method that treats each completed page as an independent compression unit. PuzzleKV decomposes pages within each layer and KV head, computes attention directly over dense and factorized pages, and incrementally compresses newly eligible pages during autoregressive decoding. Experiments across models, context lengths, and benchmarks demonstrate the effectiveness of PuzzleKV under matched storage budgets. At approximately 60% of the original KV cache storage, PuzzleKV achieves more than 96% of Full KV performance across both evaluated models and all benchmark settings, with substantial gains over Global SVD on RULER and competitive performance on LongBench. To achieve a more aggressive compression ratio, PuzzleKV can be further combined with quantization while retaining more than 93% of Full KV performance using only 18.7% of the original storage.

The key-value (KV) cache is a primary capacity and bandwidth bottleneck in long-context LLM serving. We present Minima-KV, a retention-preserving hierarchy for mixed-format paged attention. Recent and protected Anchor pages remain in FP8, while older non-anchor pages move to packed TQ3; every live-request page remains addressable. Format-specific kernels compute partial attention states and combine them through a globally normalized online-softmax merge, enabling direct heterogeneous decode without a cache-sized dense shadow. Across separate, configuration-bound Qwen3.6-27B profiles on a single 96-GB NVIDIA RTX PRO 6000 Blackwell GPU, deployment accounting reports 18.3 KiB of attention KV per live token, corresponding to 3.50x compression relative to BF16 and 1.75x relative to FP8. A materializing quality profile matches its dense control on 16K RULER needle-in-a-haystack tasks. On the same 503-question LongBench v2 set, measured deltas are -0.80, -0.60, and -0.40 percentage points at 16K, 32K, and 64K. A separate single-pair direct-decode canary with two 59,008-token requests measures 3.625x active-KV compression and 0.9821x throughput relative to its control, routes all 16 full-attention layers without fallback, and retains no dense shadow. These results establish a practical mixed-format path for compressing long-context state without evicting live-request KV pages.

An LLM serving engine sizes its key-value (KV) cache once, at startup, permanently setting aside a reserve for the worst-case prefill activation. During decode-dominant phases that reserve sits idle, yet it cannot be handed to the KV pool because it is exactly the memory a large prefill needs. We ask whether this reserve is reclaimable, and build a mechanism to test it. Our elastic KV cache lends the reserve to the KV pool during decode and returns it before prefill, driven by the scheduler's one-step-ahead view of the next batch. It is pure userspace on the CUDA virtual-memory path: two physical handles mapped into one contiguous virtual range per layer, so the attention kernel is unchanged and no driver patch is required. It decommits in a few milliseconds and recommits in tens of milliseconds, works with CUDA graphs and prefix caching, and never triggers an out-of-memory event. A static commit of the same memory is unsafe, crashing on prefill bursts, which makes the dynamic toggle necessary. Having built the mechanism, we test the premise it rests on and report an honest negative result. It only pays off if a small prefill chunk size badly hurts prefill latency. In a controlled experiment injecting long prompts into a live decode load, that penalty is small (median time-to-first-token differs by about 1% between chunk sizes of 8192 and 32768 tokens), because prefill is compute bound and decode consumes only about one token per sequence per step. Simply lowering max_num_batched_tokens recovers more KV than the controller does, at nearly equal latency. The reserve also dilutes under tensor parallelism, from 16% of KV at TP1 to 2.7% at TP4. We state precisely when reclaiming the reserve could still help, and release the mechanism as a reusable userspace elastic-VMM allocator.

Grammatical knowledge and how it is empirically tested are typically considered robust to the frequency of the lexical items in the expressions. However, neural network-based models of grammaticality exhibit high sensitivity to lexical frequency. We draw upon Complementary Learning Systems theory to test the hypothesis that robustness to lexical frequency can arise via a hippocampal episodic memory mechanism, which enables rapid encoding and retrieval of specific experiences and allows learners to leverage them when processing rare patterns. We use retrieval-augmented language models as an instantiation of such an episodic memory mechanism (specifically, $k$-nearest-neighbor language models that augment parametric models with explicit instance storage), and test whether this augmentation helps close the lexical frequency gap that vanilla language models exhibit in syntactic contrast tests. Using syntactic contrasts with frequency-stratified test items, we find that retrieval augmentation narrows the performance gap between high- and low-frequency items, consistent with episodic memory compensating for weak parametric representations. This benefit is consistent across different syntactic phenomena and across models pretrained on child-realistic and large-scale data. Additionally, we show that structural information is critical for effective retrieval, whereas semantic similarity alone provides little benefit. While these are promising proof-of-concept results supporting our hypothesis, the frequency gap is narrowed rather than fully closed. Based on our analyses, we propose preferential reweighting of retrieved instances, better representations and retrieval strategies for structural information, and flexible configurations of storage and retrieval as promising future directions for improving the implementation of episodic memory in language models.

Aligning large language models to human-centered objectives is difficult when targets are non-executable and context-dependent, limiting reliable verification and scalable supervision. Although synthetic data expands coverage, weak verification shifts the bottleneck from generation to selection. Noisy signals destabilize iterative refinement and can cause silent regressions. We propose Agentic Data Evolution (ADE), a data-centric framework that organizes synthetic supervision as evolving data snapshots. ADE improves data snapshots through a closed-loop Observation-Variation-Selection (OVS) procedure, where a steady-state admission mechanism acts as a quality ratchet that conservatively gates updates for sustained cross-round improvement. We validate these improvements through complementary intrinsic trend tracking and extrinsic post-training evaluation. On DEV300, ADE raises the intrinsic win rate from 50% to 75.81% and the extrinsic win rate from 55.20% to 68.86%, consistent performance gains across diverse benchmarks. Blind expert evaluation further confirms this, with a 66.11% preference for evolved answers. These gains extend across post-training methods, model scales, and tasks beyond the target weakly verifiable educational objectives. Resources are available at https://github.com/ZeroLoss-Lab/Agentic-Data-Evolution.

Transformer-based language models are widely used as models of human language processing, yet their attention mechanisms allow lossless access to the full preceding context, unlike the limited memory systems of humans. We hypothesize that installing memory constraints into transformers' attention mechanisms can improve their fit to human behavioral data. While previous work has explored individual constraints in isolation, we conduct a systematic comparison of multiple attention-based memory mechanisms across different model sizes and training corpora, evaluating both psychometric predictive power for human reading times and grammatical competence. We additionally compare static constraints, in which the constraint strength is fixed throughout training, to dynamic memory curricula. We find that constraints that are sensitive to the content of intervening tokens consistently achieve the highest alignment with human reading times, outperforming distance-based constraints. We observe a dissociation between psychometric fit and grammatical competence under dynamic memory curricula, suggesting that Transformers cannot serve as a one-size-fits-all cognitive model.

Data mixing is a central design problem in large language model pretraining: given a fixed token budget, practitioners must decide how much data to allocate to each domain. Recent proxy-based methods address this problem by training small models on candidate mixtures, fitting a response model, and using the response to select mixtures for larger-scale training. We show that this workflow has the structure of a classical mixture experiment. Under this view, data domains are mixture components, token shares are component proportions, proxy-training runs are experimental design points, and validation loss defines a response surface over the probability simplex. We develop this formulation using sparse second-order Scheffé response-surface models and construct model-robust $\mathcal{I}$-optimal designs for proxy data-mixing experiments. Using RegMix as an empirical case study, we demonstrate how the framework can both interpret observed mixture responses and design more efficient proxy experiments. The Scheffé analysis shows that domain value is strongly relational: several domains that are weak under additive effects become favourable through pairwise interactions, especially through combinations with web-derived text. The sparse Scheffé model preserves mixture rankings across model scales and remains competitive with a flexible machine-learning predictor while providing an explicit decomposition of additive and interaction effects. In a simulation study calibrated to observed proxy-training responses, model-robust $\mathcal{I}$-optimal designs recover the relevant mixture ordering after removing about 25% of the original proxy runs. These results suggest that LLM data mixing should be treated not only as a prediction problem, but also as an experimental-design problem in which the proxy mixtures themselves can be chosen to improve statistical efficiency.

Large Language Models (LLMs) demonstrate strong capabilities in automated essay scoring (AES), but contemporary approaches typically employ fixed prompt selection, failing to address operational cost concerns and evolving optimal configurations. We propose a novel cost-aware approach that treats each prompt type as an arm in a multi-armed bandit (MAB) controller, enabling adaptive selection of optimal prompting strategies during inference. Our experiments on IELTS Writing Task 2 essays show that the MAB framework achieves comparable scoring accuracy to exhaustive grid search while reducing LLM calls by 78.4% to find the best grading approach. We implemented four distinct grading recipes (multi-step vs. single-step assessment, with vs. without calibration examples) and found that the multi-step approach with examples achieves the highest accuracy. By tracking token usage and latency alongside agreement metrics, we produce the first cost-reliability learning curves for essay scoring, providing actionable insights for educational technology platforms that must balance operational costs against assessment validity. This work represents the first application of online control mechanisms to adaptively select prompting strategies in AES, transforming prompt selection from an offline hyperparameter optimization problem into an efficient online learning task.

Large language models (LLMs) are known to exhibit social sycophancy, often validating or agreeing with users in socially sensitive contexts. Existing evaluations typically measure sycophancy under a fixed prompt formulation, leaving unclear whether such behavior is stable when the same underlying situation is presented with different sycophancy-relevant prompt variants. In this work, we study sycophancy prompt sensitivity: the extent to which changes in user confidence, emotional framing, social consensus, or validation-seeking language alter a model's sycophantic behavior. We refer to our evaluation framework as SyPS, short for Sycophancy Prompt Sensitivity. Building on existing social sycophancy evaluation settings, SyPS constructs controlled prompt variants that preserve the same underlying user situation while varying sycophancy-relevant social cues. We introduce the Sycophancy Prompt Sensitivity Score (SPSS), an instance-level measure of sycophancy variation across paired prompt variants. Unlike aggregate sycophancy rates, SPSS separates baseline sycophancy from prompt-induced shifts, enabling model-level comparisons of robustness to sycophancy-relevant social cues. Empirically, we find that sycophancy prompt sensitivity is socially structured: validation-seeking and emotional-pressure cues often increase sycophancy, whereas counter-framing and anti-sycophancy prompts tend to reduce it. Our framework highlights whether LLMs maintain stable social judgments while adapting appropriately in tone.

Large language models (LLMs) are becoming routine instruments of scientific research, assisting with literature synthesis, hypothesis development, coding, and formal reasoning. Their use raises a central epistemic question: when parts of scientific reasoning are delegated to an artificial system, what conditions must remain under human control for the resulting knowledge claims to retain epistemic legitimacy and accountable authorship? This paper develops a normative and conceptual framework for analyzing such delegation. Scientific reasoning is treated as a distributed process in which the origin of a contribution may vary between human and machine, while responsibility for its acceptance into the scientific record remains human. The framework distinguishes content origin $O(g)$, completion of human verification $V(g)$, responsibility assignment $R(g)$, accountable human ownership $M(g)$, and epistemic outcome $E(g)$. These constructs separate the provenance of a claim from the process by which it is checked, the epistemic outcome of that checking, and the human responsibility attached to its disposition. The central proposition is that the ethical boundary of LLM-assisted research is determined primarily by adequate verification and accountable human ownership rather than by the degree of machine involvement itself. On this basis, the paper develops the notion of an epistemic audit: a structured record of delegation, verification, provenance, and responsibility intended to make AI-assisted reasoning transparent and reviewable. The resulting framework provides a formal vocabulary for distinguishing responsible cognitive delegation from the transfer or neglect of epistemic responsibility in scientific research.

Large language models (LLMs) are increasingly deployed in education, healthcare, policy advising, and other interactive settings, where users engage them as sustained social interlocutors rather than one-shot query engines. This shift makes jailbreaks a growing safety threat, yet most research emphasizes single-turn prompt optimization or iterative attack refinement, leaving psychologically grounded multi-turn vulnerabilities underexplored. We present PsychJail, a psychology-guided framework for red teaming aligned LLMs through theory-grounded, multi-turn persuasion. PsychJail maps established social-psychological persuasion techniques into a tactic-conditioned attack policy. It factorizes each attacker action into a Change-of-Meaning analysis, tactic selection, and victim-visible message, operationalizing the Persuasion Knowledge Model (PKM). The policy is refined with trajectory-level reinforcement learning using a PKM-gated reward that credits early jailbreak success only when every turn contains a well-formed Change-of-Meaning analysis. Across four aligned victim models, PsychJail achieves the highest average attack success rate (87.3%) and outperforms strong single-turn and multi-turn baselines on every model. We also measure susceptibility at the action that breaks each victim, revealing four distinct model-level fingerprints that identify which persuasion levers affect each model and how broadly. These fingerprints help explain cross-model transfer asymmetry. We interpret them as four candidate psychological profiles-rationalist, credibility-driven, narrative-monoculture, and broadly persuadable-while treating this interpretation as a conjecture requiring future validation. Our findings establish psychological jailbreaks as a distinct red-teaming frontier for increasingly interactive LLMs.

Self-reflection is a powerful mechanism for credit assignment in human learning, converting sparse outcome feedback into actionable guidance. However, its potential for post-training Large Language Models (LLMs) remains underexplored. We propose Self-Reflective Policy Optimization (SRPO), a framework that internalizes this capability. SRPO enables LLMs to analyze their own completed trajectories, synthesize errors into concise "reflection patches," and use reflection-conditioned teacher scores on student on-policy rollouts as dense token-level training signals. This process effectively transforms sparse terminal supervision into dense, token-level learning signals without requiring external critics, separate reward models, or larger teacher models. We demonstrate that SRPO achieves state-of-the-art performance across mathematical reasoning and long-horizon agentic benchmarks with exceptional data efficiency. Using a Qwen3-8B base model, SRPO attains 73.3% on AIME'24 using only 8% (0.08x) of the training FLOPs required by scaled supervised fine-tuning, while significantly improving success rates on WebShop (64.7%), ALFWorld (76.8%), and SWE-Bench-Lite (31.2%). Code is available at https://github.com/Galleons2029/SRPO

Group-based reinforcement learning methods such as GRPO for large language models avoid training a critic by sampling multiple responses for each prompt. A reliable critic could instead estimate token-level advantages from one response, but standard critic-based training recipes are often unstable. We study this instability and develop **Best Practice Critic Optimization (BPCO)**, a recipe that combines DPPO, value predictions bounded to the reward range, Monte Carlo value targets, unnormalized policy advantages, and length-adaptive generalized advantage estimation. Because the critic is used only during training, BPCO can also condition it on reward-defining information, such as a reference answer or grading rubric, that is hidden from the policy. Controlled experiments isolate the effect of each design choice. Across mathematical reasoning tasks with models ranging from 1.5B parameters to 30B-A3B mixtures of experts, BPCO improves a strong critic-based baseline consistently, and matches or exceeds a group-based baseline while sampling one response per prompt. The same recipe also improves learning with rubric-based rewards. These results show that a carefully designed critic provides a reliable alternative to group-relative advantage estimation. Code is available at https://github.com/QPHutu/golden_critic.

Fast and reliable plasma equilibrium prediction is essential for real-time tokamak operation and control, but conventional Grad-Shafranov (GS) solvers are often too costly for real-time deployment. We develop an AI surrogate framework and benchmark five architectures (MLP, CNN, FNO, Transformer, and KAN) on a numerical GS database with 100,000 IID and 10,000 OOD samples. Under a unified protocol, we evaluate accuracy, inference efficiency, model scaling, and robustness. We also establish device-level validation on the EXL-50U tokamak by linking numerical GS solutions, surrogate predictions, and the standard Shape Editor reference to assess simulation-to-device consistency. The surrogates achieve errors of $10^{-3}$-$10^{-2}$ relative to GS solutions, while the GS-to-device discrepancy remains at $10^{-3}$. Transformer gives the best IID accuracy, whereas CNN offers the best balance of accuracy, robustness, and speed, reaching 0.7 ms TensorRT latency. On unseen plasma geometries and parameter regimes, CNN and FNO show the strongest extrapolation stability, with 4%-5% relative $L_2$ error, while models with weaker inductive biases degrade more substantially. Scaling data and model capacity improves interpolation but not necessarily extrapolation, revealing a trade-off between capacity and OOD generalization. Overall, this work provides a systematic, device-consistent benchmark for AI-based GS prediction and practical guidance for selecting reliable surrogates for real-time plasma control and fusion applications.

Modern AI models such as tabular foundation models and gradient-boosted ensembles can outpredict classical methods, but provide little basis for reasoning about their predictions. High-stakes decisions call for models that are both accurate and interpretable as built. Local linear modeling offers a path forward: a smooth regression function is locally well approximated by a linear one, allowing a linear fit near each query point to achieve high accuracy without sacrificing transparency. The challenges lie in learning what is "local" and developing statistical tools for interpretation. Here, we propose local distillation, in which a black-box "teacher" guides a regularized linear "student" model at each query point. The teacher (1) defines locality by upweighting training observations with similar predicted outcomes, and (2) anchors the fit with its prediction at the query point, included as a pseudo-observation whose weight is estimated from the data. For interpretation, we add a small amount of Gaussian randomization to the local objective and use refits to assess stability: selection frequencies identify reliable features at a query point, and clustering the randomized fits identifies stable subgroups across the data. Under the lasso penalty, we prove that this randomization yields feature-selection probabilities that are stable under small perturbations of the training responses. Across 17 benchmark datasets, local distillation nearly matches its AI teacher's accuracy while producing a sparse linear model at each test point. In a high-dimensional cancer gene expression example, the framework identifies patient subgroups whose local models use different genes; this heterogeneity is invisible to a global linear model, and difficult to surface in a black-box model.

Iterative preference optimization is essential for aligning Large Language Models on mathematical reasoning tasks, yet its efficiency is often throttled by signal scarcity: as the model improves, static problem sets become increasingly mismatched to the model's evolving competence, producing rollouts that are either too easy or too hard and therefore non-informative, which leads to a scarcity of valid preference pairs. We propose DIAG, a Diagnostic Iterative Alignment and Generation framework that adaptively reshapes the practice distribution to increase informative supervision and focus training near the student's current competence boundary. DIAG consists of two phases: (1) diagnosing valid preference-pair yield to calibrate the exploration-exploitation trade-off and allocate topic quotas via an Empirical Bayes shrinkage estimator, thereby prioritizing high-yield concepts; and (2) generating targeted practice, where a teacher synthesizes variants from the student's failure traces. We further provide a theoretical view interpreting DIAG as a teacher-mediated approximation to KL-regularized reweighting of the practice distribution toward the student's competence boundary, where valid preference-pair yield is maximized. Experiments show that DIAG boosts yield across iterations and delivers stronger reasoning performance under an iso-effective training budget, demonstrating that it can distill more informative preference supervision for mathematical reasoning.