Multilingual speech technologies still fail to reach most of the world's languages because existing corpora are limited to a few hundred well-resourced languages and narrow domains like parliamentary proceedings or Bible readings. ELARSpeech addresses this gap by curating 30,202 hours of speech across 675 underrepresented languages from the Endangered Languages Archive, yielding both labeled and unlabeled subsets with rich acoustic and communicative diversity. It matters because it provides a large-scale, publicly available testbed that reveals existing self-supervised speech models struggle on the true linguistic long tail despite massive pretraining.

Base models are hard to evaluate on agentic tasks because standard benchmarks assume instruction-following chat interfaces. ASAP adapts existing agent scaffolds to pretrained base models by treating execution as text continuation, using a Markdown+XML format, a short reference trajectory prefix, and minimal failure recovery. The framework yields discriminative end-to-end scores across models and checkpoints, and exposes execution trajectories that enable fine-grained behavioral diagnosis before post-training.

STRATA tackles the problem of scalable, self-maintaining memory for test-time self-evolving LLM agents. Instead of a flat, monotonic pool of episode summaries, it organizes experience as a typed three-layer directed graph of atomic skills, task workflows, and meta-frameworks, maintained by five mechanisms including multi-dimensional similarity, adaptive reflection thresholds, community detection, pressure-triggered garbage collection with articulation-point protection, and contrast edges that warn against directionally opposite experiences. The authors show that this structured memory attains the best average accuracy across four interactive benchmarks while simultaneously using the lowest latency, fewest environment steps, and fewer injected memory lines than seven memory-augmented baselines.

Configuring distributed LLM training over hundreds of GPUs involves a combinatorial explosion over tensor, pipeline, expert, context, and data parallelism plus micro-batch size, recomputation, and sharding, where a single byte of activation accounting can separate a legal run from an OOM crash. This paper proposes Primus Projection, a hybrid engine that benchmarks compute kernels on as few as one GPU and analytically reconstructs out-of-scope communication and pipeline schedules at the target scale, coupled with an LLM-driven tuning agent that searches the space in minutes instead of days. The authors validate across dense and MoE models on two AMD GPU generations, reporting projection errors around $5\%$ at up to 1,024 GPUs and a $+27\%$ throughput gain discovered in under 30 minutes of single-node exploration.

ixi-sLM tackles the problem of running an AI call agent entirely on a mobile device under constraints that cloud deployments never face: roughly 1 GiB of memory, in-call thermal limits, static-integer NPU execution, Korean STT input riddled with disfluencies, and user expectations set by cloud LLMs. The core idea is to take a 1.2B-parameter Korean sLM and apply two service-specific optimizations—domain-adaptive vocabulary trimming and tokenizer merging to shrink footprint and token counts, followed by task-aware mixed-precision quantization calibrated on real calls—to squeeze the model onto Qualcomm and Apple NPUs. It matters because the authors claim a live production deployment serving several million subscribers at 94% of an unnamed cloud LLM's quality, cutting per-call energy to 23% of a CPU baseline.

LLM agents face an expanded attack surface because malicious content can hide in user queries, tool descriptions, and environment observations. AgentSafeEvo casts agent safety as an attacker-defender co-evolutionary game over multi-turn tool interactions, where the attacker constructs full adversarial tasks and the defender learns to execute tasks safely. A novel learnability-guided task weighting mechanism uses within-task reward variance to shape updates for both players. Experiments on Qwen3-8B show strong safety gains and broad preservation of general capabilities.

EvoWeave addresses how interactive agents can reuse past experience without importing entire workflows or discarding step dependencies. It proposes a graph memory where nodes are abstract states and edges are reusable operations annotated with contracts, provenance, and capability references; a task planner guides retrieval and bounded graph search to assemble experience fragments. Experiments across four benchmarks and three LLM backbones show heterogeneous gains and highlight tradeoffs between retrieval precision and runtime overhead.

Program synthesis and library learning generate useful operations, but the resulting artifacts are often treated merely as derivations or search traces rather than reusable operational objects. This paper introduces FormCraft, a framework that represents both primitive and composed operations as first-class capabilities in a uniform well-formed space $\mathfrak{F}$, separating instance-independent formation and public callability from optional persistence through a verification gate. By retaining formation provenance while exposing composites through the same public interface as primitives, the framework aims to make persistent library growth possible and to eliminate the need to reconstruct historical derivation chains at reuse time, a claim backed by a paired experiment showing orders-of-magnitude reductions in search work.

Choosing a parallelism configuration for distributed LLM inference is challenging because the prefill and decode phases impose conflicting demands on time-to-first-token (TTFT) and inter-token latency (ITL). This paper develops a systematic framework that defines parallelism operating regions via TTFT-ITL Pareto frontiers and formulates SLO-aware selection using goodput $G_p(w,T_s)=\lambda_p(w)F_p(T_s|w)$. The core finding is a workload-dependent transition: tensor-parallel (TP) configurations excel for decode-heavy workloads and relaxed latency requirements, while pipeline-parallel (PP) configurations dominate for prefill-heavy workloads, stringent TTFT targets, and long contexts. The work matters because it moves parallelism selection from a static deployment choice to a workload- and SLO-dependent operating decision.

Long-horizon language agents frequently re-issue identical tool calls and receive redundant observations, a failure mode the authors call a loop. The paper shows that compressing the agent's history to fit a context window dramatically increases these loops, which in turn strongly predict task failure. DeLoop solves this by mining preference pairs directly from the agent's own rollouts—treating the looping action as a rejected example and the first later non-looping action as a chosen one—and training with direct preference optimization without any external rewards, gold answers, or LLM judges.

KoAGE investigates whether large language models can know, learn from context, and coin Korean neologisms. The authors introduce NeoTrace, an automated pipeline that collects and dates 4,271 neologisms across six genres, and propose a cutoff-aware benchmark that separates lexical recall from contextual inference and generation. By comparing model outputs with attested human coinages under original materials and simulated community settings, the paper offers a temporally grounded view of how models adapt to evolving vocabulary.

This paper exposes a critical flaw in how compressed memory is evaluated for large language models: standard ablation tests like zeroing memory vectors do not prove that stored document content improves predictions. The authors distinguish three empirically separable questions—whether a model depends on memory after training, whether training with memory outperforms training without it, and whether gains actually require document content—and argue that evaluation claims must be matched to specific controlled comparisons. Across nine memory systems and multiple QA benchmarks, they show these comparisons frequently lead to different conclusions, calling for stricter evidentiary standards in context-compression research.

Policy-Verified Self-Training asks whether a runtime policy gate's discarded verdicts can be recycled as training signals for the same agent. The paper proposes PVST: an agent rolls out trajectories, a deterministic gate compiled from written policy judges compliance, and the verdicts become acceptance filters and preference labels. On AgentDojo prompt-injection benchmarks, post-training Qwen3-8B with these verifier-derived rewards cuts violating trajectories from 37.2% to 10.3% and successful attacks to 0.5% without any stronger teacher model, human preference labels, or attack-specific supervision. The result suggests that compiled policy checkers can serve not only as runtime constraints but as a scalable, auditable source of self-training supervision.

LLM agents frequently fail in multi-turn settings by violating policies, using tools incorrectly, or acting prematurely. This paper proposes Critique-Augmented Agentic Fine-Tuning (CAFT), which rolls out a student agent, has an expert label each wrong turn with both a textual critique and a corrected action, and then fine-tunes the student to generate that critique and correction while keeping the original mistake in context. The central finding is that these natural-language explanations provide learning signal beyond action-only supervision, improving task success across four models and two benchmarks.

Multimodal agents need persistent memory to act coherently over long horizons, yet forming and managing memory across heterogeneous, high-dimensional observations—video, audio, 3D geometry, and GUIs—has lacked a unifying analytical lens. This survey introduces a two-axis lifecycle framework that classifies how experience is represented (source, textual, structured, latent, or parametric) and how memory states and policies evolve ($T1$ Static, $T2$ Dynamic, or $T3$ Self-Evolving). By analyzing 297 architectures, the authors identify key trade-offs between fidelity and efficiency and expose the striking scarcity of systems that evolve their own memory policies at deployment time.

DangerLens asks whether two language models with identically safe refusal behavior can still differ in how much hazardous knowledge remains internally accessible. The authors train paired Qwen1.5 checkpoints—one directly safety-aligned, the other first exposed to harmful data then aligned to the same standard—and introduce a vocabulary-space probe that uses a Jacobian lens and blinded semantic scoring to count harmful goals and operational means inside hidden states. The work matters because current safety audits rely on surface behavior alone, leaving latent dangerous representations undetected before downstream fine-tuning or attacks.

Best-of-N sampling improves LLM output quality by generating multiple candidates and selecting the highest reward, yet its scaling behavior under imperfect reward models is poorly understood. This paper recasts BoN as extreme-value competition between the upper tails of correct and incorrect reward distributions, yielding phase transitions that explain both monotonic gains and reward hacking. Using extreme value theory, the authors derive finite-sample scaling laws showing polynomial rather than exponential convergence, and propose a lightweight tail-tension criterion for label-free early stopping. The work matters because it replaces vague intuitions about reward hacking with precise, testable distributional conditions and a practical inference-time safeguard.

Activation steering is used to steer language models toward named behaviors, but standard baselines cannot tell whether a vector controls a specific concept or a neighboring one. This paper introduces the competing-target test, which compares a steering vector with the matched vector for a competing target and bounds their difference with an equivalence test. In three open-weight models, the authors show that vectors for Kantian deontology and act utilitarianism are largely interchangeable---they move both framework scores alike---whereas explicit prompting separates them, and the same construction does separate optimistic and pessimistic tones.

This paper studies the infinite-generation limit of autoregressive transformers. Rather than fixing sequence length, the authors let generation time tend to infinity and ask whether the model's empirical state converges. Their core result is that, for a broad class of architectures, the stochastic generation process tracks a deterministic ODE on a slow time scale, and every fixed transformer has a finite temperature threshold above which it converges to a unique limit. However, no uniform temperature works for all models; for any finite temperature, some transformer can fail to converge, even with just one layer and three tokens.

AI agents chain many LLM calls to complete tasks, yet models are usually compared per call. This paper argues the correct functional unit is the successfully completed task, because failed attempts waste energy, cost and time. It derives crossover task lengths that predict when a smaller model stops being greener, cheaper or faster, and shows with a 15-model pilot that the greenest choice shifts from 1-3B models for short tasks to 8-27B models for long ones.