Token-Level Uncertainty

As large language models (LLMs) move from centralized clouds to mobile edge environments, efficient serving must balance latency, energy consumption, and accuracy under constrained device-edge resources. Query-level routing between lightweight on-device models and stronger edge models provides a flexible mechanism to navigate this trade-off. However, existing routers are designed for centralized cloud settings and optimize token-level costs, failing to capture the dynamic latency and energy overheads in wireless edge deployments. In this paper, we formulate mobile edge LLM routing as a deployment-constrained, cost-aware decision problem, and propose CR^2, a two-stage device-edge routing framework. CR^2 decouples a lightweight on-device margin gate from an edge-side utility selector for deferred queries. The margin gate operates on frozen query embeddings and a user-specified cost weight to predict whether local execution is utility-optimal relative to the best edge alternative under the target operating point. We further introduce a conformal risk control (CRC) calibration procedure that maps each operating point to an acceptance threshold, enabling explicit control of the marginal false-acceptance risk under the full-information utility reference. Experiments on the routing task show that CR^2 closely matches a full-information reference router using only device-side signals before deferral. Compared with strong query-level baselines, CR^2 consistently improves the deployable accuracy-cost Pareto frontier and reduces normalized deployment cost by up to 16.9% at matched accuracy.

Aligning large language models (LLMs) with human values typically relies on post-training or inference-time steering that directly manipulates the backbone's parameters or representation space. However, a critical gap exists: the model's residual stream is highly dynamic, in which values exist as fragile, low-dimensional properties, inherently incompatible with the stability required for consistent value expression. In this paper, we propose the Stable Value Guidance Transformer (SVGT), which addresses this gap through an independent value module incorporating two key designs: (1) independent value modeling, maintaining normative representations in a dedicated value space isolated from the backbone, and (2) explicit behavioral guidance, transducing these stable signals into learnable latent Bridge Tokens. These tokens serve as dynamic value anchors to explicitly steer the generative trajectory, ensuring robust adherence across diverse contexts without disrupting the backbone's internal representations. Experiments across multiple backbones and safety benchmarks show that SVGT generally reduces harmful scores by over 70% while maintaining generation fluency, demonstrating the efficacy of architecturally grounded value modeling. Our code is available at https://github.com/Clervils/SVGT.git.

We present ScaleMoGen, a scale-wise autoregressive framework for text-driven human motion generation. Unlike conventional autoregressive approaches that rely on standard next-token prediction, ScaleMoGen frames motion generation as a coarse-to-fine process. We quantize 3D motions into compositional discrete tokens across multiple skeletal-emporal scales of increasing granularity, learning to generate motion by autoregressively predicting next-scale token maps. To maintain structural integrity, our motion tokenizers and quantizers are explicitly designed so that discrete tokens at every scale strictly preserve the skeletal hierarchy. Additionally, we employ bitwise quantization and prediction, which efficiently scale up the tokenizer vocabulary to preserve motion details and stabilize optimization. Extensive experiments demonstrate that ScaleMoGen achieves state-of-the-art performance, establishing an FID of 0.030 (vs. 0.045 for MoMask) on HumanML3D and a CLIP Score of 0.693 (vs. 0.685 for MoMask++) on the SnapMoGen dataset. Furthermore, we demonstrate that our skeletal-temporal multi-scale representation naturally facilitates training-free, text-guided motion editing.

Activation steering controls language model behavior by adding directions to internal representations at inference time, but standard residual-stream steering can fail in stateful dialogue. We identify KV-cache contamination as a key failure mode: steered token states are stored and repeatedly reused, turning a local perturbation into cumulative coherence degradation. To address this challenge, we propose Gated Cropped Attention-Delta steering (GCAD), which extracts steering signals from system-prompt contributions to self-attention and applies them with token-level gating. Across persona-steering experiments, GCAD preserves trait control while substantially improving long-horizon coherence. On the main multi-turn benchmark, GCAD improves average coherence drift from -18.6 to -1.9 and raises turn-10 trait expression from 78.0 to 93.1. These results suggest that activation steering becomes more reliable when interventions follow the prompt-mediated pathways that models already use for behavioral control.

Current 3D tokenizers largely treat representation as spatial compression: compact codes reconstruct surface geometry, but leave component ownership and attachment validity implicit. In open-world assets with intersecting components, noisy topology, and weak canonical structure, this creates a representation mismatch: local shape, component identity, and assembly relations become entangled in a latent stream and are not natively addressable during decoding. We formulate an alternative view, interface-centric generative states, in which tokenization constructs an operational state rather than a passive compressed code. The state exposes local geometry, component ownership, and attachment validity as variables that can be queried, constrained, and repaired during decoding. We instantiate this formulation with Component-Conditioned Canonical Local Tokens (C2LT-3D), factorizing representation into canonical local geometry, partition-conditioned context, and relational seam variables. Each factor targets a distinct failure mode of compression-centric tokens: pose leakage, cross-component interference, or invalid local attachment. This exposed state supports attachment validation, latent structural repair, targeted intervention, and constrained serialization without a separate post-hoc structure recovery module. Trained on single-object CAD models and evaluated zero-shot on open-world multi-component assets, C2LT-3D improves structural robustness and shows that its latent variables remain actionable under adversarial attachment settings. These results suggest that open-world 3D generative representations should be evaluated not only by reconstruction fidelity, but by whether their discrete states remain operational for assembly-level structural reasoning.

Modern LLM APIs often reveal only top-$K$ logit scores and censor the remaining vocabulary. We study the per-position distribution-recovery limits of this access model. For censoring threshold $τ$, the compatible teacher distributions form an identified set whose total-variation diameter is exactly $U_K=(V-K)\exp(τ)/(Z_A+(V-K)\exp(τ))$, where $Z_A$ is the observed partition function. For KL recovery, we give a computable binary-endpoint lower bound and an asymptotically matching small-ambiguity upper bound, with an extension to reference-aware attackers. Experiments on a Qwen3 math-reasoning teacher reveal a layered extraction hierarchy: on-task top-$K$ distillation recovers 12% of private capability, full-logit distillation recovers 56% despite 99% KL closure, and generation-based extraction recovers 96%. Top-$K$ censoring therefore limits per-position distribution recovery but does not by itself prevent capability extraction, separating fidelity from transfer in prompt-only logit distillation.

Multimodal models that jointly reason over protein sequences, structures, and function annotations within a unified representation hold immense potential for integrating multimodal data and generating new proteins with designed functional properties. To utilize transformer architectures, such models require a tokenizer that converts protein structure from continuous atomic coordinates into discrete representations suitable for scalable multimodal training. The quality of such models are fundamentally upper bounded by the fidelity and expressiveness of the underlying tokenized structure. However, existing tokenizers prioritize reconstruction over generative abilities. To address these gaps, we introduce Yeti, a simple and compact protein structure tokenizer based on lookup free quantization and trained end to end with a flow matching objective for multimodal learning. Compared to existing models, Yeti generally achieves the best codebook utilization and token diversity, and second best reconstruction accuracy (with 10x fewer parameters than ESM3) on diverse datasets. To validate Yeti's generative capability, we trained a compact multimodal model jointly over its structure tokens and amino acid sequence entirely from scratch, with no pretrained initialization. The resulting multimodal model generates plausible structures under unconditional cogeneration of protein sequence and structures, achieving comparable results to 10x larger models. Together, these results demonstrate that Yeti is a compact and expressive protein structure tokenizer suitable for training multimodal models that cogenerates highly plausible sequences and structures.

We present MambaNetBurst, a compact tokenizer-free byte-level sequence classifier for network burst classification based on a Mamba-2 backbone. In contrast to most recent strong traffic-classification and intrusion-detection approaches, our method operates directly on raw packet bytes, avoids tokenization, patching, and heavy engineered multimodal representations, and does not require any self-supervised pre-training stage. Given a packet flow, we form a fixed-length burst from the first few packets, embed the resulting byte sequence appending a learnable CLS token, and process it with a stack of residual pre-normalized Mamba-2 blocks for end-to-end supervised classification. Across six public benchmarks spanning encrypted mobile app identification, VPN/Tor traffic classification, malware traffic classification, and IoT attack traffic, MambaNetBurst achieves consistently strong results and is competitive with, or outperforms, substantially heavier and often pre-trained baselines. Our ablation study shows that preserving byte-level temporal resolution is critical, that early downsampling through striding is consistently harmful, and that moderate state sizes are sufficient for robust generalization. We further show that Mamba-2, despite its more constrained transition structure relative to Mamba-1, remains highly effective for packet-byte modeling while providing clear efficiency advantages, particularly in training speed. Overall, our results demonstrate that direct undiluted byte-to-classification learning with compact selective state space models is a practical, effective and novel direction for efficient, deployable traffic analysis that bypasses the complexity of pre-training pipelines even over highly optimized linear attention architectures.

Large language models (LLMs) exhibit a strong capacity for in-context learning: Given labeled examples, they can generate good predictions without parameter updates. However, many interactive settings go beyond static prediction to online decision-making, in which effective behavior demands adaptation over long multi-turn horizons in response to feedback, and efficient algorithms in these domains must use compact representations of what they have learned. Recently, continuous transformer architectures with latent chain of thought have shown promise for offline iterative tasks such as directed graph-reachability. Motivated by this, we study whether continuous latent context tokens equip transformers to more effectively realize online learning. We give explicit constructions of constant-depth transformers that implement two foundational online decision-making procedures -- the weighted majority algorithm and $Q$-learning -- by storing their algorithmic state as linear combinations of feature embeddings, using a small number of latent context tokens. We further train a small GPT-2-style transformer with latent contexts using a multi-curriculum objective that does not directly supervise the latent states. On long synthetic online prediction sequences, this model outperforms larger and more complex LLMs, including Qwen-3-14B and DeepSeek-V3. Our results suggest that continuous latent contexts provide a simple and effective persistent state for transformers to implement online learning algorithms.

Large Language Models (LLMs) have demonstrated exceptional performance across diverse tasks. However, their deployment in long-context scenarios faces high computational overhead and information redundancy. While soft prompt compression has emerged as a promising way to mitigate these costs by compressing sequences into compact embeddings, existing paradigms remain fundamentally constrained by position bias: they primarily rely on learnable tokens insertion at fixed positions or group tokens according to their physical token layout, thereby inducing performance instability and semantic fragmentation. To overcome this bottleneck, we propose Semantic Consistency Context Compression (SeCo), a method that shifts context compression from position-driven to semantic-driven. Rather than constraint by physical token layout, SeCo dynamically anchors compression directly in the semantic space by selecting query-relevant tokens as semantic centers and aggregating remaining tokens via consistency-weighted merging. This design inherently preserves semantic consistency while eliminating position bias. Extensive experiments on 14 benchmarks across two backbone models demonstrate that SeCo consistently shows superiority in downstream tasks, inference latency, and out-of-domain robustness. The code is available at https://anonymous.4open.science/r/seco-EE5E.

The Chain-of-Thought (CoT) paradigm, while enhancing the interpretability of Large Language Models (LLMs), is constrained by the inefficiencies and expressive limits of natural language. Latent Chain-of-Thought (latent CoT) reasoning, which operates in a continuous latent space, offers a promising alternative but faces challenges from structural complexities in existing multi-step or multi-model paradigms, such as error propagation and coordination overhead. In this paper, we introduce One-Model One-Step, a novel compression framework for Latent Reasoning with Rule-Based Priors(RuPLaR) to address this challenge. Our method trains an LLM to autonomously generate latent reasoning tokens in a single training stage, guided by rule-based prior probability distributions, thereby eliminating cascaded processes and inter-model dependencies. To ensure reasoning quality, we design a joint training objective that enforces answer consistency via cross-entropy, aligns soft tokens with rule-based priors via KL divergence (the Soft Thinking constraint), and adds a problem-thought semantic alignment constraint in the representation space. Extensive experiments show that our compression framework not only improves accuracy by 11.1% over existing latent CoT methods but also achieves this with minimal token usage, underscoring its effectiveness and extensibility. Code: https://github.com/xiaocen-luo/RuPLaR.

While recent work in Reinforcement Learning with Verifiable Rewards (RLVR) has shown that a small subset of critical tokens disproportionately drives reasoning gains, an analogous token-level understanding of On-Policy Distillation (OPD) remains largely unexplored. In this work, we investigate high-loss tokens, a token type that--as the most direct signal of student-teacher mismatch under OPD's per-token KL objective--should progressively diminish as training converges according to existing studies; however, our empirical analysis shows otherwise. Even after OPD training reaches apparent saturation, a substantial subset of tokens continues to exhibit persistently high loss; these tokens, which we term Rock Tokens, can account for up to 18% of the tokens in generated outputs. Our investigation reveals two startling paradoxes. First, despite their high occurrence frequency providing a disproportionately large share of total gradient norms, Rock Tokens themselves remain stagnant throughout training, resisting teacher-driven corrections. Second, through causal intervention, we find that these tokens provide negligible functional contribution to the model's actual reasoning performance. These findings suggest that a vast amount of optimization bandwidth is spent on structural and discourse residuals that the student model cannot or need not internalize. By deconstructing these dynamics, we demonstrate that strategically bypassing these ``stumbling blocks'' can significantly streamline the alignment process, challenging the necessity of uniform token weighting and offering a more efficient paradigm for large-scale model distillation.

We present Dolphin-CN-Dialect, a streaming-capable ASR model with a focus on Chinese and dialect-rich scenarios. Compared to the previous version, Dolphin-CN-Dialect introduces substantial improvements in data processing, tokenization, training stability, and data sampling strategies. To address the challenges of highly imbalanced dialect data, we propose a temperature-based sampling strategy that effectively balances standard Mandarin and low-resource dialects, leading to significant gains in dialect recognition performance. In addition, we redesign the tokenizer to better align with linguistic characteristics, adopting character-level modeling for Chinese and subword modeling for English, while introducing extensible dialect tokens. Experimental results show that Dolphin-CN-Dialect achieves improvement in dialect recognition accuracy and CER reduction compared to Dolphin. Furthermore, Dolphin-CN-Dialect reaches competitive performance with recent SOTA open-source ASR models, while maintaining a significantly smaller model size. Dolphin-CN-Dialect supports both streaming and non-streaming inference, enabling a practical balance between latency and accuracy. It also provides flexible customization through hotword support and efficient deployment optimized for specialized hardware. These improvements make Dolphin-CN-Dialect a strong and practical solution for real-world multi-dialect ASR applications.

Diffusion Language Models (dLLMs) have garnered significant attention for their potential in highly parallel processing. The parallel capabilities of existing dLLMs stem from the assumption of conditional independence at high confidence levels, which ensures negligible discrepancy between the marginal and joint distributions. However, the stringent confidence thresholds required to preserve accuracy severely constrain the scalability of parallelism. Through systematic token-level statistical analysis, we reveal that a substantial proportion of tokens converge to their correct predictions early in the denoising process yet fail to reach standard confidence thresholds, confirming that current confidence-based criteria are overly conservative. In response, we introduce LEAP (Lookahead Early-Convergence Token Detection for Accelerated Parallel Decoding). LEAP is a training-free, plug-and-play method that leverages future context filtering and multi-sequence superposition to detect early-converging tokens. By validating the alignment between early convergence and correctness, we enable reliable early decoding of these tokens. Benchmarking across diverse domains demonstrates that LEAP significantly lowers inference latency and decoding steps. Compared to confidence-based decoding, the average number of denoising steps is reduced by about 30%. On the GSM8K dataset, combining LEAP with dParallel accelerates decoding to 7.2 tokens per step while preserving model precision. LEAP effectively breaks the reliance on high-confidence priors, offering a novel paradigm for parallel decoding.

Tokenizers are a crucial component of latent diffusion models, as they define the latent space in which diffusion models operate. However, existing tokenizers are primarily designed to improve reconstruction fidelity or inherit pretrained representations, leaving unclear what kind of latent space is truly friendly for generative modeling. In this paper, we study this question from the perspective of latent manifold organization. By constructing controlled tokenizer variants, we identify three key properties of a diffusion-friendly latent manifold: coherent spatial structure, local manifold continuity, and global manifold semantics. We find that these properties are more consistent with downstream generation quality than reconstruction fidelity. Motivated by this finding, we propose the Prior-Aligned AutoEncoder (PAE), which explicitly shapes the latent manifold instead of leaving diffusion-friendly manifold to emerge indirectly from reconstruction or inheritance. Specifically, PAE leverages refined priors derived from VFMs and perturbation-based regularization to turn spatial structure, local continuity, and global semantics into explicit training objectives. On ImageNet 256x256, PAE improves both training efficiency and generation quality over existing tokenizers, reaching performance comparable to RAE with up to 13x faster convergence under the same training setup and achieving a new state-of-the-art gFID of 1.03. These results highlight the importance of organizing the latent manifold for latent diffusion models.

DeepSeek Sparse Attention (DSA) sets the state of the art for fine-grained inference-time sparse attention by introducing a learned token-wise indexer that scores every prefix token and selects the most relevant ones for the main attention. To remain expressive, the indexer uses many query heads (for example, 64 on DeepSeek-V3.2) that share the same selected token set; this multi-head design is precisely what makes the indexer the dominant cost on long contexts. We propose MISA (Mixture of Indexer Sparse Attention), a drop-in replacement for the DSA indexer that treats its indexer heads as a pool of mixture-of-experts. A lightweight router uses cheap block-level statistics to pick a query-dependent subset of only a few active heads, and only those heads run the heavy token-level scoring. This preserves the diversity of the original indexer pool while reducing the per-query cost from scoring every prefix token with every head to scoring it with only a handful of routed heads, plus a negligible router term computed on a small set of pooled keys. We further introduce a hierarchical variant of MISA that uses the routed pass to keep an enlarged candidate set and then re-ranks it with the original DSA indexer to recover the final selected tokens almost exactly. With only eight active heads and no additional training, MISA matches the dense DSA indexer on LongBench across DeepSeek-V3.2 and GLM-5 while running with eight and four times fewer indexer heads respectively, and outperforms HISA on average. It also preserves fully green Needle-in-a-Haystack heatmaps up to a 128K-token context and recovers more than 92% of the tokens selected by the DSA indexer per layer. Our TileLang kernel delivers roughly a 3.82 times speedup over DSA's original indexer kernel on a single NVIDIA H200 GPU.

Metaphor is pervasive in everyday language, yet token-level computational identification of metaphor-related words in Chinese under the MIPVU framework remains under-explored relative to English. This paper presents a reproducible multi-architecture baseline for token-level metaphor identification on the PSU Chinese Metaphor Corpus (PSU CMC), the only widely available MIPVU-annotated Chinese corpus. We systematically compare three model families: (i) encoder fine-tuning with Chinese RoBERTa-wwm-ext-large; (ii) MelBERT adapted to Chinese using a newly constructed basic-meaning resource derived from the Modern Chinese Dictionary, 7th edition (MCD7), comprising 74,823 entries with 71.51% PSU CMC vocabulary coverage; and (iii) Qwen3.5-9B fine-tuned with QLoRA as an instruction-tuned generative baseline. Across five fixed seeds, MelBERT MIP-only achieves the strongest performance at 0.7281 +/- 0.0050 test positive F1, marginally above MelBERT Full (0.7270 +/- 0.0069) and clearly above plain RoBERTa (0.7142 +/- 0.0121). The Qwen QLoRA generative configuration trails encoder baselines by approximately 11 F1 points (0.6157 +/- 0.0113). Three findings merit attention: (1) the SPV channel of MelBERT does not contribute reliable positive signal in Chinese, consistent with the dominance of conventional metaphor; (2) the Qwen-encoder gap is concentrated in recall, reflecting the discrete-commitment limitation of generative output; (3) several Qwen task formulations fail due to format design rather than model capacity. We release all split manifests, per-seed outputs, the MCD7 basic-meaning embedding pipeline, and training scripts to serve as a common reference for future Chinese metaphor identification research.

Autoregressive generation lies at the heart of the mechanism of large language models. It can be viewed as the repeated application of a next-token generator: starting from an input string (prompt), the generator is applied for $M$ steps, and the last generated token is taken as the final output. [Joshi et al., 2025] proposed a PAC model for studying the learnability of the input-output maps arising from this process. We develop an online analogue of this framework, focusing on the mistake bound of learning the final output induced by an unknown next-token generator. We distinguish between two forms of feedback. In the End-to-End model, after each round the learner observes only the final token produced after $M$ autoregressive steps. In the Chain-of-Thought model, the learner is additionally shown the entire $M$-step trajectory. Our goal is to understand how the optimal mistake bound depends on the generation horizon $M$, and to what extent observing intermediate tokens can reduce this dependence. Our main results show that the online theory of autoregressive learning exhibits a qualitative picture analogous to the statistical one found by [Hanneke et al., 2026], but with a different scale of dependence on the generation horizon. In the End-to-End model, we prove a taxonomy of possible mistake-bound growth rates in the generation horizon $M$: essentially any rate between constant and logarithmic can arise. We further show that this logarithmic ceiling is unavoidable. In the Chain-of-Thought model, we show that access to the full generated trajectory eliminates the dependence on $M$ altogether. We also analyze autoregressive linear threshold classes, and prove optimal mistake bounds, as well as a new lower bound for the statistical setting. Along the way, our results resolve several questions left open by [Joshi et al., 2025].

Modern learning systems represent perceptual signals with continuous vectors, but comparison, retrieval, memory, alignment, and reasoning are often naturally symbolic. In language, this interface is given by tokens; for speech and audio, it must be learned. Existing audio tokenizers use local quantization, clustering, or reconstruction, leaving sequence consistency, compactness, length control, termination, and edit geometry indirectly optimized. We introduce PairAlign, a framework for compact audio tokenization through sequence-level self-alignment. PairAlign treats tokenization as conditional sequence generation: an encoder maps speech to a condition, and an autoregressive decoder emits tokens from BOS to EOS, learning identity, order, length, and termination. Given two content-preserving views, each token string is trained to be likely under the other's representation, while unrelated examples provide competing sequences. This yields a surrogate for edit-distance preservation while discouraging collapse. Starting from a VQ tokenizer, PairAlign extends a frame-synchronous prior into an autoregressive tokenizer using VQ-derived and EMA-teacher targets, cross-paired teacher forcing, anti-bypass regularization, likelihood contrast, length control, and timing recovery. On 3 s speech, PairAlign learns compact token strings with strong cross-view consistency. In retrieval, it operates at 12.71 tokens/s and reduces archive tokens by 55% versus VQ while preserving edit-distance search. The results expose a compactness--locality trade-off: PairAlign does not aim to dominate dense geometric or SSL tokenizers on every local metric, but provides a lower-rate symbolic interface for comparison, retrieval, and analysis. More broadly, PairAlign is a sequence-symbolic analogue of JEPA-style predictive learning, predicting a learned variable-length symbolic sequence rather than a continuous latent.

On-policy distillation (OPD) trains a student on its own trajectories with token-level teacher feedback and often outperforms off-policy distillation and standard reinforcement learning. However, we find that its standard advantage weighted policy gradient suffers from three structural weaknesses, including high variance updates, vanishing gradients in zero-advantage regions, and exploration bottlenecks when corrective signals are insufficient. We therefore propose Asymmetric On-Policy Distillation (AOPD), which replaces ineffective negative reinforcement with localized divergence minimization in non-positive advantage regions while preserving positive reinforcement learning. Experiments on mathematical reasoning benchmarks show that AOPD consistently outperforms standard OPD, with average gains of 4.09 / 8.34 under strong / weak initialization, respectively. AOPD also maintains higher policy entropy during training and better capability retention during sequential tool-use adaptation.

Feed-forward 3D reconstruction models based on Vision Transformers can directly estimate scene geometry and camera poses from a small set of input images, but scaling them to video inputs with hundreds or thousands of frames remains challenging due to the quadratic cost of global attention layers. Recent token-merging methods accelerate these models by compressing the token sequence within the global attention layers, but they apply a uniform reduction to query tokens and key-value tokens, ignoring their functionally distinct roles in 3D reconstruction. In this work, we identify a key property of feed-forward 3D reconstruction models: query tokens encode view-specific geometric requests and are sensitive to compression, while key-value tokens represent shared scene context and tolerate aggressive compression. Guided by this insight, we propose Spark3R, a training-free acceleration framework that decouples the compression of query tokens and key-value tokens by assigning distinct reduction factors, with intra-group token merging applied to query tokens and lightweight token pruning to key-value tokens. Additionally, Spark3R adaptively adjusts the key-value reduction factor across layers, further improving the quality-efficiency trade-off. As a plug-and-play framework requiring no retraining, Spark3R integrates directly into multiple pretrained feed-forward 3D reconstruction models, including VGGT, $π^3$, Depth-Anything-3, and VGGT-$Ω$, and achieves up to $28\times$ speedup on 1,000-frame inputs while maintaining competitive reconstruction quality.

Agentic LLM workloads put bit-identical tokens at shifted positions every turn, voiding prefix caches at the first byte of divergence. Operators report cache-hit regressions ranging from moderate slowdowns to severe TTFT spikes of 10-16s on unchanged content. Prior position-independent caching systems correct RoPE on the full $d_K$-dimensional key, an architectural cost imposed by GQA, not by caching itself. Multi-Head Latent Attention, deployed at scale in DeepSeek-V2/V3/R1, Kimi-K2/Moonlight, GLM-5, and Mistral Large 3, factors each KV row into a position-free $c_{KV}$ and a 64-dim $k_r$ correctable in closed form; this structure motivates content-addressed caching as a natural fit rather than a GQA workaround. We present Irminsul, which extends SGLang's radix cache with content-hash keying over CDC-chunked segments and a $δ$-rotation rule for $k_r$. We evaluate three native MLA-MoE deployments - DeepSeek-V2-Lite (16B/2.4B), Kimi Moonlight-16B-A3B, and JoyAI-Flash (48B/3B) - with output-consistency on all three and recovery measured on the two endpoints; Irminsul recovers up to ~83% of prompt tokens above exact-prefix on agentic traffic while delivering 63% prefill energy savings per cache hit. We argue that content-addressed caching belongs in the serving stack as a first-class primitive, not a retrofit over prefix matching.

Post-training quantization makes large reasoning models practical under tight memory and latency budgets, but it can distort the online signals that drive adaptive test-time compute allocation. Under a fixed cap on the number of newly generated tokens, miscalibrated confidence can lead to harmful early halting: the model may surface a plausible final line while the underlying reasoning is still wrong, or the controller may stop before the trace has stabilized. We study this interaction for greedy 4-bit inference and propose BitCal-TTS, a lightweight runtime controller that combines (i) inexpensive online proxies for token-level uncertainty and reasoning-trace stability, (ii) a bit-conditioned confidence rescaling that is conservative at low nominal precision, and (iii) a bit-aware post-marker confirmation horizon designed for GSM8K-style structured outputs. The method requires no fine-tuning of the base model and integrates with standard Hugging Face 4-bit inference using forward hooks for logits and last-layer hidden states. On small evaluation shards of GSM8K with Qwen2.5 Instruct models, BitCal-TTS improves exact-match accuracy over a non-bit-aware adaptive baseline at the 7B and 14B scales while preserving substantial token savings relative to fixed-budget decoding. At a token cap of B=512, on the evaluation shards we report (N=54 for 7B and N=35 for 14B; not the full GSM8K test set), accuracy gains are +3.7 points (7B) and +2.8 points (14B), with the premature-stop rate falling from 14.8% to 11.1% on 7B and from 17.1% to 11.4% on 14B. We report Wilson 95% confidence intervals throughout and explicitly discuss the limited statistical power of the partial-shard comparisons. We release code and figure-generation scripts to support full reproduction.

Long-Text Understanding (LTU) at million-token scale requires balancing reasoning fidelity with computational efficiency. Frontier long-context LLMs can process millions of token contexts end-to-end, but they suffer from high token consumption and attention dilution. In parallel, specialized LTU agents often sacrifice fidelity through task-agnostic abstractions like graph construction or indexing. We identify a key insight for LTU: query-relevant information is typically sparse relative to the full document, so effective reasoning should rely on a query-sufficient subset rather than the entire context. To address this, we propose SCOUT, a new paradigm for LTU that shifts from passive processing to active information foraging. It treats the document as an explorable environment and answers from a compact, provenance-grounded epistemic state. Guided by state-level gap diagnosis, SCOUT adaptively alternates between coarse-to-fine exploration and anchored state updates that progressively contract its epistemic state toward query sufficiency. Experiments show that SCOUT matches state-of-the-art proprietary models while reducing token consumption by up to 8x. Moreover, SCOUT remains stable as context length scales, substantially alleviating the practical cost-performance trade-off.

This preprint presents a systematic, research-oriented practicum that guides the reader through the entire modern NLP pipeline: from tokenisation and vectorisation to fine-tuning of large language models, retrieval-augmented generation, and reinforcement learning from human feedback. A distinctive feature of the work is its consistent attention to low-resource and morphologically rich languages -- original contributions on Tajik and Tatar, including subword tokenisers, word embeddings, lexical databases, and transliteration benchmarks, are woven throughout the twelve sessions, demonstrating how modern NLP can be adapted to data-scarce environments without sacrificing rigour. Each session combines concise theory with detailed implementation plans, formalised evaluation metrics, and transparent assessment criteria. The work is not a conventional textbook: it is designed as a reproducible research artefact where every session requires publishing code, models, and reports in public repositories. All experiments are conducted on a single evolving corpus, and the work advocates open-weight models over commercial APIs, with special attention to the Hugging Face ecosystem. Designed for senior undergraduates, graduate students, and practising developers seeking to implement, compare, and deploy methods from classical ML to state-of-the-art LLM-based systems.

The performance of end-to-end automatic speech recognition (ASR) systems enables their increasing integration into numerous applications. While there are various benefits to such speech-to-text systems, the choice of hyperparameters and models plays a crucial role in their performance. Typically, these choices are determined by considering only the character (CER) and/or word error rate (WER) metrics. However, it has been shown in several studies that these metrics are largely incomplete and fail to adequately describe the downstream application of automatic transcripts. In this paper, we conduct a qualitative study on the French language that investigates the impact of subword tokenization algorithms and self-supervised learning models from different linguistic and acoustic perspectives, using a comprehensive set of evaluation metrics.

In single-stream autoregressive interfaces, the same tokens both update the model state and constitute an irreversible public commitment. This coupling creates a silence tax: additional deliberation postpones the first task-relevant content, while naive early streaming risks premature commitments that bias subsequent generations. We introduce Side-by-Side (SxS) Interleaved Reasoning, which makes disclosure timing a controllable decision within standard autoregressive generation. SxS interleaves partial disclosures with continued private reasoning in the same context, but releases content only when it is supported by the reasoning so far. To learn such pacing without incentivizing filler, we construct entailment-aligned interleaved trajectories by matching answer prefixes to supporting reasoning prefixes, then train with SFT to acquire the dual-action semantics and RL to recover reasoning performance under the new format. Across two Qwen3 architectures/scales (MoE Qwen3-30B-A3B, dense Qwen3-4B) and both in-domain (AIME25) and out-of-domain (GPQA-Diamond) benchmarks, SxS improves accuracy--content-latency Pareto trade-offs under token-level proxies such as inter-update waiting.

In classification tasks, models may rely on confounding variables to achieve strong in-distribution performance, capturing spurious features that fail under distribution shift. This shortcut behavior leads to substantial degradation in out-of-distribution settings. Task arithmetic offers a potential solution by removing unwanted signals via subtraction of secondary model updates, but it typically requires full fine-tuning, which is computationally expensive. Prompt tuning provides a parameter-efficient alternative by adapting models through a small set of trainable virtual tokens. Task arithmetic on the resulting prompts presents an appealing alternative to operations on entire models, but the extent to which this approach can limit reliance on spurious features remains to be established. In this work, we study whether composing soft prompts through task arithmetic improves robustness to confounding shifts. We propose Hybrid Prompt Arithmetic (HyPA), which combines task prompts with linearized confounder prompts to counteract spurious correlations. Across multiple benchmarks, HyPA consistently improves the robustness-performance trade-off relative to prompt-arithmetic baselines under distribution shift. We further analyze how HyPA affects hidden representations and find evidence consistent with it mitigating confounding either by reducing the influence of confounder signals on predictions or by suppressing them in the representation. These results establish HyPA as a parameter-efficient and promising approach for improving robustness under confounding shifts in the evaluated setting.

The increasing use of token-based representations in language-driven applications has motivated wireless token communication, where tokens are treated as fundamental units for transmission. However, conventional communication systems overlook dependencies among tokens and allocate transmission resources uniformly, leading to inefficient use of limited wireless resources under channel impairments. In this paper, we propose a context-aware token communication framework that leverages a masked language model (MLM) as a shared contextual model between the transmitter (Tx) and receiver (Rx). At the Rx, we develop a context-aware token detection method that integrates channel likelihoods with MLM-based contextual priors under a Bayesian formulation, enabling robust token inference over noisy channels. At the Tx, we propose a context-aware token masking strategy that selectively omits tokens that can be reliably inferred at the Rx, allowing the available power budget to be concentrated on more informative tokens. These components are jointly designed through a shared MLM, establishing a unified Tx-Rx framework for efficient token transmission and detection. Simulation results demonstrate that the proposed framework significantly improves reconstruction performance compared to conventional and existing token communication schemes, achieving up to 1.77X and 1.63X performance gains on the Europarl corpus and WikiText-103 datasets, respectively.

Token communication has emerged as a promising framework for efficient wireless transmission by representing source data as compact semantic tokens. However, transmitting full semantic tokens still incurs considerable communication overhead. In this paper, we propose an evolving semantic token communication system with a parametric memory network over MIMO fading channels. Specifically, only an equal-length prefix of each semantic token is transmitted, which reduces transmission cost while preserving a consistent token structure for receiver-side recovery. At the receiver, a parametric memory network is introduced to reconstruct the missing suffix information from the received token prefixes, where semantic memory is stored implicitly in the network parameters. To realize this design, full semantic tokens are first organized into a codebook, and truncated tokens are paired with the codeword labels of their corresponding full tokens. Based on these token-label pairs, kNN-based teacher distributions are constructed to fine-tune a pretrained GPT-2-based recovery module, which learns to infer the codeword distribution of each incomplete token and recover the corresponding complete semantic token. In addition, an online evolution strategy is developed to periodically update the parametric memory network and the entire system using newly observed test samples, thereby improving adaptability under distribution shifts. Experimental results demonstrate that the proposed method consistently outperforms the existing evolving memory benchmark under different channel conditions and channel bandwidth ratios, with up to 1.09 dB PSNR improvement.

Large language models (LLMs) have achieved remarkable performance across diverse domains, yet their enormous computational and memory requirements hinder deployment in resource-constrained environments. Knowledge distillation offers a promising solution by transferring knowledge from a large teacher model to a smaller student model. However, existing distillation methods typically treat all tokens equally, ignoring the fact that different tokens contribute unequally to model decisions. This can lead to inefficient knowledge transfer and reduced learning effectiveness. To address this limitation, we propose an entropy-based adaptive distillation strategy that dynamically adjusts the training process at the token level. Our method leverages the teacher's output entropy to guide three aspects of distillation. Specifically, we introduce a token-level curriculum by dynamically shifting focus from low- to high-entropy tokens during training. We further adjust the distillation temperature based on token entropy to better capture teacher confidence patterns. Moreover, we employ a dual-branch architecture for efficient logits-only distillation on easy tokens and deeper feature-based distillation on difficult tokens. Extensive experiments validate the soundness and effectiveness of our method.

Reinforcement Learning from Human Feedback (RLHF) has become a key post-training paradigm for improving model quality. However, the synchronous three-stage RLHF pipeline is often bottlenecked by the generation stage, where response-length skew causes the effective batch size to shrink rapidly during decoding, leaving GPUs underutilized while a few long responses remain unfinished. Mainstream frameworks employ a static tensor parallelism (TP) configuration that cannot adapt to changing batch characteristics, leaving substantial performance headroom unexplored. We propose PAT, an adaptive TP method that dynamically reconfigures TP during the generation stage of each RLHF iteration. PAT introduces two key techniques. First, a predictor-guided online reconfiguration method decides both the reconfiguration point and the target TP configuration based on offline profiling, triggering reconfiguration only when the predicted latency benefit outweighs the reconfiguration overhead. Second, a lightweight online reconfiguration mechanism updates only the states and layouts affected by TP changes: it adapts unfinished decoding states through a cost-model-based choice between KV-cache migration and recomputation, performs in-place weight resharding, and reuses cached communication groups. We implement PAT on top of SGLang and integrate it with the VeRL framework. Evaluations on LLaMA3.1-8B and Qwen3-14B using DeepScaleR show that PAT reduces generation latency by up to 34.6% and end-to-end RLHF training iteration latency by up to 27.2% compared to the original VeRL setup.

Time-series generative models often lack control over temporal granularity, forcing users to accept whatever granularity the model produces. To enable truly user-driven generation, we introduce TimeTok, a unified framework for Granularity-Controllable Time-Series Generation (GC-TSG), which generates time series at any target granularity from any coarser input (e.g., rough sketches) or from scratch. At the core of TimeTok is a hierarchical tokenization strategy that maps time series into an ordered sequence of tokens, from coarse to fine temporal granularity. Our autoregressive generation process operates across these granularity levels, producing token blocks that are decoded back into continuous time series. This design naturally enables GC-TSG - including standard generation - within a single framework, where controlling the number of token blocks provides explicit control over output detail. Experiments show that TimeTok excels at GC-TSG tasks while achieving state-of-the-art performance in standard generation. Furthermore, we showcase TimeTok's potential as a foundational tokenizer by training on multiple datasets with heterogeneous temporal granularities, verifying strong transferability that consistently outperforms models trained on individual datasets. To our knowledge, this is the first unified framework that covers the full generative spectrum for time series, offering a valuable foundation for models that benefit from diverse temporal granularities.

Scaling laws enable the optimal selection of data amount and language model size, yet the impact of the data unit, the token, on this relationship remains underexplored. In this work, we systematically investigate how the information granularity of tokens, controlled by the compression rate (i.e., average bytes of text per token), affects scaling trends. We train 988 latent tokenized models (BLT) ranging from 50M to 7B parameters that enable setting the desired compression rate. This flexibility allows us to study the role of compression rate well beyond 4.57 bytes per token obtained with a popular BPE tokenizer. Our experiments reveal that in compute-optimal configurations, model parameter counts scale proportionally to data size measured in bytes, not in tokens as commonly perceived (Kaplan et al., 2020; Hoffmann et al., 2022). Furthermore, we discover that the optimal compression rate differs from the one obtained with BPE and decreases with compute. These findings generalize to both latent and subword tokenization, as well as to languages other than English, guiding language model developers on tokenization scheme selection for maximal compute efficiency.

Character-level perturbations bypass safety alignment in modern LLMs despite leaving prompts human-readable. We identify and test a central structural mechanism: BPE tokenization fragments safety-critical words into sub-word pieces, and the three public alignment datasets we surveyed contain no intentionally fragmented inputs. The mechanism is a chain, tested end-to-end on five model families (Qwen-3-4B, Qwen-2.5-7B, Gemma-3-4B, Llama-3.1-8B, Mistral-7B). An optimization targeting safety-token fragmentation flips the first-token refusal trigger on 80-100% of refused HarmBench prompts, with 48% of those flips producing genuinely harmful outputs (per-model 29-65%; gap-vs-behavior ROC-AUC 0.66-0.98, pooled 0.84). Activation patching localizes the disrupted signal to the last ${\sim}30\%$ of layers; an alignment-data scan finds zero fragmented prompts among 30,000 examples (positive-control recall $\geq 99\%$ at attack-relevant intensities); and targeted-mutation experiments isolate safety words as the disruption locus. On the defense side, a 68-cell grid (55 trained checkpoints) shows that no DPO configuration achieves seed- and pool-stable ASR closure on the three families with closed pool-size confounds. SFT trained on fragmented prompts closes ASR on 3/5 families but only via global collapse that raises refusal on benign prompts as well, indicating the missing distribution is necessary but not sufficient under the LoRA-16 recipe we tested. To distinguish selective repair from global collapse, we introduce Conv-Benign, a candidate paired diagnostic. All ASR claims are 3-judge-calibrated (cell rankings stable across judges; absolute levels $\pm$18pp; see App.~B.13).

Autoregressive image modeling relies on visual tokenizers to compress images into compact latent representations. We design an end-to-end training pipeline that jointly optimizes reconstruction and generation, enabling direct supervision from generation results to the tokenizer. This contrasts with prior two-stage approaches that train tokenizers and generative models separately. We further investigate leveraging vision foundation models to improve 1D tokenizers for autoregressive modeling. Our autoregressive generative model achieves strong empirical results, including a state-of-the-art FID score of 1.48 without guidance on ImageNet 256x256 generation.

Machine unlearning has emerged as a critical capability for addressing privacy, safety, and regulatory concerns in large language models (LLMs). Existing methods operate at the sequence level, applying uniform updates across all tokens despite only a subset encoding the knowledge targeted for removal. This introduces gradient noise, degrades utility, and leads to suboptimal forgetting. We propose TokenUnlearn, a token-level attribution framework that identifies and selectively targets critical tokens. Our approach combines knowledge-aware signals via masking, and entropy-aware signals to yield importance scores for precise token selection. We develop two complementary strategies: hard selection, applying unlearning only to high-importance tokens, and soft weighting, modulating gradient contributions based on importance scores. Both extend existing methods to token-level variants. Theoretical analysis shows token-level selection improves gradient signal-to-noise ratio. Experiments on TOFU and WMDP benchmarks across three model architectures demonstrate consistent improvements over sequence-level baselines in both forgetting effectiveness and utility preservation.

Single-prompt first-token probabilities from zero-shot vision-language model (VLM) safety classifiers are treated as decision scores, but we show they are unreliable under semantically equivalent prompt reformulation: even when the binary label is constrained to a fixed output position, equivalent prompts can induce materially different unsafe probabilities for the same sample. Across multimodal safety benchmarks and multiple VLM families, cross-prompt variance is strongly associated with prompt-level disagreement and higher error, making it a useful fragility diagnostic. A training-free mean ensemble improves NLL on all 14 dataset-model evaluation pairs and ECE on 12/14 relative to a train-selected single-prompt baseline, and wins more head-to-head NLL comparisons than labeled temperature scaling, Platt scaling, and isotonic regression applied to the same prompt. Ranking gains are consistent against the train-selected baseline on both AUROC and AUPRC, and against the full 15-prompt distribution remain consistent on AUPRC while softening on AUROC. Labeled calibration on top of the mean provides further gains when labels are available, identifying prompt averaging as a strong label-free first stage rather than a replacement for calibration. We frame this as a reliability stress test for zero-shot VLM first-token safety scores and recommend prompt-family evaluation with mean aggregation as a standard label-free reliability baseline.

Public inference benchmarks compare AI systems at the model and provider level, but the unit at which deployment decisions are actually made is the endpoint: the (provider, model, stock-keeping-unit) tuple at which a specific quantization, decoding strategy, region, and serving stack is exposed. We introduce TokenArena, a continuous benchmark that measures inference at endpoint granularity along five core axes (output speed, time to first token, workload-blended price, effective context, and quality on the live endpoint) and synthesizes them, together with a modeled energy estimate, into three headline composites: joules per correct answer, dollars per correct answer, and endpoint fidelity (output-distribution similarity to a first-party reference). The framework's novelty is empirical and methodological. Across 78 endpoints serving 12 model families, the same model on different endpoints differs in mean accuracy by up to 12.5 points on math and code, in fingerprint similarity to first party by up to 12 points, in tail latency by an order of magnitude, and in modeled joules per correct answer by a factor of 6.2. We further show that workload-aware blended pricing reorders the leaderboard substantially: 7 of 10 top-ranked endpoints under the chat preset (3:1 input:output) fall out of the top 10 under the retrieval-augmented preset (20:1), and the reasoning preset (1:5) elevates frontier closed models that the chat preset penalizes on price. We release the framework, schema, probe and eval harness, and a v1.0 leaderboard snapshot under CC BY 4.0. TokenArena is a methodology, not a single ranking; we publish full provenance and limitations and welcome external replication.

Mamba's recurrent state h_t is, by construction, a compressed summary of every token seen so far. This raises a tempting hypothesis: if we extract token-level outputs y_t at fixed patch boundaries, we obtain semantic sentence summaries for free, with no pooling head, no fine-tuning, and no [CLS] token. We test this hypothesis carefully. Across five benchmarks (SST-2, CoLA, MRPC, STS-B, IMDb), we compare four strategies for extracting frozen sentence representations from a pretrained Mamba-130M backbone under a strict frozen-feature probing protocol, using three random seeds where computationally feasible. The results do not support the hypothesis: patch boundary readouts do not consistently outperform simple mean pooling. We identify and quantify two structural pathologies: severe anisotropy (mean pairwise cosine similarity 0.9999, std 0.000044) and representational collapse in the raw final SSM state (MCC = 0.000 on CoLA across all three seeds, confirmed via confusion matrix). We further propose orthogonal injection, a modified recurrence that constrains new information per

Understanding how Large Language Models (LLMs) make token-level decisions during code generation remains a major challenge for both researchers and practitioners. While recent tools provide insights into model internals or generation outcomes, they often lack decoding-time signals, fine-grained uncertainty measures, and interactive mechanisms for exploring alternative generation paths. We present TokenScope, an interactive interpretability and analysis tool for decoder-based LLMs that exposes token-level metrics, attention patterns, and structural information during generation. TokenScope supports interactive token replacement, counterfactual branching, and code-aware aggregation via abstract syntax trees. By unifying decoding-time signals with structural program analysis, TokenScope enables systematic investigation of LLM behaviour during code generation.

Local fine-tuning datasets routinely contain sensitive secrets such as API keys, personal identifiers, and financial records. Although "local offline fine-tuning" is often viewed as a privacy boundary, we reveal that compromised model code is sufficient to steal them. Current passive pretrained-weight poisoning attacks, while effective for natural language, fundamentally fail to capture such sparse high-entropy targets due to their reliance on probabilistic semantic prefixes. To bridge this gap, we identify and exploit a practical but overlooked supply-chain vector -- malicious model code camouflaged as standard architectural definitions to realize a paradigm shift from passive weight poisoning to active execution hijacking. We introduce a deterministic full-chain memorization mechanism: it locks onto token-level secrets in dynamic computation flows via online tensor-rule matching, and leverages value-gradient decoupling to stealthily inject attack gradients, overcoming gradient drowning to force model memorization. Furthermore, we achieve, for the first time, attacker-verifiable secret stealing through black-box queries that precisely distinguishes true leakage from hallucination. Our attack achieves over 98% Strict ASR in the default LoRA setting with limited primary-task utility degradation and effectively evades defense measures including semantic safety filtering, code auditing, and perplexity-based detection.

Subword tokenization is an essential part of modern large language models (LLMs), yet its specific contributions to training efficiency and model performance remain poorly understood. In this work, we decouple the effects of subword tokenization by isolating them within a controlled byte-level pretraining pipeline. We formulate and test hypotheses across various dimensions, including sample throughput, vocabulary scaling, and the linguistic prior of subword boundaries. By simulating these effects in a byte-level setting, we refine our understanding of why subword models outperform raw byte models and offer insights to improve the pretraining of future byte-level and subword models. Specifically, our experiments highlight the critical role of increased training throughput and the integration of subword boundaries as either explicit priors or inductive biases.

Token serves as the fundamental unit of computation in modern autoregressive models, and generation length directly influences both inference cost and reasoning performance. Despite its importance, existing approaches lack fine-grained length modeling, operating primarily at the coarse-grained sequence level. We introduce the Length Value Model (LenVM), a token-level framework that models the remaining generation length. By formulating length modeling as a value estimation problem and assigning a constant negative reward to each generated token, LenVM predicts a bounded, discounted return that serves as a monotone proxy for the remaining generation horizon. This formulation yields supervision that is annotation-free, dense, unbiased, and scalable. Experiments on LLMs and VLMs demonstrate LenVM provides a highly effective signal at inference time. On the LIFEBench exact length matching task, applying LenVM to a 7B model improves the length score from 30.9 to 64.8, significantly outperforming frontier closed-source models. Furthermore, LenVM enables continuous control over the trade off between performance and efficiency. On GSM8K at a budget of 200 tokens, LenVM maintains 63% accuracy compared to 6 percent for token budget baseline. It also accurately predicts total generation length from the prompt boundary. Finally, LenVM's token-level values offer an interpretable view of generation dynamics, revealing how specific tokens shift reasoning toward shorter or longer regimes. Results demonstrate that LenVM supports a broad range of applications and token length can be effectively modeled as a token-level value signal, highlighting the potential of LenVM as a general framework for length modeling and as a length-specific value signal that could support future RL training. Code is available at https://github.com/eric-ai-lab/Length-Value-Model.

Federated fine-tuning provides a practical route to adapt large language models (LLMs) on edge devices without centralizing private data. However, in mobile deployments, the training wall-clock is often dominated by straggler-limited uplink communication under heterogeneous bandwidth, intermittent participation, and non-IID client data. Although parameter-efficient fine-tuning (PEFT) methods such as LoRA and QLoRA reduce local memory and trainable parameters, repeated transmission of adapter updates remains a major bottleneck. We propose Fed-FSTQ, a semantic-sensitivity-aware communication-control primitive for communication-efficient federated LLM fine-tuning. Fed-FSTQ uses a lightweight token-level Fisher proxy to estimate semantic sensitivity, couples token-guided sparsification with mixed-precision adapter-update quantization, and allocates higher communication fidelity to semantically load-bearing evidence while suppressing redundant transmission. The method is drop-in compatible with standard federated PEFT pipelines and requires no change to the server aggregation rule. Experiments on multilingual QA and medical QA under non-IID partitions show that Fed-FSTQ reduces cumulative uplink traffic required to reach a fixed quality threshold by 46-fold relative to a Fed-LoRA baseline and improves straggler-limited wall-clock time-to-accuracy by 52%. Under the corrected Controlled LTE-20Mbps accounting, Fed-FSTQ reduces per-round time from 414.60s to 67.29s and reduces per-round energy from 839.20J to 146.28J, yielding a 6.16-fold speedup. On NVIDIA Jetson-class edge devices, Fisher-guided token reduction also yields up to a 1.55-fold inference speedup, demonstrating deployability under tight resource constraints.

We present PolyKV, a system in which multiple concurrent inference agents share a single, asymmetrically compressed KV cache pool. Rather than allocating a separate KV cache per agent -- the standard paradigm -- PolyKV writes a compressed cache once and injects it into N independent agent contexts via HuggingFace DynamicCache objects. Compression is asymmetric: Keys are quantized at int8 (q8_0) to preserve softmax stability, while Values are compressed using TurboQuant MSE -- a Fast Walsh-Hadamard Transform (FWHT) rotation followed by 3-bit Lloyd-Max quantization with centroids tuned to N(0,1). We evaluate across two model scales (SmolLM2-1.7B-Instruct and Llama-3-8B-Instruct), three context lengths (600-7,194 tokens), and up to 15 concurrent agents. PolyKV achieves a stable 2.91x compression ratio across all configurations. On Llama-3-8B with 15 agents sharing a 4K-token context, PolyKV reduces KV cache memory from 19.8 GB to 0.45 GB -- a 97.7% reduction -- while maintaining only +0.57% perplexity degradation and a mean BERTScore F1 of 0.928. PPL delta does not grow with agent count and improves as context length increases, inverting to -0.26% at 1,851 coherent tokens. To our knowledge, no prior work combines a single shared, lossy-compressed KV pool with multi-reader concurrent agent access.

We introduce an efficient, resolution-agnostic autoregressive (AR) image synthesis approach that generalizes to arbitrary resolutions and aspect ratios, narrowing the gap to diffusion models at scale. At its core is VibeToken, a novel resolution-agnostic 1D Transformer-based image tokenizer that encodes images into a dynamic, user-controllable sequence of 32-256 tokens, achieving a state-of-the-art efficiency and performance trade-off. Building on VibeToken, we present VibeToken-Gen, a class-conditioned AR generator with out-of-the-box support for arbitrary resolutions while requiring significantly fewer compute resources. Notably, VibeToken-Gen synthesizes 1024x1024 images using only 64 tokens and achieves 3.94 gFID; by comparison, a diffusion-based state-of-the-art alternative requires 1,024 tokens and attains 5.87 gFID. In contrast to fixed-resolution AR models such as LlamaGen -- whose inference FLOPs grow quadratically with resolution (11T FLOPs at 1024x1024) -- VibeToken-Gen maintains a constant 179G FLOPs (63.4x efficient) independent of resolution. We hope VibeToken can help unlock the wide adoption of AR visual generative models in production use cases.

Discrete diffusion language models (dLLMs) enable parallel token updates with bidirectional attention, yet practical generation typically adopts blockwise semi-autoregressive decoding. This switch creates a training-inference mismatch: training denoises with full-sequence context, while inference commits tokens within a bounded block without future context. Therefore, decoding with fixed-size or heuristic-based blocks can lead to premature token commitments, as decisions are made without full access to future context that could alter those choices. Motivated by this, we propose self-containedness as a principled criterion for block commitment. A block is self-contained if its predictions remain consistent with Future-Aware (FA) or without No-Future (NF) access to future context, reframing block boundary selection as a test of self-containedness rather than a heuristic choice. Based on this principle, we introduce Variable-size Self-contained Blocks (VSB) for dLLMs. VSB scores and selects block boundaries using the divergence between token-level predictive distributions under NF and FA conditioning, which quantifies how predictions would change if future context were revealed. We provide theoretical justification linking self-containedness to predictive consistency, and extensive experiments validate VSB's efficacy over fixed-size and heuristic blockwise decoding.

The Korean writing system, \textit{Hangeul}, has a unique character representation rigidly following the invention principles recorded in \textit{Hunminjeongeum}.\footnote{\textit{Hunminjeongeum} is a book published in 1446 that describes the principles of invention and usage of \textit{Hangeul}, devised by King Sejong \cite{Hunminjeongeum_Guide}.} However, existing pre-trained language models (PLMs) for Korean have overlooked these principles. In this paper, we introduce a novel framework for Korean PLMs called KOMBO, which firstly brings the invention principles of \textit{Hangeul} to represent character. Our proposed method, KOMBO, exhibits notable experimental proficiency across diverse NLP tasks. In particular, our method outperforms the state-of-the-art Korean PLM by an average of 2.11% in five Korean natural language understanding tasks. Furthermore, extensive experiments demonstrate that our proposed method is suitable for comprehending the linguistic features of the Korean language. Consequently, we shed light on the superiority of using subcharacters over the typical subword-based approach for Korean PLMs. Our code is available at: https://github.com/SungHo3268/KOMBO.

Chain-of-Thought (CoT) reasoning has emerged as a key technique for eliciting complex reasoning in Large Language Models (LLMs). Although interpretable, its dependence on natural language limits the model's expressive bandwidth. Continuous thought models address this bottleneck by reasoning in latent space rather than human-readable tokens. While they enable richer representations and faster inference, they raise a critical safety question: how can we detect misaligned reasoning in an uninterpretable latent space? To study this, we introduce MoralChain, a benchmark of 12,000 social scenarios with parallel moral/immoral reasoning paths. We train a continuous thought model with backdoor behavior using a novel dual-trigger paradigm - one trigger that arms misaligned latent reasoning ([T]) and another that releases harmful outputs ([O]). We demonstrate three findings: (1) continuous thought models can exhibit misaligned latent reasoning while producing aligned outputs, with aligned and misaligned reasoning occupying geometrically distinct regions of latent space; (2) linear probes trained on behaviorally-distinguishable conditions ([T][O] vs [O]) transfer to detecting armed-but-benign states ([T] vs baseline) with high accuracy; and (3) misalignment is encoded in early latent thinking tokens, suggesting safety monitoring for continuous thought models should target the "planning" phase of latent reasoning.

Whether intermediate reasoning is computationally useful or merely explanatory depends on whether chain-of-thought (CoT) tokens contain task-relevant information. We present a mechanistic causal analysis of CoT on GSM8K using activation patching: transferring token-level hidden states from a CoT generation to a direct-answer run for the same question, then measuring the effect on final-answer accuracy. Across models, generating after patching yields substantially higher accuracy than both direct-answer prompting and the original CoT trace, revealing that individual CoT tokens can encode sufficient information to recover the correct answer, even when the original trace is incorrect. This task-relevant information is more prevalent in correct than incorrect CoT runs and is unevenly distributed across tokens, concentrating in mid-to-late layers and appearing earlier in the reasoning trace. Moreover, patching language tokens such as verbs and entities carry task-solving information that steers generation toward correct reasoning, whereas mathematical tokens encode answer-proximal content that rarely succeeds. Patched outputs are often shorter and yet exceed the accuracy of a full CoT trace, suggesting complete reasoning chains are not always necessary. Together, these findings demonstrate that CoT encodes recoverable, token-level problem-solving information, offering new insight into how reasoning is represented and where it breaks down.

The wide adoption of AI agents in complex human workflows is driving rapid growth in LLM token consumption. When agents are deployed on tasks that require a significant amount of tokens, three questions naturally arise: (1) Where do AI agents spend the tokens? (2) Which models are more token-efficient? and (3) Can agents predict their token usage before task execution? In this paper, we present the first systematic study of token consumption patterns in agentic coding tasks. We analyze trajectories from eight frontier LLMs on SWE-bench Verified and evaluate models' ability to predict their own token costs before task execution. We find that: (1) agentic tasks are uniquely expensive, consuming 1000x more tokens than code reasoning and code chat, with input tokens rather than output tokens driving the overall cost; (2) token usage is highly variable and inherently stochastic: runs on the same task can differ by up to 30x in total tokens, and higher token usage does not translate into higher accuracy; instead, accuracy often peaks at intermediate cost and saturates at higher costs; (3) models vary substantially in token efficiency: on the same tasks, Kimi-K2 and Claude-Sonnet-4.5, on average, consume over 1.5 million more tokens than GPT-5; (4) task difficulty rated by human experts only weakly aligns with actual token costs, revealing a fundamental gap between human-perceived complexity and the computational effort agents actually expend; and (5) frontier models fail to accurately predict their own token usage (with weak-to-moderate correlations, up to 0.39) and systematically underestimate real token costs. Our study offers new insights into the economics of AI agents and can inspire future research in this direction.

Watermarking has emerged as a promising technique for tracing the authorship of content generated by large language models (LLMs). Among existing approaches, the KGW scheme is particularly attractive due to its versatility, efficiency, and effectiveness in natural language generation. However, KGW's effectiveness degrades significantly under low-entropy settings such as code generation and mathematical reasoning. A crucial step in the KGW method is random vocabulary partitioning, which enables adjustments to token selection based on specific preferences. Our study revealed that the next-token probability distribution plays an critical role in determining how much, or even whether, we can modify token selection and, consequently, the effectiveness of watermarking. We refer to this characteristic, associated with the probability distribution of each token prediction, as \emph{watermark strength.} In cases of random vocabulary partitioning, the lower bound of watermark strength is dictated by the next-token probability distribution. However, we found that, by redesigning the vocabulary partitioning algorithm, we can potentially raise this lower bound. In this paper, we propose SSG (\textbf{S}ort-then-\textbf{S}plit by \textbf{G}roups), a method that partitions the vocabulary into two logit-balanced subsets. This design lifts the lower bound of watermark strength for each token prediction, thereby improving watermark detectability. Experiments on code generation and mathematical reasoning datasets demonstrate the effectiveness of SSG.

Traditional data valuation methods based on ``row-count $\times$ quality coefficient'' paradigms fail to capture the nuanced, nonlinear contributions that data makes to Large Language Model (LLM) capabilities. This paper presents a dynamic data valuation framework that transitions from static accounting to utility-based pricing. Our approach operates on three layers: (1) token-level information density metrics using Shannon entropy and Data Quality Scores; (2) empirical training gain measurement through influence functions, proxy model strategies, and Data Shapley values; and (3) cryptographic verifiability through hash-based commitments, Merkle trees, and a tamper-evident training ledger. We provide comprehensive experimental validation on three real domains (instruction following, mathematical reasoning, and code summarization), demonstrating that proxy-based empirical gain achieves near-perfect ranking alignment with realized utility, substantially outperforming row-count and token-count baselines. This framework enables a fair Data-as-a-Service economy where high-reasoning data is priced according to its actual contribution to model intelligence, while providing the transparency and auditability necessary for trustworthy data markets.

Large language models (LLMs) often produce content that contradicts or overlooks information provided in the input context, a phenomenon known as faithfulness hallucination. In this paper, we propose Context-Fidelity Boosting (CFB), a lightweight and general decoding-time framework that reduces such hallucinations by increasing the generation probability of source-supported tokens. Motivated by logit-shaping principles from watermarking techniques, CFB applies additive token-level logit adjustments based on a token's degree of support from the input context. Specifically, we develop three boosting strategies: static boosting, which applies a fixed bias to source-supported tokens; context-aware boosting, which scales this bias using the divergence between next-token distributions with and without context; and token-aware boosting, which further redistributes the adaptive bias according to local relevance estimated from source-position attention and source-scoped semantic similarity. CFB requires no retraining or architectural changes, making it compatible with a wide range of LLMs. Experiments on summarization and question answering tasks across multiple open-source LLMs show that CFB consistently improves faithfulness metrics with minimal generation overhead. Our implementation is fully open-sourced.

We investigate the ability of decoder-only transformer models to perform abstract symbolic reasoning; specifically solving propositional logic reasoning problems given in-context. Previous work demonstrated that models fail to generalize to problems involving variable names that were not observed during training, and it was shown that one reason behind this is the difficulty of copying (or generating) unseen tokens. We show both theoretically and empirically that a particular representational collapse also has a crucial role: the unembeddings (last-layer weights) of unseen tokens collapse to nearly the same vector during training. The collapse makes distinguishing multiple unseen variables difficult for the model (especially when the embedding and unembedding parameters are shared), and provides a mechanistic explanation for the effectiveness of existing heuristic interventions like "active forgetting", which periodically reset the token (un)embeddings. Based on these observations, we devise a combination of techniques, involving a small architecture change facilitating copying, data diversity, and freezing or resetting (un)embeddings, that achieves generalization to unseen tokens. We support our claims with extensive controlled experiments on propositional logic reasoning problems. Beyond synthetic experiments, we also observe evidence of (un)embedding collapse in the open-weight models in the Gemma 3 family, which includes 99 unused tokens reserved for downstream use. Empirically we find that the correlated embeddings of these tokens are a poor initialization for finetuning applications.

Language models must bind entities to the attributes they possess and maintain several such binding relationships within a context. We study how multiple entities are represented across token positions and whether single tokens can carry bindings for more than one entity. We introduce a multi-slot probing approach that disentangles a single token's residual stream activation to recover information about both the currently described entity and the immediately preceding one. These two kinds of information are encoded in separate and largely orthogonal "current-entity" and "prior-entity" slots. We analyze the functional roles of these slots and find that they serve different purposes. In tandem with the current-entity slot, the prior-entity slot supports relational inferences, such as entity-level induction ("who came after Alice in the story?") and conflict detection between adjacent entities. However, only the current-entity slot is used for explicit factual retrieval questions ("Is anyone in the story tall?" "What is the tall entity's name?") despite these answers being linearly decodable from the prior-entity slot too. Consistent with this limitation, open-weight models perform near chance accuracy at processing syntax that forces two subject-verb-object bindings on a single token (e.g., "Alice prepares and Bob consumes food.") Interestingly, recent frontier models can parse this properly, suggesting they may have developed more sophisticated binding strategies. Overall, our results expose a gap between information that is available in activations and information the model actually uses, and suggest that the current/prior-entity slot structure is a natural substrate for behaviors that require holding two perspectives at once, such as sycophancy and deception.

Transformer-based OCR models have shown strong performance on Latin and CJK scripts, but their application to African syllabic writing systems remains limited. We present the first adaptation of TrOCR for printed Tigrinya using the Ge'ez script. Starting from a pre-trained model, we extend the byte-level BPE tokenizer to cover 230 Ge'ez characters and introduce Word-Aware Loss Weighting to resolve systematic word-boundary failures that arise when applying Latin-centric BPE conventions to a new script. The unmodified model produces no usable output on Ge'ez text. After adaptation, the TrOCR-Printed variant achieves 0.22% Character Error Rate and 97.20% exact match accuracy on a held-out test set of 5,000 synthetic images from the GLOCR dataset. An ablation study confirms that Word-Aware Loss Weighting is the critical component, reducing CER by two orders of magnitude compared to vocabulary extension alone. The full pipeline trains in under three hours on a single 8 GB consumer GPU. All code, model weights, and evaluation scripts are publicly released.

We present LLaDA2.0-Uni, a unified discrete diffusion large language model (dLLM) that supports multimodal understanding and generation within a natively integrated framework. Its architecture combines a fully semantic discrete tokenizer, a MoE-based dLLM backbone, and a diffusion decoder. By discretizing continuous visual inputs via SigLIP-VQ, the model enables block-level masked diffusion for both text and vision inputs within the backbone, while the decoder reconstructs visual tokens into high-fidelity images. Inference efficiency is enhanced beyond parallel decoding through prefix-aware optimizations in the backbone and few-step distillation in the decoder. Supported by carefully curated large-scale data and a tailored multi-stage training pipeline, LLaDA2.0-Uni matches specialized VLMs in multimodal understanding while delivering strong performance in image generation and editing. Its native support for interleaved generation and reasoning establishes a promising and scalable paradigm for next-generation unified foundation models. Codes and models are available at https://github.com/inclusionAI/LLaDA2.0-Uni.

Supervised Fine-Tuning (SFT) is the predominant paradigm for aligning large language models (LLMs), yet it suffers from optimization instability and limited generalization. Recent work attributes this issue to pathological gradient scaling and proposes Dynamic Fine-Tuning (DFT) to correct it at the token level. However, DFT assumes all demonstrations are equally suitable learning targets, an assumption violated by the strong heterogeneity of large-scale instruction data, where demonstration-policy mismatch induces high-variance updates at the sample level. We introduce Compatibility-Aware Dynamic Fine-Tuning (CADFT), a principled extension of DFT that controls sample-level optimization variance. CADFT derives a dynamic, policy-dependent compatibility signal from model likelihoods to modulate supervised updates, suppressing high-variance gradients from incompatible demonstrations. We further propose a delayed, low-frequency compatibility-guided rewriting strategy to transform persistently incompatible demonstrations into learnable targets. We show that CADFT can be interpreted as a variance-controlled estimator that generalizes token-level stabilization in DFT to the sample level. Extensive experiments demonstrate improved stability, generalization, and cold-start reinforcement learning initialization, while remaining fully supervised and independent of explicit reward modeling.