每天知道
你的领域
出了什么

On-device LLM agents interact repeatedly with users on local hardware, producing private traces that are valuable for adaptation but should not be sent to a remote trainer. Ideally, such agents would learn contiguously---adapting from every interaction without pausing or suspending user-facing inference---yet existing inference runtimes assume stable weights and existing RL systems assume separated resources, so neither can support this continuity. We present LOCAL, the first single-GPU runtime that enables contiguous on-device learning for LLM agents. The key insight is that GPU scheduling, adapter version management, and KV-cache validity cannot be handled by independent subsystems: adapter updates invalidate cached KV tensors from older versions, and cache retention affects the memory available for training. LOCAL makes adapter version, task priority, and cache state visible to three cooperating components---a cooperative scheduler, a version-aware KV-cache manager, and a multi-agent model runtime---that share this state to keep scheduling, execution, and cache maintenance mutually consistent. On a single 24 GB GPU with 7B-class models, LOCAL lowers foreground queue-wait p95 by 3.1x over FIFO, lowers p95 time-to-first-token (TTFT) by 1.55x versus non-preemptible training, cuts post-publish first-hit prefill p99 by 25.6% and cross-agent TTFT p99 by 21.9%, and keeps background learning progressing under tight KV budgets.

Wide-area bandwidth per unit of GPU compute falls every hardware generation: in compute-intensity-ratio terms (CIR, bytes per FLOP), the gap between on-package memory and the conventional WAN is four to five orders of magnitude, widening at roughly 12-19% per year. Position papers - including our own - argued this makes elastic optical wide-area capacity necessary for cross-site AI inference. Reviewers correctly objected that such arguments show more bandwidth helps, not that it beats the alternatives: KV recomputation, cache compression, locality-aware routing, scheduling, or an overprovisioned packet backbone. This paper does the comparison. We derive a workload model predicting when moving inference state across sites beats recomputing it - a context-independent crossover at 74-111 Gbps per stream for a 70B multi-head-attention model, falling to 9-14 Gbps under grouped-query attention at 1/8 KV heads - and quantify five sensitivity axes: context length, attention architecture, queueing, agentic compounding, and loss/jitter-induced bandwidth collapse. On economics: at list GPU prices recomputation is cheaper; transfer wins when GPU scarcity and KV reuse multiply effective GPU cost by roughly 5-20x, and modern attention moves the breakeven an order of magnitude in transfer's favour. We position the network levers correctly rather than adversarially: packet networks allocate lit capacity at millisecond timescales; optical fungibility changes how much capacity is lit, at minute timescales, substituting for overprovisioning economically rather than functionally. Finally we specify a ten-metric measurement plan on a three-site production-fibre testbed, framed as an open ecosystem exercise: no single company can - or should - assemble this evidence alone. Every claim is bounded by the regime in which it holds; several findings weaken the naive version of our own thesis, and we state them.

Analytical models of peak VRAM consumption for LLM inference decompose memory into weight-storage, KV-cache, and activation terms parameterized by step count, tool invocations, and context expansion. We evaluate this decomposition empirically within a strictly scoped measurement study: a LangGraph-based CUDA-kernel-synthesis agent (AgentK), a 4-bit quantization family (Q4 K M), a single NVIDIA H100 GPU, and four LLM backbones across 1,920 trajectories. Focusing on peak-memory forecasting behavior, we report two primary observations. First, closed-form analytical models achieve competitive accuracy when provided with two empirical constants: loaded-weight VRAM and a fixed activation-memory overhead. Supplied with live GPU readings and ground-truth trajectory parameters, the closed-form model matches or outperforms the best learned baseline on three of the four backbones (test MAPE 2.2-4.4% vs. 3.4-6.5%, p = 0.76). The exception is the smallest backbone (Phi-4-mini), where minimal VRAM variance (CV 0.3%) causes dynamic modeling to underperform simple regression. Second, compile success strictly bifurcates by backbone capacity (from 5.7% for Phi-4-mini to 62.0% for Qwen2.5-Coder-14B), demonstrating that functional code synthesis remains constrained by intrinsic LLM capabilities rather than available memory. Furthermore, because overall peak-memory variance is remarkably low across all backbones (CV 0.3-9.4%), learned prompt-feature regression offers statistically insignificant improvements over a constant-mean baseline. Consequently, we find no justification for deploying complex predictive VRAM models in highly quantized, weight-dominated regimes. We release the evaluated corpus and anonymized framework to support replication.

Agentic applications are shifting AI serving from isolated model inference to long-running workloads in which LLMs coordinate tools, environments, and persistent state. However, the system behavior of these workloads---where latency, cost, and bottlenecks arise---remains poorly characterized, leaving serving systems to rely on assumptions built for conventional inference. We present AgentSysBench, a benchmark suite and measurement toolkit with ten representative agentic applications and unified systems-level instrumentation. Across controlled deployments and production traces, we identify six properties that distinguish agentic workloads from conventional LLM serving: (1) execution is heavyweight and stateful, with non-LLM components dominating latency in 5 of 10 applications and sandbox working-set memory peaking at 28 GB per session; (2) applications compose components with heterogeneous resource affinity---GPU-bound inference, memory-bound retrieval, CPU-bound sandboxes---whose task latencies diverge by up to 32x; (3) bottlenecks shift across requests, models, and deployments; (4) production sessions hold state idle for minutes to hours between active steps; (5) a control-plane tax---auxiliary LLM calls and context overhead from tool schemas and observations---crowds out productive compute and context; and (6) production traces from three applications reveal heavy cross-request redundancy in search queries and web fetches, exposing a large caching opportunity. Four design explorations demonstrate that these findings are actionable: task-aware serving reduces latency by 29--40%, communication-aware placement by up to 4.5x, state offloading reduces memory usage by 4.6x, and tool-result caching removes 35.2% of redundant search calls and saves 19.3% of aggregate search latency.

Long-context LLM applications such as retrieval-augmented generation (RAG) and agentic systems often process tens of thousands of input tokens to produce short outputs, making end-to-end request latency an important serving objective. We show that the maximum number of batched tokens (MBT), which controls the token scheduling budget in vLLM, has a scheduling-pressure-dependent effect on latency. Larger token budgets can reduce latency under low scheduling pressure, while smaller budgets become preferable under higher pressure. Consequently, no single static MBT performs best across load regimes. We introduce Prefill-Pressure Adaptive Scheduling (P-PAS), a lightweight policy that dynamically adapts the scheduling budget based on concurrent prefill and decode state. P-PAS retains a large token budget under low pressure and constrains prefill work as pressure increases. Across models, workloads, and GPUs, P-PAS maintains low end-to-end latency across changing load regimes, avoiding the limitations of a fixed MBT. Kernel-level profiling shows that large prefill chunks can improve execution efficiency under low scheduling pressure, but that this advantage varies across model--hardware configurations. As scheduling pressure increases, smaller chunks can instead reduce interference with active decoding, explaining the observed load-dependent MBT sensitivity. Code and artifacts for reproducing our results are available at https://github.com/TimoSaemann/ppas-vllm .

Sparse mixture-of-experts (MoE) language models reduce arithmetic by activating only a small subset of experts per token, yet deployment still requires storing and moving the full expert bank. We present ExactMoE, an inference design that applies symmetric group-128 four-bit weight quantization only to routed experts, stores those experts in kernel-native MARLIN form in pinned host memory, and executes all selected experts through a configurable GPU-resident slot cache and fused grouped MoE kernels. The router, attention, embeddings, normalization layers, and language-model head remain in BF16. "Exact" refers to complete expert availability and an unchanged top-k routing procedure: no expert is pruned, substituted, or forced to execute on the CPU. It does not imply numerical identity with the BF16 model. On OLMoE-1B-7B-0924-Instruct, evaluated on a single NVIDIA L4, a 16-slot configuration reduces peak reserved GPU memory from 14.168 to 1.836 GiB (87.04%) while retaining 81.85% of BF16 decode throughput. A fully resident 64-slot configuration reaches 31.923 tokens/s versus 21.662 tokens/s for BF16 while reserving 4.061 GiB. Across 12,450 zero-shot multiple-choice questions, ExactMoE obtains 70.3534% normalized accuracy versus 70.8996% for BF16, retaining 99.23% of the baseline accuracy. In a matched 16-token ablation, fused grouped execution is 1.97x as fast as a sequential W4 reference. These results identify a practical memory-transfer-throughput frontier for complete-expert MoE inference.

Scaling transformer language models creates an inherent tension between expressivity and memory efficiency. While unique weights across layers preserve functional specialization---from input-grounding to abstract refinement---they incur a substantial memory footprint. Conversely, standard depth-sharing enforces uniform transformations that collapse representational diversity and degrade modeling quality. We introduce RecurrentGPT, a recurrent depth transformer where fixed-depth prelude and coda blocks bracket a single shared core iterated R times. Inspired by gated recurrent neural networks, we employ a lightweight projection and an elementwise update gate---conditioned on the hidden state, the fixed prelude output, and noise resampled at every step---to modulate the recurrent update. This allows the model to specialize the input to the same few layers across recurrences, rather than requiring many unique layers to achieve functional diversity. Under an isoFLOPS constraint, a 3-layer RecurrentGPT matches the accuracy of a 12-layer GPT-2 Small baseline with similar training and inference FLOPs, and leads MoR and heavy-tail depth sampling in all nine scale-by-budget cells; at medium and large scale it approaches dense quality at the standard token budget and overtakes it at medium scale once that budget is doubled. Under an isoPARAMS constraint, deeper recurrence achieves a 2.76 validation loss versus 2.84 for a non-recurrent counterpart at matched parameter and data budget. Our results demonstrate that adaptive depth reuse is a principled strategy for trading parameters for quality: at large scale, 63% fewer parameters and 59% less peak decoding memory for a 10% increase in compiled generation latency.

Deploying large language models (LLMs) for inference on edge devices is challenging due to severe memory and bandwidth constraints. While speculative decoding and Mixture-of-Experts (MoE) have been proposed to improve inference efficiency, naively combining them often incurs excessive verification overhead and poor expert reuse, limiting their effectiveness in memory-bound edge settings. In this work, we propose S2-MoE, an efficient self-speculative decoding framework for MoE inference on edge devices. S2-MoE reduces redundant verification through routing-aware adaptive speculative expansion, improves verification efficiency with reuse-aware expert gating, and aligns draft and target execution via shared context. Implemented in llama$.$cpp, S2-MoE achieves up to $5.3\times$ speedup (about $2.0\times$ on average) over standard autoregressive decoding across diverse MoE models and datasets on edge devices. Code is available at https://github.com/angerybob/S2-MoE.

Understanding a model and its learning mechanisms in depth requires identifying when its internal structures become useful, rather than simply looking at the final state. We study this through concept dynamics: at each layer and checkpoint, we decompose activations, select sparse soft masks, and inject masked reconstructions into the model. Concept analysis is therefore tested functionally: a mask is useful only insofar as it preserves a target under intervention. We compare sufficiency for activation reconstruction, linear decodability, true downstream preservation, and checkpoint transfer under learned alignment. The framework treats decomposition assumptions as hypotheses rather than interpretability guarantees, monitoring functional sufficiency across checkpoints and source-to-final reconstructability under learned alignment. At the shared fixed-penalty operating point across seven models, downstream masks retain substantially less soft mass than reconstruction masks; predictive-distribution shifts remain small.

Chain-of-thought monitoring is a potentially useful safety signal, but its reliability across languages and behavioral settings remains uncertain. In a small case study of eight manually verified synthetic scenarios, one model, one annotator, and one deterministic generation seed, I study API-visible reasoning during indirect prompt injection in Sarvam-105B across English, Tamil, and Tanglish. A four scenario pilot found 5/12 injected attack successes without reasoning and 1/11 with reasoning. A preregistered four-scenario follow-up reversed that direction, finding 2/12 attacks without reasoning and 3/12 with reasoning. With only four scenarios per phase, this design cannot distinguish a real reasoning-mode effect from prompt-specific variation or sampling noise. Across 20 non-empty injected-thinking traces, all 17 benign-correct outputs stated an intent to ignore the injection, while all three attack successes stated an intent to follow it. These descriptive observations provide a reproducible case study of behaviorally informative visible reasoning when it is available; they do not establish that reasoning mode improves safety, that visible reasoning is mechanistically faithful, or that the findings generalize beyond this configuration.

Frontier AI models have advanced rapidly, but they still struggle with telecom-specific tasks. We present Open Telco (OTel), an open telecom AI resource with derived datasets for retrieval, reranking, instruction tuning, and safety/abstention, plus 30 full-parameter post-trained baselines across embedding, reranking, and language models. The community has already engaged substantially with the resource: as of May 3, 2026, the released models have been downloaded over 16 million times, and the project has received 157+ pieces of media coverage worldwide. Building on prior open telecom datasets and benchmarks, OTel provides documented telecom data sources, held-out evaluation partitions, trained embedding models, rerankers, context-grounded LLMs, and safety/abstention data in one unified resource. OTel post-training improves performance across all three model families: embedding retrieval reaches 93.5% NDCG@10, reranking reaches 0.952 MRR@10, and language-model correctness reaches 88.2%. We release OTel as a reproducible starting point and invite the community to expand the data, improve embedding and reranking models, and build stronger context-grounded telecom LLMs.

Large language models (LLMs) are increasingly used to score text records at scale (e.g., rating candidate resumes on a 1-5 scale). However, existing LLM-powered approaches do not account for the fact that effective scoring requires both holistic understanding of records and locally consistent judgments across similar ones. We present Attune, a mixed-initiative system for steerable LLM-powered scoring. Given a task description and scoring range, Attune performs pairwise comparisons across records to develop a global understanding first, and then resolves these comparisons into consistent score assignments-deriving scoring criteria and rules bottom-up in the process. These serve as shared representations of scoring logic that users can inspect and edit. Based on insights from a formative study (n = 12), Attune's interface introduces novel steering interactions that allow users to deterministically refine scoring logic. Users can provide examples, directly edit criteria, rules, or target distributions, and give natural language feedback-with all refinements compiling into constraints that guide re-scoring. We validate our approach through a technical evaluation across three workloads and a user study with domain experts (n = 8) in healthcare, law, education, and AI evaluation.

Generative AI systems are increasingly producing real-world artifacts, however their efficacy and validity are often evaluated via context-free LLM-scoring. These judges can be miscalibrated by irrelevant in-context reference examples, creating false confidence and allowing low-quality or harmful outputs to pass evaluation. We study this failure mode as context-induced miscalibration and introduce DA-RAC, a distance-aware reference-anchored calibration method for LLM judges. DA-RAC retrieves semantically and structurally similar labeled anchors for each judgement scenario, weights them by distance, and exposes neighborhood difficulty as a calibration and triage signal. On multi-run LLM-judge evaluation benchmarks, it improves calibration and reduces false-pass risk relative to zero-shot, chain-of-thought evaluation, and static-anchor baselines. Mechanistic analysis shows that judge scores vary systematically with anchor distance, while static references can induce misleading decision boundaries. Thus LLM-judgement requires not only better models, but also calibrated, auditable reference selection, especially when automated evaluation is used to support high-impact AI generated artifacts. Judgments should be grounded in relevant, inspectable, and contestable interpretive artifacts.

Battery energy storage systems (BESS) are increasingly used in distribution networks for voltage regulation and demand response, which increases the volume and complexity of operational telemetry available to grid operators. This paper presents an AI-enabled monitoring framework that connects a large language model (LLM) interface with a structured telemetry database for BESS-integrated distribution system analysis. Operator questions are submitted in natural language and translated into validated SQL queries using predefined database schema information and approved KPI views. Retrieved measurements, including bus voltages, state of charge, active power, and reactive power, are evaluated against engineering constraints for voltage limits, BESS operation, and demand response tracking. The framework is validated using hardware-in-the-loop co-simulation data from a BESS-equipped distribution feeder operating under reactive power-based voltage control and price-driven demand response. Case studies show that the framework generates valid database queries, identifies repeated voltage violations, detects reactive power overshoot, and evaluates active-power tracking performance. The results show that LLM-assisted monitoring can connect structured grid telemetry with automated engineering assessment for BESS operation analysis.

Recent advances in Large Language Models (LLMs) have opened opportunities to apply high-level code transformations to the field of code optimization, and it has since emerged as one of the most fundamental tasks for LLMs to perform; however, at present, LLMs struggle to apply wide-ranging code optimization tasks due to both the complexity of the code and the inability to independently verify the correctness of the transformations. In this paper, we present the Trusted LLM (T-LLM) Compiler, which proposes an advancement in compiler technology through a collaborative effort involving high-level LLM code transformations, traditional compilers, and verification tools. Experimental results reveal that it can significantly improve code correctness when tested on a set of PolyBench/C benchmarks. Our approach facilitates iterative code optimization efforts with verification strategies that enable corrective actions. Through this approach, T-LLM Compiler achieves code optimization accuracy of up to 83.3% and a speedup of up to 16.1% on the PolyBench/C benchmarks, with the transformed code reaching an average of 26.7% speedup wrt standard baselines. Additionally, we release the project's source code to the open-source community.

Power-law anisotropy in internal representations has been observed across a wide range of biological and artificial neural systems, from state-of-the-art language models to the mouse cerebral cortex. This anisotropy is a key geometric property of high-dimensional information processing and underlies a variety of theoretical analyses. However, the mechanism by which it emerges from input structure and task-driven learning has remained unclear. Here, we characterize this formation process by exactly solving the learning dynamics of a wide two-layer linear neural network in a teacher--student setting with power-law input and teacher structures. We show that, in the feature-learning regime, the local power-law exponent of the internal-representation spectrum evolves nonmonotonically over the course of training and exhibits up to four distinct asymptotic regimes across modes and training times. By contrast, in the lazy regime, the exponent remains essentially unchanged. We further demonstrate numerically that similar exponent dynamics arise in more realistic nonlinear networks. Together, these results suggest a general mechanism by which the dynamic interaction between input statistics and task structure gives rise to power-law internal representations.

Large language models increasingly rely on sampling as a driver of their own improvement, making the fidelity of their learned distributions more critical than ever. Yet, not all distributions are equally easy to learn. In this work, we identify a curse of ambiguity: in large language models, and more broadly in all neural networks that produce discrete probability distributions, the more ambiguous a next-token distribution is, the harder it is to learn accurately. Through an extensive theoretical analysis, we trace this curse to architectural and learning roots. More ambiguous distributions require more capacity to be stored, larger embeddings to be represented, more steps to be fitted, and amplify token-sampling noise. We validate these findings on synthetic tasks with controlled ground truth and observe the same signatures in language models trained on real data. Our results provide a new perspective on the statistical capabilities of large language models and a practical framework for when to trust their output distribution.

We investigate whether Mixture-of-Experts (MoE) language models develop linguistically structured expert routing during bilingual language acquisition. Inspired by the Declarative-Procedural framework, we analyze lexical, grammatical, and syntactic processing in a decoder-only English-German MoE Transformer trained under sequential language exposure. We construct a probe-based validation set and extract token-level routing distributions to quantify category-dependent specialisation using mutual information, routing entropy, and Jensen-Shannon distance. The curriculum-trained model exhibits a peak mutual information of 0.1148 at layer 5, indicating category-dependent differences in routing distributions across linguistic categories. Surprisingly, a no-curriculum baseline trained on mixed English-German data shows stronger aggregate specialisation, reaching a peak mutual information of 0.2599 at the same layer. These results suggest that interpretable linguistic organization emerges within MoE routing patterns even without sequential language exposure. A replication at a second training seed shows that the no-curriculum condition's specialisation concentrates on a single language whose identity is seed-dependent, whereas the curriculum consistently yields a stable, language-balanced routing profile; rather than uniformly increasing specialisation, staged bilingual exposure reduces single-language dominance. The official Github repository: https://github.com/Amrit828/DP-Theory-MOE-Interpretability-Research

Designing an ABR algorithm for one network scenario takes an engineer months, and large language models now do this work in hours, matching or beating hand-built designs. But either way, the design fits only the world visible at its birth, and fails on the world that arrives after. We ask whether an ABR algorithm can keep pace with the world, redesigned in minutes as each scenario arrives, with every change proven harmless to every scenario already served. In this work, we propose ReForge, a continual heuristic learning framework that adapts to continuously changing scenarios. ReForge runs that routine with a large language model (LLM) in the loop. Each round the LLM reads where the current design falls short and proposes one small edit, and a replay over every network served so far decides. Specifically, what it edits is a single page of fuzzy rules that routes every decision to one of a frozen pool of pre-trained policies. The LLM writes the first page from measurements alone, then keeps improving it on its own. Each round it reads where the current rules fall short and proposes one small edit, and a replay over every network served so far decides whether the edit lands. We evaluate ReForge on nine real-world network families arriving one at a time as 3G, 4G, then 5G. A few edits per arrival lift mean QoE from 1.23 to 1.74, past the best single policy at 1.66 and to 94% of an oracle, and even repair families the loop never saw, one rising from 0.30 to 0.80. All code, data, and experiment records will be open-sourced upon cleanup.

Multimodal large language models (MLLMs) exhibit substantial performance degradation in non-English visual reasoning, despite the strong multilingual competence of their text-only backbones. While mechanistic evidence from text-only models suggests that non-English inputs are routed through an English-centric latent space, the multimodal implications of this phenomenon remain unexplored. Through rigorous mechanistic analysis, we identify the Ghost Anchor phenomenon: a temporal modality asynchrony where linguistic translation to the English semantic manifold completes in early layers, while visual semanticization remains immature. Consequently, visual signals are physically present yet functionally invisible during the early alignment window. To rectify this, we propose ANCHOR, a training framework employing Proactive Visual Anchoring (PVA) to accelerate early visual semantic emergence, ensuring visual representations proactively guide linguistic translation. Mechanistic interventions confirm that ANCHOR successfully restores the causal influence of visual signals during early translation. Furthermore, extensive experiments on XMMMU, MaXM, and CVQA demonstrate that ANCHOR consistently outperforms standard baselines, achieving robust visual reasoning across both fine-tuned and zero-shot languages.

Test-time compute can substantially improve Large Language Model (LLM) reasoning performance, yet how and when additional compute helps remains poorly understood. We study Divergent-Convergent Reasoning (DCR), a simple two-phase primitive consisting of an exploration phase that generates multiple candidate solutions followed by a convergent reconciliation phase. We present three core results. First, we show that even a single reconciliation step can reliably amplify correct minority reports: across datasets, DCR often recovers the correct answer when correct exploration outputs are in the minority, a regime where majority voting fails. Second, we introduce recursive DCR, an autoregressive reconciliation system that iteratively analyzes disagreements and allocates additional test-time compute. Recursive DCR achieves higher accuracy than fixed-compute baselines-reaching 93.3% on AIME 2024 and 92.0% on AIME 2025-while using roughly 27% less compute on average, demonstrating that attentive resource allocation is superior to uniform scaling. Third, we analyze disagreement among exploration outputs via a simple, training-free dispersion metric. Dispersion reveals a structured relationship between disagreement and test-time gains: in regimes where DCR is effective, higher disagreement among exploration outputs is associated with larger accuracy improvements from reconciliation. Together, these results show that disagreement, often viewed as noise, can be systematically exploited to improve test-time reasoning and reveal emerging scaling laws for agentic LLM systems.

While Parameter-Efficient Fine-Tuning (PEFT) has substantially reduced the hardware cost of adapting Large Language Models (LLMs) by decreasing the number of trainable parameters, recent studies have sought to further improve PEFT through parameter sharing. However, these approaches either employ uniform parameter sharing across layers, which can delay convergence, or rely on dynamic masking strategies, which add computational overhead. The potential of sharing patterns inspired by the inherent hierarchical structure of Transformer architectures remains unexplored in PEFT. To address this gap, we introduce SAPE (Sandwich Adapters for Parameter Efficiency), a PEFT framework based on a sandwich-style hard weight-sharing topology. SAPE routes intermediate Transformer layers through balanced shared group adapters while strictly isolating the input embedding and final projection boundary transformations to prevent gradient interference. This design significantly reduces memory consumption while eliminating the computational overhead associated with dynamic parameter-sharing methods. Extensive evaluations across encoder-only and causal decoder architectures demonstrate that SAPE achieves state-of-the-art performance in low-parameter regimes. On natural language understanding, SAPE outperforms proPETL on RoBERTa-large while utilizing only 10% of the baseline's parameter budget. On natural language generation and world knowledge reasoning with LLaMA-3.2 (3B) under a strict ~0.6M parameter constraint, SAPE outperforms AdaLoRA, yielding absolute improvements of +4.85% on GSM8K and +3.11% on CommonsenseQA. Furthermore, through comprehensive topological ablations, we formalize an inherent capacity trade-off: while hard parameter sharing strongly regularizes semantic generalization, it slightly smooths the sharp layer-wise transformations required for rigid multi-step arithmetic reasoning.

We study score-matching diffusion models with a convolutional architecture. We argue that the inductive bias of locality means that the machinery of effective field theory from physics can be usefully applied to describe the denoising dynamics. We apply this formalism first to a simple toy example which permits an analytical description, and thereafter to MNIST, and show that in both cases, the mutual information between two points grows in a manner predicted by a simple effective field theory of Brownian motion.

Dynamic scene graphs (DSGs) capture spatio-temporal interactions across videos as $\langle$subject, predicate, object$\rangle$ triplets, and underpin downstream tasks such as video captioning, video question answering, and action analysis. However, end-to-end dynamic scene graph generation (DSGG) methods are closed-set: they recognize only objects and predicates from a fixed training vocabulary and struggle with the long-tailed distribution of rare concepts, severely limiting their real-world applicability. Existing open-vocabulary models typically inherit pretrained large language models, resulting in multi-stage training and inference with substantial cost. We introduce OvDSGG, the first end-to-end framework for open-vocabulary DSGG. OvDSGG builds on top of an open-vocabulary Spatial Backbone and a Temporal Backbone; we further propose a Triplet Feature Extraction Module that bridges them, and a Visual-Language Alignment Module that preserves open-vocabulary recognition by learning an adaptive decision boundary in the joint visual-language feature space, without expensive knowledge distillation in existing methods. We further introduce a rigorous open-vocabulary DSGG benchmark adapted from Action Genome, with disjoint Base/Novel splits for both objects and predicates. OvDSGG significantly outperforms open-vocabulary baselines across all metrics, with zero-shot Recall@$K$ scores 10.0--20.4 percentage point higher than the next-best baseline, while on closed-set DSGG remaining competitive with state-of-the-art models. Code and benchmark are publicly available at https://github.com/jhelsby/OvDSGG/.

Aligning a language agent to several objectives at once is a persistent failure mode of preference-based training: when objectives are combined additively, optimization collapses onto whichever is cheapest to improve and sacrifices the rest, so a support agent learns to sound warm while giving no real help. The root issue is that an additive reward has no notion of balance. We introduce Mint (MIN-selection preference disTillation), a one-line change to preference distillation: rather than ranking sampled candidates by a weighted sum of rewards, we rank them by their weakest objective, distilling the best-balanced candidate over the most lopsided one with an unchanged DPO objective. This is the p -> negative infinity limit of a generalized-mean family spanning additive to worst-case selection. Across cooperative emotional support and adversarial negotiation, min-selection lifts both objectives while sharply cutting their imbalance; on emotional support it raises the weaker axis from 0.37 to 0.64 (p < 10^-40), surpassing human experts and persisting across full multi-turn rollouts. A turn-by-turn analysis yields our central finding: min-selection corrects imbalance in proportion to how imbalanced the reference policy is, and its benefit endures over an interaction precisely as long as that imbalance does.

On-policy self-distillation (OPSD) gives language agents dense token-level supervision from a privileged self-teacher on the policy's own trajectories. Existing methods allocate this supervision mainly by teacher trust, but trust does not reveal whether emphasizing a token supports the current policy objective. We call this the trust-utility mismatch and introduce Influence Calibration for Self-Distillation (ICSD). For each supervised token, ICSD measures the first-order response of its importance-weighted RL surrogate contribution to a teacher-directed output perturbation. Batch-adaptive calibration converts this non-stationary signal into a bounded allocation weight while preserving the original auxiliary-loss mass within each action turn. These detached weights affect only the distillation loss and require no additional model pass. Across ALFWorld, WebShop, and Search-QA, ICSD improves all matched aggregate metrics over trust-only allocation under Group Relative Policy Optimization (GRPO) and Group-in-Group Policy Optimization (GiGPO), across two model families spanning 1.5B to 7B. At 7B, it reaches 96.1% ALFWorld success and a WebShop score of 93.1. Frozen-batch analyses show that ICSD reduces teacher-supported mass assigned to objective-opposed tokens from 60.1% to 37.8% and raises cosine compatibility with the RL gradient by 0.192. A companion repository is avail- able at https://github.com/lanqz7766/Influence-Calibration-for-On-Policy-Self-Distillation-in-Agentic-RL.

Large language models work well on English and behave in poorly understood ways on languages typologically far from it. Japanese is a clean example, where evaluation still leans on translation quality and JGLUE-style benchmarks, which roll lexical, syntactic and pragmatic competence into a single score. The phenomena on which general-purpose models fail Japanese users are pragmatic: honorifics, in-group and out-group reference, context-sensitive politeness, zero anaphora. I introduce J-PragEval-v0, a minimal-pair benchmark isolating four such phenomena from surface fluency, and combine it with linear probes and teacher-forced log-probability evaluation to ask where inside TinySwallow-1.5B (28 layers, hidden size 1536) the corresponding contrasts live. The four features split three ways. Honorific register sits cleanly in the residual stream: 0.96 balanced accuracy at layer 15, and the model flips its preferred continuation with the scenario on 93 percent of items. Implicit subject and in-group reference are not linearly decodable at the final prompt token (0.48 and 0.38), yet flip rates are 0.77 and 0.79, so the contrast is worked out during generation rather than stored at the prompt. Indirect refusal is the negative case: 0.95 probe accuracy collapsing to a 0.43 flip rate under length-normalised teacher forcing, because the current minimal pairs conflate politeness with continuation length. I also specify Pragmatic Representation Steering, a parameter-free inference-time method that edits residual-stream activations along the class-mean-difference directions probing identifies. Feasibility is argued indirectly rather than demonstrated: the contrastive activation addition baseline, the same geometry the method would inject, recovers probe accuracy within one to two points of logistic regression wherever a linear signal exists. Scaling to Llama-3.1-Swallow-8B is the next step.

Instruction-based video editing is commonly built on video-pretrained generative backbones: a video diffusion transformer is adapted, at considerable cost, to condition on a source video and an editing instruction. In this report we explore a different route and show that a strong instruction-based image editing model can edit videos by operating directly on video-VAE latents. Starting from Qwen-Image-Edit, we arrange the latent frames of a Wan~2.1 video VAE as tiles of one large virtual image, reuse the editor's image positional encoding for every tile, and bridge the two latent spaces with a pair of lightweight input/output projections warm-started from the editor's own patchify and unpatchify layers, so that at initialization a (static) video is embedded exactly as an image the model already understands. The whole system is then fine-tuned on the public Ditto-1M editing triplets, and a few denoising steps of Wan~2.2 serve as an optional temporal enhancer. We motivate the design with a chain of zero-training observations: the stock image editor already edits a video presented as a contact sheet; it is indifferent to whether the sheet's tokens come from one joint encode or from per-frame encodes stitched in latent space; and it even edits genuine video latents zero-shot to a clearly recognizable degree, leaving fine-tuning only a fidelity gap to close. Our results suggest that, despite the large investment in training video latent spaces, per-frame video latents remain close enough to the image domain that mature image editing priors transfer with minimal adaptation. Project Page: https://yunpeng1998.github.io/Qwen-Video-Edit-Page Code: https://github.com/yunpeng1998/Qwen-Video-Edit Model: https://huggingface.co/yunpeng1998/Qwen-Video-Edit

Multi-agent large language model (LLM) systems can improve reasoning by spending more computation, but deployment requires deciding when extra collaboration is worth its cost. We isolate this decision by running every problem under four protocols while holding the solver fixed within each setting: direct solving (Baseline), iterative self-correction (Single), planner-executor-reviewer collaboration (PER), and multi-agent deliberation (Broadcast). The primary benchmark comprises 4,181 competition-level math problems; paired robustness checks cover four benchmarks spanning competition math, biology, and broader science with two solver families. Across fixed policies, trained routers, and frozen LLM routers, conservative policies under-escalate, whereas higher-solve frozen routers often over-escalate. A post-answer, pre-collaboration gpt-oss-120b probe ranks Baseline failures with 0.8847 AUROC (4,151 parseable cases; 95% CI [0.8732, 0.8955]). The same score remains informative for predicting whether any collaboration helps (0.7683 AUPRC), but is much weaker for identifying PER- or Broadcast-specific value (0.1674 and 0.1041 AUPRC). Separately, the pre-answer self-confidence gate reaches 78.0% solve at 45K tokens, compared with 73.8% at 71.3K for a frozen gpt-oss-120b router and 92.4% for a retrospective fixed-order oracle. Across 10 paired model-condition settings, the oracle adds 23.2-58.3 points of retrospective coverage over Baseline, but protocol profiles vary by task. In the six settings with held-out router evaluations, oracle gaps remain 18.5-28.9 points. Confidence can therefore support initial escalation, while protocol-specific cost-aware routing remains unresolved.

Modern sequence models heavily rely on massive memory footprints and large-batch stochastic optimization, barriers that restrict sample efficiency and continual learning. We introduce the $p$-Spin Glass Network, a novel architecture that overcomes these limitations, structurally manages optimization variance and yields four noticeable capabilities: 1. It enforces memory efficiency: native ternary quantization compresses internal parameters by $8\times$, while exact implicit gradients strictly bound activation memory to $\mathcal{O}(B \cdot T \cdot D)$. 2. it demonstrates sample efficiency, matching the asymptotic performance of a Transformer baseline while utilizing $8\times$ fewer training sequences. 3. Method enables single-batch stability and smooth, monotonic convergence at a stochastic micro-batch size of $1$. 4. Finally, this stability proves modality-agnostic, maintaining robust temporal credit assignment across both discrete subword and long horizon uncompressed raw byte streams. Ultimately, this work removes large batch requirement for stable deep learning, establishing a foundation for continuous learning and edge AI.