Two-lens editorial ranking of AI research papers under consideration for upcoming AI Post Transformers episodes. https://podcast.do-not-panic.com en Sun, 01 Nov 2026 01:24:57 +0000 http://arxiv.org/abs/2610.38981v1 Speculative decoding (SD) accelerates autoregressive inference by leveraging the memory bound nature of decoding, verifying multiple draft tokens in a single forward pass. For mixture-of-experts (MoE) models, prior work predicts a concave speedup-batch size relationship, with gains peaking at intermediate batch sizes due to expert saturation. We evaluate this across 10 sparse MoE models (30B--1T parameters, 8--512 experts) and find recent fine-grained models break this trend, showing higher s... Public: 0.512 | Memory: 0.531 | Status: Cover now Why now: Fine-grained MoE models (DeepSeek/Kimi-style, up to 1T parameters) now dominate frontier releases, and this paper shows that the standard speculative-decoding speedup model for MoE is wrong for them. Anyone deploying SD on these models needs the corrected picture. Wed, 28 Oct 2026 22:04:53 +0000 2610.38981 http://arxiv.org/abs/2606.06302v2 Multi-turn LLM serving accumulates dialogue history whose Key-Value (KV) cache grows with every turn and every user, quickly exceeding the model weights themselves and making memory -- not compute -- the binding constraint on throughput. Non-uniform KV compression, which allocates heterogeneous budgets across attention heads, preserves accuracy far better than uniform schemes, yet remains impractical: modern serving stacks assume identical KV lengths across heads, so heterogeneity traps freed... Public: 0.510 | Memory: 0.648 | Status: Cover now Why now: Multi-turn serving is dominated by KV cache growth, and non-uniform head-wise compression has been stuck on serving-stack assumptions of uniform KV lengths. Tangram shows that offline-calibrated, static budgets (reservation, ragged paging, ahead-of-time load balancing) make it practical in vLLM, with up to 2.6x throughput and open code. Thu, 04 Jun 2026 15:41:27 +0000 2606.06302 http://arxiv.org/abs/2610.38355v1 KV cache reuse accelerates LLM serving, but existing systems manage host DRAM and NVMe SSD as a tiered hierarchy where DRAM acts as an intermediate cache. Production LLM workloads break both premises of this hierarchy. First, bimodal reuse defeats DRAM temporal filtering, since workloads are dominated either by one-shot blocks with long reuse distances or by multi-gigabyte agentic sessions that exceed DRAM capacity, putting SSD on the critical read path. Second, DRAM staging and eviction-orde... Public: 0.482 | Memory: 0.801 | Status: Cover now Why now: Agentic coding workloads (Claude Code traces) produce multi-gigabyte sessions that overflow DRAM, and FlatCake shows the DRAM-as-cache tier hierarchy breaks down for KV reuse. It replaces that hierarchy with flat placement, session-contiguous layout and GPUDirect Storage reads, which bears directly on how production KV caching gets built. Wed, 28 Oct 2026 06:54:51 +0000 2610.38355 http://arxiv.org/abs/2606.06256v4 As the input length of large language model (LLM) serving continues to grow, the KV cache has become a dominant bottleneck in AI infrastructure. It limits GPU memory capacity, serving concurrency, cache reuse, and distributed scalability. Multiple important problems, including position-independent KV cache, prefix KV cache compression, hot/cold KV cache separation, and distributed KV cache management, all depend on how the KV cache is represented and managed. However, existing serving systems... Public: 0.414 | Memory: 0.620 | Status: Cover now Why now: KV cache is the dominant bottleneck in long-context serving, and RedKnot unifies position-independent reuse, prefix compression, hot/cold tiering, and distributed placement through one head-level decomposition. This closely matches the show's memory/storage lens (paging, tiering, KV reuse) and the ongoing interest in CacheBlend-style PIC and sparse KV work. Thu, 04 Jun 2026 14:57:07 +0000 2606.06256 http://arxiv.org/abs/2606.06448v2 LLM agents are increasingly deployed on long-horizon tasks requiring sustained reasoning over extended interaction histories. Realizing this at scale requires agents to persistently store, retrieve, and update their own memory across sessions. A rich ecosystem of agent memory systems has emerged spanning flat retrieval, LLM-mediated extraction, consolidating fact stores, and agentic control flows. Yet, their system-level behavior remains uncharacterized. We present the first systems character... Public: 0.610 | Memory: 0.412 | Status: Cover now Why now: Agent memory systems (Mem0, A-Mem, MemGPT-style stores, and similar) are spreading quickly, and this is the first systems-level cost characterization of them. It splits cost across construction, retrieval, and generation, and gives concrete recommendations on construction scheduling, amortization, and fleet-scale management. Thu, 04 Jun 2026 17:44:18 +0000 2606.06448 http://arxiv.org/abs/2608.02691v4 The key-value (KV) cache has become a major memory and bandwidth bottleneck in long-context large language model inference, making ultra-low-bit quantization increasingly important. However, existing rotation-based INT2 methods optimize cache statistics or proxy errors before the complete attention readout, even though the model is ultimately affected by the error propagated through attention and the output projection $W_O$. To address this mismatch, we propose \textit{OptR}, an output-aware ... Public: 0.475 | Memory: 0.458 | Status: Monitor Why now: INT2 KV-cache quantization is a live frontier for long-context serving, and this work targets the actual attention-output error through W_O instead of proxy metrics. It also keeps the paged KV-cache format with negligible overhead, which makes it deployable. Mon, 03 Aug 2026 09:24:23 +0000 2608.02691 http://arxiv.org/abs/2606.05951v1 NVSHMEM is NVIDIA's OpenSHMEM-based PGAS communication library for GPU clusters, enabling GPU-initiated, one-sided communication through symmetric memory. Despite its growing adoption, a system-level understanding of its design and behavior remains scattered across documentation, source code, and application experience. This paper presents a concise study of NVSHMEM's programming model, implementation, and performance characteristics, focusing on symmetric memory, one-sided operations, and de... Public: 0.468 | Memory: 0.446 | Status: Monitor Why now: NVSHMEM and device-initiated communication underpin expert-parallel MoE stacks like DeepEP, which are central to current large-scale serving and training. A consolidated system-level explainer of symmetric memory and one-sided GPU communication is a useful reference as MoE and disaggregated serving spread. Thu, 04 Jun 2026 09:50:16 +0000 2606.05951 http://arxiv.org/abs/2610.02800v2 Speculative decoding accelerates autoregressive generation by using a lightweight draft to propose multiple tokens for parallel verification. However, existing methods often require an additional draft model or weight representation, introducing non-negligible memory overhead on resource-constrained devices. Self-speculative approaches reduce this overhead, yet still face trade-offs between draft quality, target quality, and storage efficiency. We propose BitNest, a bit-nested speculative dec... Public: 0.486 | Memory: 0.433 | Status: Monitor Why now: Speculative decoding on edge devices is a hot serving topic, and BitNest's nested-precision trick removes the separate draft model's memory cost by sharing one weight representation. Extending the same idea to the KV cache gives it a real memory-footprint angle. Fri, 02 Oct 2026 04:43:42 +0000 2610.02800 http://arxiv.org/abs/2608.01651v1 Hybrid-attention large language models combine full attention with recurrent linear attention to reduce long-context inference costs, yet their autoregressive decoding remains memory-bound. Tree speculative decoding offers an attractive acceleration path, but existing tree-speculation systems are designed around the key--value caches of full-attention models. On hybrid models, they traverse recurrent layers branch by branch and materialize a full state for every proposal node, causing verific... Public: 0.456 | Memory: 0.558 | Status: Monitor Why now: Hybrid-attention models (full attention plus linear attention/recurrent state) are increasingly used in production, and tree speculative decoding is a standard decode accelerator. Bole addresses the gap between the two, with an SGLang integration, a 4.72x decode speedup and an 82-99x cut in transient state memory. Mon, 03 Aug 2026 03:43:14 +0000 2608.01651 http://arxiv.org/abs/2606.05538v2 Mixture-of-Experts (MoE) models achieve strong performance through conditional computation, but their large parameter footprint poses deployment challenges. Prior MoE compression approaches catastrophically fail when evaluated on general-purpose benchmarks beyond commonsense reasoning. We trace this failure to the granularity of compression: important capabilities are distributed across experts but concentrated in FFN sparse intermediate dimensions. To identify these dimensions, we use Fisher... Public: 0.551 | Memory: 0.413 | Status: Monitor Why now: MoE models are the dominant open-weight architecture, and deployment footprint is a pressing practical problem. The paper shows that prior expert-level pruning fails on general benchmarks, and that capability concentrates in a tiny set of FFN intermediate dimensions (12 of 1.35M collapse GSM8K). That finding is surprising and would make a good teaching point. Thu, 04 Jun 2026 00:43:20 +0000 2606.05538 http://arxiv.org/abs/2610.39302v1 Whether models trained on different modalities converge to the same representation of the world is central to AI. The Platonic Representation Hypothesis (PRH) conjectures that they do as they scale, but without a proof it offers neither a guarantee nor a recipe. We turn the hypothesis into a theorem. First, we prove a \emph{Platonic Representation Theorem}: Joint-Embedding Predictive Architectures (JEPAs) trained separately on any number of modalities learn the same representation up to one r... Public: 0.799 | Memory: 0.152 | Status: Cover now Why now: LeCun and Balestriero turn the widely discussed Platonic Representation Hypothesis into a theorem. They also show that a 109M-parameter text JEPA aligns with vision encoders better than language models up to 72B. This directly challenges the 'scale is all you need' narrative and ties into the JEPA vs. next-token prediction debate. Thu, 29 Oct 2026 03:49:53 +0000 2610.39302 http://arxiv.org/abs/2610.02200v1 We show that multimodal models possess strong reasoning abilities and that an appropriate harness can unlock their potential to solve tasks across diverse interactive environments. We introduce VISTA, a visual harness that gives a general-purpose multimodal model long-horizon vision. VISTA allows the model to directly perceive the environment through visual observations and maintains a lossless visual memory that preserves past observations in their original form. The model can actively retri... Public: 0.774 | Memory: 0.266 | Status: Monitor Why now: Agent harnesses are a hot topic, and a Kaiming He-authored paper claiming a perfect Relative Human Action Efficiency score on ARC-AGI-3 will draw broad attention in the next 30-90 days. The lossless visual memory with active retrieval is a concrete design that fits the current debate about agent memory and context management. Thu, 01 Oct 2026 17:59:45 +0000 2610.02200 http://arxiv.org/abs/2607.04011v1 While decoders have rapidly scaled, encoders have remained largely unchanged since BERT. We revisit this disparity by frozen backbone evaluation via probing. Under this lens, the representations of BERT encoders become increasingly $\textit{unexploitable}$ by frozen probes, despite improved perplexity. The misalignment originates in BERT's flat design, which couples representation learning to the token reconstruction loss. We propose $\textbf{CrossBERT}$, a two-part architecture that separate... Public: 0.699 | Memory: 0.144 | Status: Monitor Why now: Encoders have barely changed since BERT while decoders scaled rapidly. Yann LeCun's co-authorship and the frozen-probe finding that BERT representations get worse with scale should draw broad attention, and the paper also bears on embedding and retrieval models. Sat, 04 Jul 2026 20:13:38 +0000 2607.04011 http://arxiv.org/abs/2606.05806v1 Existing benchmarks evaluate Tool-Integrated Reasoning (TIR) in LLMs on idealized ''happy paths'', largely overlooking real-world tool failures. We introduce ToolMaze, a benchmark for dynamic path discovery and error recovery in TIR agents. To separate systematic replanning from blind trial-and-error, ToolMaze adopts a two-dimensional design: DAG-based topological complexity and a $2 \times 2$ taxonomy of tool perturbations (explicit/implicit, transient/permanent). Evaluations show that pertu... Public: 0.655 | Memory: 0.174 | Status: Monitor Why now: Agent reliability under real-world tool failures is a hot topic as tool-using agents move into production. The finding that implicit semantic failures cause about a 37% drop in recovery, and that fault tolerance scales 3.66x slower than basic task execution, is a quotable result for a broad audience. Thu, 04 Jun 2026 07:38:46 +0000 2606.05806 http://arxiv.org/abs/2606.05661v1 Continual learning, the ability of AI systems to improve through sequential experience, has attracted substantial interest, but no high-quality benchmark exists to evaluate it. We introduce Continual Learning Bench (CL-Bench), the first difficult, expert-validated benchmark designed to measure whether LLM-based systems genuinely improve with experience. CL-Bench spans six diverse domains (software engineering, signal processing, disease outbreak forecasting, database querying, strategic game-... Public: 0.645 | Memory: 0.250 | Status: Monitor Why now: Continual learning and agent memory are hot topics, and this benchmark finds that naive in-context learning beats dedicated memory systems. That result challenges the assumptions behind the current wave of agent-memory products. Thu, 04 Jun 2026 03:43:28 +0000 2606.05661 http://arxiv.org/abs/2609.01437v1 As agents move from research prototypes to deployed tools, their capability increasingly depends on model-external execution infrastructure, commonly termed the agent harness. Changing this harness while holding model weights fixed can substantially alter task performance. Current agent evaluations typically report downstream performance under a chosen harness, leaving a model's ability to develop the harness itself comparatively underexplored. We introduce HarnessDev, a benchmark that shifts... Public: 0.623 | Memory: 0.214 | Status: Monitor Why now: Agent harnesses are a hot topic as deployed agents depend on external infrastructure. HarnessDev asks whether LLMs can build and improve that infrastructure themselves, which fits the self-improving-agent debate. Tue, 01 Sep 2026 15:45:33 +0000 2609.01437 http://arxiv.org/abs/2610.38342v1 Financial agents must form forecasts and investment judgments from decision-time evidence, with predictions evaluated against subsequent outcomes. Existing benchmarks assess financial knowledge, research workflows, or forecasting, but evaluating the full research process requires a strictly frozen information set and professional evaluation criteria. We introduce Investor's Last Exam (ILE), an equity-research benchmark led by experienced buy-side investors and fund managers. Its three compone... Public: 0.634 | Memory: 0.109 | Status: Monitor Why now: Finance is a high-visibility agent use case, and ILE tackles look-ahead leakage with a timestamp-frozen, versioned corpus (3.43M document versions) and expert-built rubrics. Its headline finding is that frontier models do well on research tasks but poorly on forecasts (52.5 for the top model on Forecast & Model), which is a useful lesson for evaluating agents. Wed, 28 Oct 2026 06:39:24 +0000 2610.38342 http://arxiv.org/abs/2608.01918v1 Harness design plays a critical role in agent performance by shaping how large language models (LLMs) perceive, reason over, and act within executable environments. Recent work has proposed automatic harness evolution, which iteratively improves the harness from agent--environment interactions. However, existing methods often overfit to the evolution tasks, rely exclusively on trajectory-derived signals, and optimize harness components jointly, causing interference across components. We propo... Public: 0.599 | Memory: 0.147 | Status: Monitor Why now: Agent harness design is a hot topic. This paper shows a 54% to 66% Pass@1 gain on SWE-bench Verified in 5 iterations, and the evolved harness transfers to other models and held-out tasks. That speaks directly to the question of how much of agent performance comes from the harness rather than the model. Mon, 03 Aug 2026 08:51:50 +0000 2608.01918 http://arxiv.org/abs/2609.01343v3 Looped Transformers increase effective depth by iterating a shared block of layers, but most evaluations compare at fixed model size, conflating architectural advantage with extra FLOPs. We study looping on Mixture-of-Experts Transformers while closely matching per-token FLOPs, total non-embedding parameters, and KV cache. Through a series of ablations, we arrive at a recipe we call SMELT (Sparse MoE Transformer, middle layers Loop Twice), which loops the middle half of layers twice while mat... Public: 0.613 | Memory: 0.285 | Status: Monitor Why now: Looped and recurrent-depth Transformers are getting renewed attention, and this is a rare compute-, parameter- and KV-cache-matched comparison on MoE models up to 54B with fitted scaling laws. It reports 6.8-18.0% training FLOP savings on the compute-optimal frontier. Tue, 01 Sep 2026 14:52:57 +0000 2609.01343 http://arxiv.org/abs/2607.03691v2 Coding agents, autonomous systems that use large language models (LLMs) to resolve software engineering tasks, rely on agent harness: a middleware layer in between a developer and a large language model that orchestrates system prompts, tool execution, context management, and iterative reasoning loops. While these agent harnesses evolve at extreme velocities, no study has examined how this evolution affects agent quality (i.e., effectiveness and efficiency) over time. Practitioners regularly ... Public: 0.621 | Memory: 0.193 | Status: Monitor Why now: Coding agents are in heavy daily use, and practitioners blame the model when quality drops after an update. This paper offers controlled evidence that the harness is a major, under-examined cause. The point is timely and easy to explain to a broad audience. Sat, 04 Jul 2026 03:55:25 +0000 2607.03691 http://arxiv.org/abs/2608.02352v1 Native computer use offers a general interface for agents to operate almost any software available to people, but requires long-horizon state tracking, large-scale interactive experience, and learning from sparse yet verifiable outcomes. We introduce Qwen-CUA, a native computer-use agent with a 397B-A17B Qwen mixture-of-experts backbone. It observes only screenshots and acts through keyboard and mouse events, without DOM trees, accessibility metadata, or task-specific APIs. Its scaffold maint... Public: 0.637 | Memory: 0.251 | Status: Cover now Why now: Qwen-CUA reports 86.2 on OSWorld-Verified and 87.6 for the 1T+ Qwen-CUA-Max, which puts open-weight native computer use near the leading proprietary systems. The paper also describes the RL infrastructure behind it: about 100k vCPUs, tens of thousands of concurrent environments, and roughly 40k verifiable tasks. Computer-use agents are a hot area, so a broad audience is likely to care within 30-90 days. Mon, 03 Aug 2026 15:04:20 +0000 2608.02352 http://arxiv.org/abs/2610.38681v1 Self-distillation offers a way to improve reasoning models when a suitable external teacher is unavailable. However, common approaches rely on privileged information to construct a stronger self-teacher, limiting their use when only unlabeled prompts are available. In this paper, we introduce Weakening-Induced On-Policy Self-Distillation (WISD), which instead constructs a weaker counterpart of the original model and uses the contrast between the two as a training signal without requiring priv... Public: 0.724 | Memory: 0.161 | Status: Cover now Why now: Label-free self-improvement of reasoning models is a hot topic, and WISD gets large gains (e.g. +12.8 points on HMMT 2025 for Qwen3.5-4B) without privileged information or an external teacher. The counterintuitive idea of weakening a model through quantization or pruning to make it stronger fits an engaging episode, and Song Han's group authored it. Wed, 28 Oct 2026 18:55:38 +0000 2610.38681 http://arxiv.org/abs/2610.40000v1 Smartphone agents increasingly work as teams, in which a large planner directs parallel workers across long sessions. Running such a team on the device requires 200B-class models and contexts that exceed phone DRAM many times over, yet existing tiers treat flash as storage: UFS starves per-token weight streaming, and mobile High-Bandwidth Flash (HBF) needs a redesigned SoC and cannot absorb continuous KV writes. We show that a non-volatile memory can join an unmodified LPDDR6 bus as a plain r... Public: 0.400 | Memory: 0.617 | Status: Monitor Why now: On-device multi-agent LLMs with 200B-class weights and 256K+ token contexts push past phone DRAM. This paper proposes a concrete non-volatile main-memory tier (3D FeNOR) on an unmodified LPDDR6 bus. It sits directly in the memory-tiering and KV-placement space and competes with HBF. Thu, 29 Oct 2026 13:19:45 +0000 2610.40000 http://arxiv.org/abs/2609.02027v1 The major workloads in modern large language model (LLM) serving systems have shifted from single-shot LLM calls to multi-turn conversations, where new responses are generated based on the whole conversation history across all previous turns. The hit ratio, i.e., the average fraction of KV caches accessed directly from existing caches stored in high-bandwidth memory (HBM), is hence a crucial metric that governs system performance. Estimating the hit ratio is a highly nontrivial task due to th... Public: 0.409 | Memory: 0.588 | Status: Monitor Why now: Multi-turn conversations and agentic workloads now dominate LLM serving, and KV cache hit ratio under finite HBM is the key quantity governing prefix-cache capacity planning. This paper offers a closed-form, mean-field hit-ratio estimator for LRU eviction, validated on Qwen3-8B on Ascend NPUs. Wed, 02 Sep 2026 02:57:16 +0000 2609.02027 http://arxiv.org/abs/2607.04031v1 Large Language Model (LLM) inference is bottlenecked by the capacity and bandwidth of GPU High-Bandwidth Memory (HBM). Recent proposals, such as High-Bandwidth Flash (HBF) and RoMe, offer higher capacity or bandwidth than HBM, but require a minimum access granularity of kilobytes. We show that these Large-Granularity Memory Systems (LGMS) can degrade the performance of tiled matrix-multiplication, which is the dominant operation in LLM inference, by up to an order of magnitude. The root cause... Public: 0.334 | Memory: 0.594 | Status: Monitor Why now: High-Bandwidth Flash (HBF) and RoMe-style large-granularity memories are an emerging answer to the HBM capacity wall in LLM inference. This paper shows that tiled GEMM read amplification can cause up to 10x slowdowns on them, and that a tile-major layout plus a prefetcher brings performance to within 1% of HBM. Sat, 04 Jul 2026 21:19:49 +0000 2607.04031 http://arxiv.org/abs/2610.02652v1 SSD-resident approximate nearest-neighbor search is essential when vector collections exceed DRAM capacity. The challenge is to reduce SSD reads and hide I/O latency through concurrent reads and overlap with computation. However, for graph-based search, progressive candidate discovery limits advance I/O planning while IVF search reads inverted lists in full, incurring unnecessary SSD reads. In this paper, we present RaBitQ-SSD, an extension of IVF-RaBitQ for SSD-resident vector search. Specif... Public: 0.370 | Memory: 0.546 | Status: Monitor Why now: SSD-resident vector search is a core building block for RAG and retrieval at billion-scale, and the paper offers a concrete storage-aware design: partial codes held in DRAM, finer-grained SSD reads, and pipelined I/O. Its results go up to 10B vectors on a single SSD machine. Fri, 02 Oct 2026 01:20:20 +0000 2610.02652 http://arxiv.org/abs/2610.39216v1 Local deployment of MoE-based LLMs on CPU-GPU nodes is attractive for cost and privacy reasons, yet limited GPU memory forces most experts into CPU memory. Prefill is then compute-bound and benefits from aggressive GPU loading, while decode is I/O-bound and requires conservative PCIe transfers. These demands cannot be amortized under low batching---typical of local assistants and document tools---forcing phase-agnostic schedulers into a seesaw: favoring prefill's aggressive loading injects un... Public: 0.372 | Memory: 0.585 | Status: Monitor Why now: Local MoE inference on CPU-GPU nodes with experts offloaded to CPU memory is a growing deployment pattern. PDScope's phase-aware expert prefetching and placement over PCIe addresses it directly and gives a clear framing of the prefill/decode I/O tension. Thu, 29 Oct 2026 02:37:09 +0000 2610.39216 http://arxiv.org/abs/2610.02732v1 Evaluating distributed LLM serving designs on physical clusters is costly. Yet existing simulators provide only subsets of the capabilities needed for realistic design exploration: stateful closed-loop execution, timing prediction without profiling target hardware, and direct execution of unmodified serving benchmarks. We present ServeTwin, a closed-loop simulator that couples specification-driven analytical timing with a stateful serving loop. This coupling captures feedback among request co... Public: 0.387 | Memory: 0.529 | Status: Monitor Why now: Distributed LLM serving, prefill-decode disaggregation, and multi-turn agent workloads are central to current infrastructure planning. ServeTwin offers a vLLM-compatible simulator that tracks KV-cache evolution, and it was validated against InferenceX (3.6% mean error) and LMBenchmark (9.9% error). Fri, 02 Oct 2026 03:06:15 +0000 2610.02732 http://arxiv.org/abs/2610.39999v1 Autoscaling is a key mechanism in model-as-a-service systems, where fast scale-up is essential because the scale-up time directly determines how long a model service remains overloaded, which in turn governs serving quality. The state-of-the-art approach, model sleeping, adopted by mainstream serving engines including vLLM and SGLang, achieves near-zero-cost scale-up by pausing pre-spawned engines on the GPU. However, sleeping is not a free lunch: each sleeping model must keep its GPU context... Public: 0.389 | Memory: 0.511 | Status: Monitor Why now: Autoscaling and multi-model serving with vLLM/SGLang sleep mode is a live production concern. This paper identifies a concrete, previously overlooked GPU memory cost: each sleeping model keeps a resident CUDA context. It removes that cost by deduplicating the context across models. Thu, 29 Oct 2026 13:19:00 +0000 2610.39999 http://arxiv.org/abs/2610.02522v1 Within the shared industry vision of AI-RAN, AI-and-RAN seeks to co-locate virtualized radio access network (vRAN) workloads and AI services on shared GPUs. This sharing is inherently asymmetric: vRAN workload is latency-critical, whereas the machine learning (ML) workload is a throughput-oriented, best-effort co-tenant. We present Beaver, a GPU sharing system that jointly manages compute and memory resources while protecting the vRAN's strict processing deadline. Beaver sizes the vRAN's stre... Public: 0.394 | Memory: 0.497 | Status: Monitor Why now: AI-RAN is an emerging industry push (NVIDIA Aerial) to co-locate ML serving with latency-critical 5G workloads on shared GPUs. Beaver shows concrete SM partitioning and HBM-bandwidth yielding via compiled-kernel rewriting, with real H200 results (74% of Llama-3.3-70B throughput at p99.9 within the 1.5ms deadline). Thu, 01 Oct 2026 21:52:30 +0000 2610.02522 http://arxiv.org/abs/2610.03935v2 General decision models, such as Jev, have recently emerged as efficient alternatives to LLMs for structured judgment and selection. But what kinds of decisions can these models reliably make, and how does their behavior change when individual decisions are composed into larger systems? To study this, we introduce JEVal, a bilingual benchmark comprising 11,257 instances from 36 datasets across 10 application domains, and evaluate 25 model configurations spanning general decision models and ge... Public: 0.442 | Memory: 0.155 | Status: Monitor Why now: Fast single-pass decision models are being pitched as cheap alternatives to LLMs for judgment and selection. This benchmark shows where they work (evidence-resolvable decisions) and where they fail: overconfident probabilities and errors that compound in long agent trajectories on τ-bench. Fri, 02 Oct 2026 18:47:00 +0000 2610.03935 http://arxiv.org/abs/2609.02786v1 The performance of LLM-based agents is jointly shaped by the base model and the harness used when interacting with the environment. This exposes them to safety risks in both harmful final responses and multi-step execution trajectories. Existing safety alignment mechanisms often rely on either external harness updates or policy optimization, yet applying either paradigm in isolation fails to bridge runtime control with intrinsic safety. We propose SafeEvolve, an experience-driven self-evolvin... Public: 0.562 | Memory: 0.136 | Status: Monitor Why now: Agent safety is a hot topic, and the paper combines runtime harness updates (safety prompts and skills) with SFT+RL policy training. That combination fits how agents are actually built today. The headline result is a 3x ASR reduction on AgentDojo with a small utility gain. Wed, 02 Sep 2026 16:19:54 +0000 2609.02786 http://arxiv.org/abs/2607.03702v1 Large language model (LLM) agents have shown strong decision-making capabilities in long-horizon interactive tasks, yet they still struggle to effectively leverage failed trajectories: full retries incur high interaction costs, while experience retrieval tends to dilute critical experience signals. To address this, we propose PivoARL, a self-feedback retry framework for experience exploitation in LLM agents. PivoARL identifies the pivotal erroneous turn through structured reflection and perfo... Public: 0.448 | Memory: 0.132 | Status: Monitor Why now: LLM agents are a hot area, and PivoARL tackles how to reuse failed trajectories. It retries locally from the pivotal erroneous turn, which cuts interaction cost by about 42% against full retries. The credit-assignment idea (reward the correct prefix, isolate the bad suffix) could transfer to other agentic RL training. Sat, 04 Jul 2026 04:45:05 +0000 2607.03702 http://arxiv.org/abs/2608.02139v2 Large language models (LLMs) capable of self-improvement require not only effective policy optimization, but also a principled mechanism for transforming transient interaction experience into persistent model capabilities. Existing self-improvement paradigms remain fragmented: test-time methods can explicitly extract experience but cannot internalize it into model parameters, whereas training-time optimization methods can update model parameters but lack an explicit mechanism for accumulating... Public: 0.532 | Memory: 0.166 | Status: Monitor Why now: Self-improvement, on-policy self-distillation, and experience-based learning are active themes. SPEE offers a clear pipeline that turns test-time experience into parameter updates before RL. Mon, 03 Aug 2026 12:27:32 +0000 2608.02139 http://arxiv.org/abs/2609.02783v1 Evaluating LLM agents is essential for guiding their development, yet it has grown prohibitively expensive: a single pass of a frontier model over an agentic benchmark can cost hundreds to thousands of dollars, a price paid repeatedly across iterative development cycles. Prior efforts, centered on benchmark distillation, reduce the number of evaluation tasks but leave the cost of executing each retained task untouched. In this work, we introduce early outcome prediction, a complementary axis ... Public: 0.610 | Memory: 0.140 | Status: Monitor Why now: Agent benchmark costs (hundreds to thousands of dollars per pass) are a real pain point for anyone iterating on agents. Early stopping inside each task is a practical complement to benchmark distillation, and the paper reports 13-26% fewer steps and up to 44% fewer input tokens. Wed, 02 Sep 2026 16:15:18 +0000 2609.02783 http://arxiv.org/abs/2606.05548v1 The rapid proliferation of Agent Development Kits (ADKs), SDK-level frameworks for building LLM-powered autonomous agents, has outpaced any empirical understanding of how framework choice affects agent performance. We propose \textbf{LLM-as-a-Developer}, a methodology that replaces human developers with an LLM coding agent that learns each framework's API from documentation, writes agent code, and iteratively repairs it through a validate-and-feedback loop until tests pass. By holding the dev... Public: 0.557 | Memory: 0.114 | Status: Monitor Why now: Agent frameworks are proliferating rapidly and developers are struggling to choose among them. A controlled comparison across 51 ADKs, using an LLM coding agent as the constant developer, speaks directly to that choice. The finding that documentation, source code and parametric knowledge are largely substitutable is a useful result for anyone building agent tooling. Thu, 04 Jun 2026 01:00:54 +0000 2606.05548 http://arxiv.org/abs/2608.01589v1 On-policy self-distillation (OPSD) improves reasoning by using a privileged view of a model conditioned on reference solutions to supervise a student view that observes only the question. However, the teacher-provided token-level targets may depend on reference-specific information unavailable at inference time. We propose Problem-Space-Guided OPSD (PS-OPSD), which replaces the complete solution with trajectory-grounded guidance describing the initial state, goal conditions, constraints, and ... Public: 0.498 | Memory: 0.113 | Status: Monitor Why now: On-policy self-distillation with privileged information is an active post-training topic. This paper offers a concrete design lesson: give the teacher structured problem-solving guidance instead of full reference solutions. Mon, 03 Aug 2026 01:47:25 +0000 2608.01589 http://arxiv.org/abs/2608.01837v1 Large language model agents have shown strong potential in complex interactive tasks, yet their reinforcement learning (RL) is often hindered by sparse rewards, as a long multi-turn trajectory may receive only a single outcome-level signal. On-policy self-distillation (OPSD) provides dense token-level supervision from a privileged teacher, but the teacher may not be reliable at every position. Existing methods commonly rely on isolated token-level discrepancies, which can be sensitive to nois... Public: 0.475 | Memory: 0.150 | Status: Monitor Why now: Dense supervision for sparse-reward agentic RL is a live topic, and on-policy self-distillation with a privileged teacher is gaining traction. PCSD gives a concrete recipe for deciding where to trust the teacher, with large ALFWorld gains over GRPO. Mon, 03 Aug 2026 07:47:59 +0000 2608.01837 http://arxiv.org/abs/2606.06090v2 LLM-based agents increasingly tackle long-horizon tasks with interdependent decisions, where each action reshapes future constraints and intermediate errors can cascade. Existing RAG and agent memory systems organize histories by semantic similarity, retrieving content-relevant entries at decision time. We argue that this design mismatches execution-state dependencies: it fragments decision trajectories and mixes valid and erroneous traces, hindering coherent state reconstruction and error is... Public: 0.606 | Memory: 0.254 | Status: Monitor Why now: Long-horizon agents are a hot topic, and context bloat and error cascades are a pain point for anyone building them. MAGE reframes agent memory as execution-state management: a hierarchical state tree with branch revision. It reports a 55% token reduction alongside higher task success. Thu, 04 Jun 2026 12:26:42 +0000 2606.06090 http://arxiv.org/abs/2606.05152v2 Reasoning models have advanced rapidly, but the dominant reinforcement learning from verifiable rewards (RLVR) recipe remains surprisingly narrow: sample many responses and reward each with a single bit indicating whether the final answer is correct. Yet many settings provide rich feedback, including execution traces, tool outputs, expert corrections, and model self-evaluations. We study how to use such feedback through a distributional variant of the classic imitation learning algorithm DAgg... Public: 0.530 | Memory: 0.113 | Status: Monitor Why now: RLVR is the dominant reasoning-model post-training recipe, and this paper offers a principled alternative that uses rich feedback such as execution traces and expert corrections. It also shows that the popular reverse-KL and JSD self-distillation objectives can fail to improve the policy monotonically. Wed, 03 Jun 2026 17:54:04 +0000 2606.05152 http://arxiv.org/abs/2609.02108v1 Diffusion language models (DLMs) have emerged as a promising alternative to the auto-regressive paradigm. With bidirectional attention and any-order generation, DLMs naturally fit infilling tasks, which require generating a middle span conditioned on both the prefix and the suffix. However, infilling is sensitive to the length of the span, while DLMs require the length to be fixed before generation. Although prior studies extend DLMs to dynamic lengths, they still suffer from two limitations.... Public: 0.479 | Memory: 0.162 | Status: Monitor Why now: Diffusion LMs are getting growing attention as an alternative to autoregressive models. Adaptive-length infilling is a real practical limitation for them, and PILL addresses it with 1.82x lower cost than the strongest baseline. Wed, 02 Sep 2026 04:52:01 +0000 2609.02108 http://arxiv.org/abs/2608.01847v1 Large language models (LLMs) achieve remarkable performance but are expensive to deploy due to their enormous size. FP4 quantization, with formats such as MXFP4 and NVFP4, offers an appealing solution with native hardware support on modern accelerators. However, maintaining accuracy under FP4 precision remains difficult. A key bottleneck lies in scale optimization: existing methods tightly couple the quantization and dequantization scales, forcing both to conform to the discrete low-precision... Public: 0.497 | Memory: 0.341 | Status: Monitor Why now: MXFP4 and NVFP4 now have native support on current accelerators, so accuracy-preserving FP4 post-training quantization is a live deployment question. FOCUS adds no inference overhead, because the quantization scale is relaxed at calibration time and never stored. Mon, 03 Aug 2026 07:54:56 +0000 2608.01847 http://arxiv.org/abs/2606.06547v2 Diffusion Large Language Models (dLLMs) refine tokens iteratively but commit them irreversibly, leading to a "stability lag" where early decisions remain fragile even after being written. We reveal that Post-Training Quantization (PTQ) error easily flips these borderline decisions at the write frontier, which are then permanently locked in and amplified. To address this, we propose Frontier-Aware Instability-Reweighted Calibration (FAIR-Calib), a two-stage PTQ framework for dLLMs. Stage I pro... Public: 0.439 | Memory: 0.303 | Status: Monitor Why now: Diffusion LLMs (LLaDA, Dream) are a rising alternative to autoregressive decoding, and W4A4 quantization is a prerequisite for making them practical to deploy. The paper identifies a failure mode specific to dLLMs: quantization error flips borderline tokens at the write frontier, and the irreversible commit locks those flips in. Thu, 04 Jun 2026 08:00:51 +0000 2606.06547 http://arxiv.org/abs/2607.03236v1 Diffusion language models (DLLMs) generate text by iteratively denoising masked positions, exposing a trajectory of predictive distributions rather than a single instantaneous belief. Most existing decoders ignore this trajectory and commit tokens from the current snapshot alone, conflating confidence with commitment readiness: a transient top-1 peak under incomplete context can be locked in, while candidates with consistent cross-step support are delayed. We propose Trajectory-Aware Commit G... Public: 0.463 | Memory: 0.166 | Status: Monitor Why now: Diffusion LLMs are an emerging alternative to autoregressive decoding, and decoding efficiency (fewer steps, higher tokens per forward) is a live topic. TACG is training-free and tested on LLaDA, Dream and LLaDA2-Mini. Fri, 03 Jul 2026 11:45:21 +0000 2607.03236 http://arxiv.org/abs/2610.02076v3 Jev-style decision models have gained surging traction by making decision-making dramatically faster and more efficient, returning categorical probability distributions over predefined options without generating free-form text. However, because Jev's architecture and training methodologies remain proprietary, how to replicate these capabilities effectively remains an open question. In this work, we investigate the extent to which general-purpose LLMs already possess this capability out of the... Public: 0.454 | Memory: 0.168 | Status: Monitor Why now: Fast single-pass categorical decision models (routing, classification, moderation) are an active practical trend. This paper shows that off-the-shelf LLMs already work as such models, and that a training-free readout may be enough. Thu, 01 Oct 2026 17:15:24 +0000 2610.02076 http://arxiv.org/abs/2606.05684v1 A central challenge for language agents is utilizing past experience to adapt to dynamic test-time conditions. While recent work demonstrates the promise of agentic memory mechanisms, most systems restrict retrieval to episode initiation. Consequently, agents are forced to rely on static guidance that becomes increasingly misaligned as long-horizon tasks unfold. To address this rigidity, we propose the Adaptive Memory Agent (AdaMEM), a novel framework for agent test-time adaptation. Without u... Public: 0.545 | Memory: 0.176 | Status: Monitor Why now: Agent memory is a hot topic, and AdaMEM shows that retrieving and synthesizing strategy mid-episode, not just at the start, improves long-horizon agents. It treats this as a new test-time compute scaling axis, which suits current interest in agentic self-improvement. Thu, 04 Jun 2026 04:06:08 +0000 2606.05684 http://arxiv.org/abs/2609.01457v1 Large language model (LLM) fine-tuning is increasingly shifting toward data generated on edge devices, where memory, computation, bandwidth, and connectivity constraints make conventional federated learning difficult to sustain. Split federated fine-tuning (SFT) improves client-side efficiency by offloading most model parameters and computation to the server but requires step-by-step bidirectional communication loop across the split interface and forces continuous client involvement throughou... Public: 0.424 | Memory: 0.352 | Status: Monitor Why now: On-device and edge fine-tuning of LLMs is growing, and this paper removes the per-step bidirectional communication in split federated fine-tuning by exploiting weight tying. It is tested on a real testbed with smartphones and Jetson boards. Tue, 01 Sep 2026 16:00:29 +0000 2609.01457 http://arxiv.org/abs/2606.06087v3 Agent systems increasingly use textual skills to encode reusable task procedures, but injecting these skills into the prompt at every step incurs substantial context overhead and exposes skill content as plaintext. We present LatentSkill, a framework that converts textual skills into plug-and-play LoRA adapters through a pretrained hypernetwork. LatentSkill stores skill knowledge in weight space rather than context space, removing per-step skill tokens while preserving modular loading, scalin... Public: 0.562 | Memory: 0.313 | Status: Monitor Why now: Agent skills (SKILL.md-style procedures) are a hot topic, and LatentSkill offers an alternative to paying prompt tokens for them: a hypernetwork turns skill text into LoRA adapters. It reports about 64-72% fewer prefill tokens with better task accuracy. Thu, 04 Jun 2026 12:26:09 +0000 2606.06087 http://arxiv.org/abs/2610.39264v1 While frontier vision-language models demonstrate strong capabilities in user interface (UI) navigation, their high inference costs prohibit everyday use. Conversely, smaller models trained via supervised fine-tuning (SFT) or offline reinforcement learning often memorize static trajectories, leading to brittle execution and an inability to recover from errors. To address these limitations, we propose a novel online reinforcement learning framework that trains UI navigation agents directly on ... Public: 0.566 | Memory: 0.117 | Status: Monitor Why now: GUI and UI-navigation agents are a hot area. This paper shows an 8B model trained with online dense RL matching Gemini-3.1-Pro on real iOS apps, which speaks to the cost of using frontier models for everyday agents. The infrastructure for running hundreds of concurrent virtual iOS devices, and the GiGPPO critic-stabilization idea, are reusable by others. Thu, 29 Oct 2026 03:17:33 +0000 2610.39264 http://arxiv.org/abs/2608.01414v1 With the rapid release of open-weight large foundation models, safety threats are shifting from black-box jailbreaks to neuron-level white-box attacks that directly identify and manipulate safety-related neurons. Existing alignment methods often investigate the safety behavior on a small number of neurons, creating fragile single point of failure with limited redundancy. To address this issue, we propose distributed safety alignment (DSA), which redundantly encodes safety capabilities across ... Public: 0.528 | Memory: 0.125 | Status: Monitor Why now: Open-weight releases make neuron-level white-box attacks on safety (ablating or steering refusal neurons) a practical threat. A training-time defense that spreads safety across many neurons is a timely angle on open-weight safety. Sun, 02 Aug 2026 17:49:07 +0000 2608.01414 http://arxiv.org/abs/2610.39633v1 One-step and few-step text-to-image generation must accurately render the details specified in the prompt despite limited opportunities for refinement. Distribution matching distillation (DMD) uses score differences to bring the generated distribution closer to a reference distribution. Following prior work, we estimate this difference with internal velocity branches, avoiding a separate fake-score network. We ask whether an explicit moment constraint improves this learned score supervision. ... Public: 0.536 | Memory: 0.040 | Status: Deferred this cycle Why now: Few-step text-to-image distillation is an active area, and IDM reports top LongText-Bench scores at two network evaluations on Z-Image-Turbo 6B and Qwen-Image 20B. It also finds that generator EMA features match an external DINOv2 encoder, which removes a dependency. Thu, 29 Oct 2026 08:57:17 +0000 2610.39633 http://arxiv.org/abs/2606.06080v1 Reinforcement learning with verifiable rewards is widely used for post-training reasoning models, but sparse outcome rewards make exploration difficult. A complementary approach is to optimize inference-time objectives such as pass@K and max@K directly, yet existing policy-gradient estimators for these objectives use different signals, baselines, and normalizations, making their relationships unclear. We study this issue through baseline design and advantage centering. Starting from the advan... Public: 0.487 | Memory: 0.085 | Status: Deferred this cycle Why now: RLVR post-training for reasoning models is a hot area, and pass@K/max@K objectives are a live approach to the exploration problem. The paper gives a clean unification of existing advantage estimators and a centered, unbiased Leave-Two-Out baseline that drops into group-based RL like GRPO. Thu, 04 Jun 2026 12:16:39 +0000 2606.06080 http://arxiv.org/abs/2606.05749v1 Iterative retrieval-reasoning agents have recently shown promise for multimodal long-document question answering. However, most existing systems maintain a single growing context that mixes retrieval traces, observations, and intermediate reasoning. As interactions accumulate, key evidence becomes scattered and diluted, making multi-hop reasoning noisy. We propose MARDoc, a Memory-Aware Refinement Agent framework that decouples long-document QA into three specialized agents: an Explorer for m... Public: 0.427 | Memory: 0.220 | Status: Deferred this cycle Why now: Agentic document QA with structured, refined memory is a visible trend, and separating retrieval, refinement and reflection is a pattern that applies to other long-context agents. Even so, nothing in the abstract suggests a shift in how systems are built or evaluated in the next 30-90 days. Thu, 04 Jun 2026 06:23:01 +0000 2606.05749 http://arxiv.org/abs/2608.01619v1 Memory-augmented agents can know that a user's stored state is outdated and still plan around the old value. The STALE benchmark calls this the implicit policy adaptation (IPA) gap. We identify one structural contributor: draft-anchored verification checks what a response says, and in an open-ended response the stale dependency is usually unsaid. StateAuditor therefore audits in the opposite direction, from stored state to draft. An LLM proposes candidate old-to-new transitions from timestamp... Public: 0.425 | Memory: 0.177 | Status: Deferred this cycle Why now: Stale personal state is a real failure mode for agent memory products, and this paper gives a concrete, measured fix on the STALE benchmark. The +5.0-point paired gain is backed by a matched control and a second judge. Mon, 03 Aug 2026 02:49:08 +0000 2608.01619 http://arxiv.org/abs/2610.02851v1 3D Gaussian Splatting (3DGS) enables photorealistic scene reconstruction, but training large-scale scenes requires substantial memory and computation. Distributing training across multiple GPUs increases available memory capacity, yet its performance is strictly constrained by data movement. We identify two dominant bottlenecks: repeated off-chip memory accesses during forward and backward rasterization, and inter-GPU communication of partial Gaussian gradients. To address these bottlenecks, ... Public: 0.317 | Memory: 0.403 | Status: Deferred this cycle Why now: Distributed 3DGS training is bottlenecked by data movement, and ByteSplat shows concrete reductions: 63.4% less off-chip traffic and 65.8% less backward communication, for up to 6.1x speedup on 8 GPUs. It is a clean example of kernel fusion and sparse gradient communication. Fri, 02 Oct 2026 05:37:59 +0000 2610.02851 http://arxiv.org/abs/2607.03876v1 With the rise of small quantized GGUF-based language models and their increasing use for on-device inference tasks, we have seen the growing need for an approach capable of reliably delivering these models at scale even under severe memory bandwidth constraints such as those imposed by pure CPU implementations. Fixed-depth speculative decoding has emerged as one promising technique, but in practice, it often leads to performance degradation due to either bandwidth saturation, instability, or ... Public: 0.352 | Memory: 0.348 | Status: Deferred this cycle Why now: On-device CPU inference with small quantized GGUF models is growing, and speculative decoding under memory-bandwidth limits is a real systems problem. The paper's focus on stability and wasted draft compute, rather than raw throughput, is a useful framing. Sat, 04 Jul 2026 13:39:40 +0000 2607.03876 http://arxiv.org/abs/2608.02208v1 Whole-body dual-energy X-ray absorptiometry (DXA) scans are routinely acquired to measure bone density and regional body composition, leaving their spatial structure largely unused. Here, we show that self-supervised learning (SSL) can convert raw DXA images into representations of systemic health. We introduce LeDXA, a vision model based on a joint-embedding predictive architecture (JEPA) that learns by predicting latent representations rather than reconstructing pixels. Trained from scratch... Public: 0.516 | Memory: 0.052 | Status: Out of scope Why now: LeCun-backed JEPA applied to DXA scans shows that small self-supervised models trained from scratch can beat DINOv3 on a clinical domain, with external validation on 47k UK Biobank scans. It is a vivid data point for the 'latent prediction beats scale' argument. Mon, 03 Aug 2026 13:33:04 +0000 2608.02208 http://arxiv.org/abs/2610.40599v1 The hair of stylized characters in games and avatars is built from hair cards: artist-placed ribbons whose geometry is the visible hair, animated by spring bones. Existing image-based methods produce realistic hair or a single fused mesh, while CHARM, the only generator of stylized hair cards, decodes a hairstyle autoregressively as a long sequence of control points. Observing that a card hairstyle is an unordered set of artist-defined primitives, we propose HairDeck, which generates it as on... Public: 0.474 | Memory: 0.083 | Status: Out of scope Why now: A set-diffusion approach to generating animation-ready hair cards from one image is a neat result for game and avatar pipelines, with large gains over CHARM (44% lower Chamfer distance, 33x faster). It is a niche graphics contribution with little pull for the podcast's audience. Thu, 29 Oct 2026 17:50:26 +0000 2610.40599 http://arxiv.org/abs/2606.06147v1 End-to-end Vision-Language-Action (VLA) models have shown promise in UAV navigation. However, existing approaches typically rely on historical observations to directly predict actions, often struggling in dense urban environments where severe occlusions and sharp turns result in drastic viewpoint transitions. We argue that the ability to "imagine" future states -- inherent in World Models -- is critical for robust decision-making under such partial observability. To address this, we construct... Public: 0.440 | Memory: 0.098 | Status: Out of scope Why now: World-model-based VLA policies are a hot embodied-AI direction, and joint video and action generation via flow matching is a timely idea. UAV urban navigation is a vivid demo, but the paper's reach stays within that niche. Thu, 04 Jun 2026 13:23:05 +0000 2606.06147 http://arxiv.org/abs/2606.05718v1 On-policy distillation (OPD) improves reasoning by training a student on trajectories sampled from its own policy under supervision from a teacher. In multimodal reasoning, a common extension is to use a privileged teacher that observes training-time-only signals such as reference answers or rationales. However, such answer-side privilege creates a train-test mismatch: the teacher's supervision may depend on signals unavailable to the student, encouraging shortcut imitation rather than visual... Public: 0.420 | Memory: 0.137 | Status: Out of scope Why now: On-policy distillation is a live topic, and the paper argues that the design of teacher privilege matters as much as teacher strength. That argument could transfer beyond vision. The evidence is limited to multimodal reasoning on Qwen3-VL. Thu, 04 Jun 2026 05:18:13 +0000 2606.05718 http://arxiv.org/abs/2606.05626v2 Machine-generated text (MGT) attribution aims to identify the specific generator responsible for a given text, thereby providing fine-grained evidence for model accountability and misuse investigation. As new large language models continue to emerge, attribution models must continuously incorporate new generators while preserving their ability to recognize previously seen ones. Prior works have shown that this lifelong MGT attribution setting is challenging, and existing methods often struggl... Public: 0.378 | Memory: 0.121 | Status: Out of scope Why now: New LLMs keep appearing, so attributing text to a specific generator is a real accountability problem. The replay-free ridge update is a tidy continual-learning recipe. Thu, 04 Jun 2026 02:50:58 +0000 2606.05626