Retrieval-augmented generation systems can precompute and store key-value caches of retrieved documents to avoid re-encoding context at every query. Quantizing these caches further reduces storage, but no prior work asks whether compression damages faithfulness, whether responses remain grounded in the retrieved evidence. Faithfulness and accuracy are not equivalent: a model can produce a correct answer that is no longer supported by the context it was given. We evaluate Qwen2.5-7B-Instruct under INT8 and INT4 quantization on RGB and HotpotQA, measuring both accuracy and faithfulness with a hallucination detector, NLI entailment, and an LLM judge. INT8 is near-lossless across both metrics. INT4 reduces accuracy and, more critically, even among answers that remain factually correct, over 90% of faithfulness changes are negative, i.e., accuracy metrics are blind to this regression. The harm grows under noisy retrieval and with more retrieved chunks. Faithfulness must be audited before compressed caches are deployed.
Personality is increasingly important in large language models (LLMs), as it shapes users' trust, engagement, and emotional experiences. While the Myers--Briggs Type Indicator (MBTI) has emerged as a common framework for assessing LLMs' personality, existing studies focus primarily on full-precision models and evaluate only final outputs. They overlook the widespread deployment of quantized LLMs requiring low memory footprints, whose personality traits remain underexplored. In this work, we present a systematic MBTI analysis of open-source LLMs across multiple precisions, including mainstream 4-bit methods (GPTQ, AWQ) and extreme 2-bit settings (AQLM variants). Beyond output-level evaluation, we examine how personality emerges across layers through option-level entropy and confidence-gap dynamics, and introduce Uncertainty-Amplified Layer Decoding (UALD) to study decoding-induced personality drift at inference time. Our results reveal a key insight: LLMs' personality is not a static property, but an emergent, layer-dependent decision process sensitive to quantization, prompting, and decoding. Specifically, we find that (1) ENFJ remains dominant across model families and precisions; (2) 4-bit quantization largely preserves coarse personality structure, while 2-bit quantization disrupts fine-grained prompt consistency and cross-precision agreement; (3) personality decisions emerges in upper layers, following substantial ambiguity in early layers; and (4) inference decoding can shift personality, while personality-aligned conditioning improves robustness. These findings provide a new perspective on the behavioral reliability of quantized LLMs and highlight the importance of considering internal dynamics and inference strategies in personality-sensitive chatbot applications.
Ismail Hossain, Nafi Ullah Shafin, Mohammad Abdullah Al Mumincs.CL
Post-training quantization lowers the memory footprint of Large Language Models (LLMs) and speeds up inference, which is why it is now common for on-device deployment. Most of what we know about its effects, however, comes from English benchmarks. It is not clear whether the same holds for morphologically complex, low-resource languages such as Bangla, and this gap is what we address here. We evaluate three model families---Qwen-2.5-7B, LLaMA-3.1-8B, and GPT-OSS-20B---in full precision and in three quantized formats (GPTQ-Int8, GPTQ-Q8, GGUF-W8A16) across five Bangla natural language understanding benchmarks (Bangla MMLU, CommonsenseQA-BN, OpenBookQA-BN, PIQA-BN, and BoolQ-BN), using zero-shot evaluation through lm-evaluation-harness. To our knowledge this is the first controlled comparison of quantization formats on Bangla NLU. The three families do not respond the same way: GPT-OSS loses up to 57.35% accuracy on reasoning-heavy tasks under GGUF-W8A16, while Qwen and LLaMA hold steady under GPTQ, and in a few cases the quantized version edges out the full-precision one. BoolQ-BN, a comprehension task, stays stable across all three families regardless of format. Taken together, these results suggest quantization can work well for Bangla deployment, but the choice of architecture and quantization method matters more than the bit width alone. We discuss what this means for practitioners choosing a model to run on constrained hardware.
A 4-bit quantized weight specifies a rounding cell rather than a single full-precision value. We introduce in-cell learning, a paradigm for writing new knowledge only within these cells, so that re-quantizing the served weights reproduces the released integer codes and scales exactly. CellFill implements this idea with bounded trainable positions inside frozen quantization cells and ships the update as a separate, subtractively revocable file. Across published NF4 and W4A16 releases of Qwen3 and Gemma from 1.7B to 32B parameters, CellFill writes 83-99% of a real-fact corpus while returning the stored code on every constrained weight. The injected facts generalize to paraphrases and composition, and answer 78-88% of selected PopQA questions that the released model misses. Sequential experiments show that rehearsal preserves earlier knowledge, whereas available room and new-task plasticity decline across updates. Consolidation re-quantizes the learned weights to produce a declared major version, restoring room at a measured capability cost. A six-task write-rehearse-consolidate cycle retains at least 92.8% of first learning in two 8B runs and records zero code violations over 6.9 billion constrained weights at every fold. These results define a version-management protocol in which minor updates preserve the released quantized artifact bitwise and major updates are explicit, measurable, and verifiable.
Aggressive quantization disproportionately harms multilingual capability: in the sub-4B INT3 GPTQ regime, we measure 2-4x larger perplexity degradation on non-English languages than on English. We propose Language-Conditional Dequantization (LCD), a post-hoc method that attaches per-language rank-2 LoRA corrections to the linear layers of an already-quantized model, adding 0.12% parameters per language and training in under 20 minutes on a single GPU. Across Qwen2.5-3B and Llama-3.2-3B, LCD recovers 70-83% of the perplexity gap for non-Latin script languages and 17-28% of the GlobalMMLU accuracy gap, outperforming a language-agnostic correction of equal capacity by 3-9 points on typologically distant languages and a data-free low-rank baseline (LQER) by an order of magnitude. We further identify a perplexity-accuracy disconnect and trace it to where quantization concentrates damage: early-depth errors (Llama) propagate downstream and resist local correction, while late-depth errors (Qwen) do not. A layer-restricted variant of LCD validates this mechanism directly.
Multilingual reasoning models are commonly evaluated by whether they arrive at the correct answer, but not by whether they preserve the intended language while reasoning and responding. This omission conceals important multilingual behaviors that emerge as tasks become harder. In this paper, we study task difficulty, task accuracy, thinking-language consistency (TC), and answer-language consistency (AC) across reasoning models using PolyMath benchmark in eight languages and four difficulty levels. We uncover four findings: (1) language consistency exhibits four difficulty-dependent behaviors: output-language consistency remains aligned with input, remains misaligned, degrades gradually, or collapses abruptly. (2) We identify the language consistency breakdown effect, where increasing difficulty can cause a sudden drop in output-language consistency, especially in less strongly represented and non-Latin-script languages. (3) Due to this breakdown effect, accuracy can be preserved or even improved at a harder difficulty level as the model shifts to its internal dominant language. (4) Quantization can improve or degrade output-language consistency independently of its effect on accuracy, with GPTQ and AWQ often outperforming AutoRound under tolerance-based voting with ε = 1.0. These results show that multilingual capability cannot be characterized by accuracy alone; reliable evaluation should jointly consider task accuracy, language consistency, and task difficulty for multilingual benchmarks.
Quantized orthogonal fine-tuning (qoft) enables parameter-efficient adaptation of low-bit language models by learning structured activation rotations before frozen quantized weights. However, its task-specific updates remain constrained to linear orthogonal transformations, limiting input-dependent nonlinear corrections. We introduce AuroOFT, which keeps qoft as a stable quantization-compatible branch while attaching a zero-start gated low-rank nonlinear residual to each adapted linear layer. AuroOFT maps activations into an RMS-normalized compact latent space and uses adaptive nonlinear bases with bounded or token-dependent gating. The zero-initialized up projection makes AuroOFT functionally identical to qoft at initialization, while orthogonality remains a branch-level stability property rather than a property of the combined nonlinear layer. Under matched data, optimization, decoding, and parser protocols, AuroOFT improves Macro-6 over matched qoft by 1.30-2.70% on the 1.5B/3B Qwen2.5 settings, exceeds QLoRA by 6.52-10.62%, and saves 32.3-44.7% trainable parameters relative to QLoRA in representative scales. The small exam-style multiple-choice math set is treated only as a protocol-sensitivity diagnostic. Our code is available at the anonymous repository: https://anonymous.4open.science/r/AuroOFT-F3FD.
We introduce OptGear, a foundation model designed for efficient on-device deployment, real-tim inference, and strong task capability. It includes a dense model (1M, 270M, and 1B) with a context length of 64K. We designed a new hybrid architecture that combines a convolutional key-value gated mixer with local-global attention to reduce the KV-cache memory that tends to increase exponentially with long context. This architecture delivers up to X4.9 faster prefill and decoding speeds on the NPUs compared to models of a similar scale models. From a 2T tokens candidate corpus, OptGear is trained on a curated 0.5T tokens subset without knowledge distillation. This is the most data-efficient of the existing foundation models. All models are released with open weights and deployment binaries for ONNX, Qualcomm NPU, and Apple ANE making OptGear a practical base for edge applications that need fast, memory-efficient inference and strong task capabilities. Furthermore, to expand the ecosystem of on-device generative language models, we are introducing the OptGear-1M that can be deployed on Micro-Controller Units (MCUs), a Tiny Language Model (TLM). OptGear-1M is the first generative language model to achieve 20 TPS with W4A32 quantization on the ARM Cortex-M7 CPU of the STM32H747I-DISCO.
Large Language Models (LLMs) deployed in dynamic financial environments face a critical challenge: maintaining factual accuracy as market conditions, regulations, and corporate facts change continuously. While 4-bit quantization enables efficient deployment, it severely limits the viability of sequential memory editing: existing methods undergo catastrophic performance degradation under this "quantization stability crisis." We introduce CACHE-UK (Contextual Adaptive Continual Hybrid Editor for UK Finance), a stability-aware memory editing framework specifically designed for domain-specific, quantized LLMs. CACHE-UK integrates three components: a rank-1 LoRA perturbation mechanism that confines edits to the low-rank adapter subspace, a financial domain prioritization module for content-adaptive edit strength, and a closed-loop Stability Controller that tracks "degradation debt" to prevent catastrophic forgetting across sequential updates. Evaluated on a 4-bit quantized OpenLLaMA-3B model with a curated UK financial corpus of 88,021 documents, CACHE-UK reduces knowledge degradation by 11-17% relative to adapted baselines under identical 4-bit constraints -- its most robust effect -- while attaining the highest test success (generalization) rate observed in our setting (28%, a 6 percentage point improvement over the strongest adapted baseline). These results indicate that stability-aware editing can improve factual maintenance in resource-constrained financial LLM deployments, though absolute generalization rates remain low.
El Hassane Ettifouri, Ayoub Belfatmi, Mahaman Sanoussi Yahaya Alassan +1cs.AI
Quantized small autoregressive reasoning models can enter long, repetitive, or unproductive trajectories, yet inference-time compute is usually allocated without observing how a trajectory develops. Building on an earlier token-level e-CUSUM controller, we develop MGT-B (Monitoring-Guided Test-time Backtracking), a revised external controller that maps overlapping windows of pre-sampling uncertainty and degeneration features to position-conditional empirical tail probabilities, accumulates mixture betting factors with a CUSUM-shaped reset, and responds to an alarm by estimating a rollback point, restoring token and key-value-cache state, and performing constrained re-decoding. To audit whether the effect persists on problem identities first observed after the manual choice of log threshold h = 10, we retrospectively exclude 260 IDs present in pre-threshold artifacts and retain the chronologically first post-threshold pair for each remaining ID, yielding a 240-pair chronology-audit set. On this set, accuracy changes from 82/240 to 88/240 (+2.50 percentage points; 13 corrections, 7 regressions; exact McNemar p = 0.2632; paired bootstrap 95% interval [-1.25, +6.25]). A broader 467-pair historical-coverage set of seed-matched pairs changes accuracy from 146/467 to 167/467 (+4.50 points; McNemar p = 0.000753), but includes 200 seed-1 IDs available before or during threshold selection and is reported only as an exploratory estimate. All 316 no-alarm outputs in the 467-pair set are identical to vanilla, while the 151 alarmed trajectories contain 29 corrections and 8 regressions. Neither analysis is confirmatory, and the empirical factors are not established as a valid e-process or e-detector. The results support a selective monitoring-and-repair mechanism for the studied MATH-500 setting, rather than a general or theoretically certified reasoning improvement.
Low-bit quantization makes small reasoning models inexpensive to deploy but can degrade their chains of thought. This motivates decoder-side monitors that intervene when generation becomes unreliable. We show that a natural candidate, the centered token log-probability increment $\log p(w_t)+H_t$, is the wrong observable for this purpose. Under the model's own sampling law it is a mean-zero martingale by construction, so it measures sampling self-consistency rather than trajectory health and is nearly silent during confident repetition, where both $\log p(w_t)$ and entropy are close to zero. We introduce a training-free decoding controller that combines (i) a degeneration-aware alarm score fusing token uncertainty with explicit verbatim repetition and (ii) a calibrated e-process-inspired sequential detector. The raw product process is Ville-valid under a conditional-mean null, while the deployed CUSUM-floored statistic is treated as an empirical change detector because the score is history-dependent and autocorrelated. On GSM8K with DeepSeek-R1-Distill-Qwen-1.5B in FP16 and INT4, calibration turns a monitor that fires on 93--95% of generations into a selective detector of failing traces ($φ\approx 0.3$, precision $\approx 0.6$ against a 0.38 base rate). In this pilot, the controller reduces measured verbatim-degeneration signals and yields a positive but statistically inconclusive INT4 accuracy change from 63% to 69% (paired McNemar $p=0.18$, $n=100$), at a 28% token-budget cost. We also find that non-termination, rather than looping, is the dominant failure mode on GSM8K. The main contribution is methodological: an explanation of why centered token log-probability is inadequate for decoder monitoring and a calibrated, cautiously evaluated replacement.
Understanding which parameters are influential in Large Language Models (LLMs) is central to improving their efficiency, reliability, and interpretability. We introduce Weight-Adjusted Gradients (WAG), a simple yet effective approach for estimating parameter importance that explicitly captures the interaction between model weights and first-order gradient information and identifies parameters that disproportionately influence model behavior, such as those responsible for collapse phenomena in LLMs. Across a range of models and settings, we show that WAG surfaces a tiny but critical subset of parameters whose modification leads to dramatic degradation in performance, a failure mode that existing importance metrics overlook. These findings reveal a previously underexplored interplay between weights and gradients, suggesting that parameter importance cannot be fully understood through either signal alone. The surprising effectiveness of WAG points to fundamental structural properties of trained networks and motivates new open questions about the role of zeroth-order and first-order information in deep learning. We demonstrate the practical utility of WAG across multiple applications, including expert allocation in mixture-of-expert architectures, parameter-specific unlearning, mixed-precision quantization, and layer selection for knowledge editing. Our results position WAG as a unified approach for analyzing, debugging, and controlling LLMs, and opens new directions for principled model-level interpretation.
Renuka Oladri, Mohan Vamsi Varadaraju Priya, Jerry Wucs.CL cs.LG
We show that post-training quantization can silently alter how large language models reason even when task accuracy is preserved. Using a six-category failure taxonomy validated by two independent human annotators (Cohen's $κ$ = 0.906), we classify 30,000 chain-of-thought outputs from five instruction-tuned LLMs (3B--14B parameters) across three quantization precisions (FP32, FP16, NF4) and four reasoning benchmarks. We find that while accuracy is robust across precisions (maximum 3.1 pp drop), Hollow Convergence (correct answers reached through incomplete or unverifiable reasoning) shows a significant size-dependent shift under NF4, dropping sharply for the two smallest models tested but remaining invariant for models at 12B parameters and above. This effect is also benchmark-specific: GSM8K is categorically immune while LogiQA and ARC-Challenge show the largest shifts. Furthermore, under NF4, Shortcut Collapse rises from 44% to 78% of wrong-answer failures in LLaMA 3.2-3B while Confidence Snowballing collapses from 15.8% to near zero, a qualitative shift invisible to accuracy metrics. Finally, we show Hollow Convergence cannot be reliably detected from surface-level text features (best F1 = 0.53), establishing it as a deployment-relevant failure mode that standard evaluation pipelines cannot catch.
This report studies on-device English-to-Traditional-Chinese subtitle translation for Taiwan under short inputs, short outputs, batch-size-one inference, low latency, and privacy constraints. These conditions limit the value of optimizations designed for long-context or high-throughput language-model serving. Starting from LMT-60-0.6B, preliminary profiling suggests that vocabulary projection becomes a more important decode-time cost after GGUF quantization reduces the relative cost of Transformer blocks. We replace the original 151k-token vocabulary with a 64k-token subtitle-domain tokenizer, migrate the embedding space, and adapt the model through embedding calibration followed by full supervised fine-tuning. On an OpenSubtitles2024 test set, LocalSubs achieves a 59.2% tie-excluded win rate against Google Translate under GPT-4o pairwise judging. Performance is strongest on short cues and declines as cue length increases. In a separate preliminary Apple M2 Metal profiling run, LocalSubs shows a 1.63x speedup over a 151k-vocabulary baseline. The code is available on https://github.com/aiden1020/localsubs .
LLM-based text embedders have substantially improved retrieval and semantic representation quality, but their deployment remains costly: large backbone models slow down embedding inference, while high-dimensional full-precision embeddings impose substantial storage and bandwidth overhead on large-scale indexes. In this paper, we present BITEMBED, an extreme low-bit framework for LLM-based text embedding that jointly targets encoding efficiency and vector storage. BITEMBED converts pretrained LLM backbones into BitNet-style embedding encoders with ternary weights, quantized activations, and lightweight normalization refinement. The converted model is adapted to representation learning through continual contrastive pre-training, followed by supervised contrastive fine-tuning with both similarity-distribution distillation and attention-relation distillation from a full-precision teacher. Beyond quantizing the backbone, BITEMBED further trains output embeddings to support multiple storage precisions meeting different storage needs in various scenarios. Experiments on MMTEB (eng, v2) with Qwen3-0.6B and Gemma3-270M show that BITEMBED is largely comparable to full precision teacher embedders. Moreover, BITEMBED flexibly obtains text embeddings of various precisions, achieving a trade-off between performance and storage cost.
While 4-bit weight quantization is critical for deploying Small Language Models (SLMs) on edge devices, evaluations of the resulting performance degradation-the quantization tax-remain overwhelmingly English-centric. We present a zero-shot multilingual evaluation of 4-bit quantization across the Gemma 4 and Qwen 3.5 architectures. Evaluating on eight typo-logically diverse languages using MMLU ProX Lite and GlobalPIQA, we show parameter truncation exposes deep pre-training inequalities. We identify four phenomena: (1) Typological Fragility: low-resource and specific non-Latin scripts suffer representational collapse via architecture-specific double dissociations, failing to generate valid task logits; (2) Home Language Fragility Paradox: foundational pre-training pathways provide limited precision loss protection; (3) Domain-Specific Forgetting: multi-step cross-lingual routing degrades while associative soft-science recall remains robust; and (4) Quantization Resistance: highly saturated, typologically aligned domains resist deterministic degradation, with post-quantization performance gains bounded by statistical noise.
Luke Ztz Hu, Hongbing Lang, Songping Maics.LG cs.AI
Edge Internet of Things (IoT) agents are often constrained by memory capacity, privacy requirements, communication latency, and recurring inference cost. Current smart-home assistants commonly rely on API-level command interfaces or cloud-based language models that remain difficult to deploy on edge devices. This paper addresses edge IoT command generation as a many-to-one structured output task, where multiple natural-language instructions map to the same canonical command string for deterministic smart-home parsing. To support this setting, we propose Semantic-Conditioned Edge-Aware Neural Framework for Structured IoT Command Generation (SCENIC), an end-to-end framework covering model architecture selection, Smart Home Instruct data generation, triplet-loss contrastive supervised fine-tuning, pruning and quantization, and deployment-oriented export. We evaluate sub-0.2B-scale transformer backbones, which are, to the best of our knowledge, among the smallest language-model backbones studied for edge IoT structured command generation. On Smart Home Instruct-Bench, the strongest dense decoder-only row reaches 99.0% EM@1, while the encoder-decoder model retains stronger high-sparsity behavior. A representative pruned INT8 encoder-decoder export preserves 91.0% EM@1 and 99.0% EM@5 while reducing exported model size by 25.38%. TensorRT profiling of the NVIDIA 2:4 sparse encoder export further shows up to 1.8x encoder-component speedup, indicating that the selected encoder-decoder deployment path can retain structured command accuracy under edge-oriented compression while hardware acceleration evidence remains component-level. The SCENIC code and experimental artifacts are open sourced to support reproducibility.
Massive activation spikes in Large Language Models (LLMs) severely degrade quantization by stretching dynamic ranges. While prior hypotheses characterize these as high-level scalar biases, we argue that they are merely the scalar intermediates of rigid, structural vector biases in the spike-carrying tokens. We show that these tokens converge to constant vectors after normalization that drive the attention sink and value-state drain mechanisms. We geometrically substantiate this by analyzing the coordination of projection weights: $W_K$ contrastively amplifies the vector, $W_Q$ aligns semantic tokens toward it, and $W_V$ projects it into the spectral null-space. Furthermore, we reveal that the model actively preserves these structural biases against Rotary Positional Embedding (RoPE) perturbations by localizing them in "zones of rotational stability" utilizing low-frequency bands and coherent channel pairs. Leveraging this, we propose INSERTQUANT, a post-training quantization (PTQ) framework that clamps spikes and restores their function via pre-computed template vectors. This renders activations strictly spike-free, enabling robust low-bit quantization with high fidelity. INSERTQUANT achieves parity with state-of-the-art per-tensor quantization methods on LLMs and uniquely generalizes beyond text to other modalities such as ViTs.
Hy-MT2 is a family of fast-thinking multilingual translation models designed for complex real-world scenarios. It includes three model sizes: 1.8B, 7B, and 30B-A3B (MoE), all of which support translation among 33 languages and effectively follow translation instructions in multiple languages. For on-device deployment, with AngelSlim 1.25-bit extreme quantization, the 1.8B model requires only 440 MB of storage and improves inference speed by 1.5x. Multi-dimensional evaluations show that Hy-MT2 delivers outstanding performance across general, real-world business, domain-specific, and instruction-following translation tasks. The 7B and 30B models outperform open-source models such as DeepSeek-V4-Pro and Kimi K2.6 in fast-thinking mode, while the lightweight 1.8B model also surpasses mainstream commercial APIs from providers such as Microsoft and Doubao overall.
Arian Eamaz, Farhang Yeganegi, Mojtaba Soltanaliancs.LG
Understanding whether deep neural networks are effectively optimized remains challenging, as training occurs in highly nonconvex landscapes and standard metrics provide limited visibility into layer-wise learning quality. This challenge is particularly acute for transformer-based language models, where training is expensive, models are often reused in frozen form, and poorly optimized layers can silently degrade performance. We propose a layer-wise peeling framework for monitoring training dynamics, in which each transformer layer is locally optimized against intermediate representations of the trained model. By constructing lightweight, layer-specific reference solutions and projecting layers onto multiple intermediate outputs via different permutations, we obtain achievable baselines that enable fine-grained diagnosis of under-optimized layers. Experiments on decoder-only transformer models show that these layer-wise reference bounds can match or even surpass the trained model at various stages of training, exposing inefficiencies that remain hidden in aggregate loss curves. We further demonstrate that this analysis remains effective under binarization and quantized settings, where training dynamics are particularly fragile. Across all numerical results, the proposed bounds consistently separate apparent convergence from effective optimality, highlighting optimization opportunities that are invisible when relying on training loss alone.
Pretraining optimizers are tuned to produce the strongest possible base model, on the assumption that a stronger starting point yields a stronger model after subsequent changes like post-training and quantization. This overlooks the geometry of the base model which controls how much of the base model's capabilities survive subsequent parameter updates. We study three pretraining optimization approaches that bias optimization toward flatter minima: Sharpness-Aware Minimization (SAM), large learning rates, and shortened learning rate annealing periods. Across model sizes ranging from 20M to 150M parameters, we find that these interventions consistently improve downstream performance after post-training on five common datasets with up to 80% less forgetting. These principles hold at scale: a short SAM mid-training phase applied to an existing OLMo-2-1B checkpoint reduces forgetting by 31% after MetaMath post-training and by 40% after 4-bit quantization.