Opt.Gear Technical Report

We introduce Opt.Gear, 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, Opt.Gear 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 Opt.Gear 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 Opt.Gear-1M that can be deployed on Micro-Controller Units (MCUs), a Tiny Language Model (TLM). Opt.Gear-1M is the first generative language model to achieve 20 TPS with W4A32 quantization on the ARM Cortex-M7 CPU of the STM32H747I-DISCO.

FoRA: Fisher-orthogonal Rank Adaptation for Parameter-Efficient Fine-Tuning

Parameter-efficient fine-tuning(PEFT) has largely focused on LoRA and its accuracy-oriented variants, leaving the original goal of reducing trainable parameters has receivedcomparatively little attention. We introduce FoRA, which revisits this goal by reducing the number of adapted layers rather than adapter rank. FoRA selects task-informative layers via a single-pass diagonal Fisher score (under 1% of training cost) and trains the LoRA down-projection at selected layers on the Stiefel manifold, preserving column orthonormality and effective rank. FoRA consistently outperforms LoRA and DoRA at half their parameter budget, and falls within 0.7-0.8 accuracy points of AdaLoRA at one-quarter its parameter count, across five LLaMA-family backbones. Cross-architecture experiments on twelve backbones from the LLaMA, Qwen3, and Gemma families confirm consistent gains from 270M to 32B parameters. The two components combine super-additively: Fisher selection alone matches rank reduction at the same budget, while the Stiefel constraint provides the decisive additional gain.

Rethinking the Rank Threshold for LoRA Fine-Tuning

A recent landscape analysis of LoRA fine-tuning in the neural tangent kernel regime establishes a sufficient condition $r(r+1)/2 > KN$ on the LoRA rank $r$ for the absence of spurious local minima under squared-error loss, prescribing $r \geq 12$ on canonical few-shot RoBERTa setups. The condition is stated for general output dimension $K$, so its sharpness in any particular regime, and its practical implication for the cross-entropy loss actually used in fine-tuning, are open. We give three results that together reduce the prescribed rank to $r = 1$ for binary classification in this regime. First, replacing the symmetric Sard-form count with the non-symmetric LoRA manifold dimension yields a strictly weaker capacity requirement, $r(m+n) - r^2 > C^* \cdot KN$ with $C^* \approx 1.35$ under Gaussian-iid features, satisfied at $r = 1$ on canonical setups. Second, in the cross-entropy setting the Polyak--Łojasiewicz inequality removes the rank threshold entirely. Third, a Rademacher-complexity bound predicts rank-one variance optimality precisely when the bias term is saturated, which is the case for binary classification but not for $K > 2$. Empirically, across four GLUE-style binary tasks, three encoder architectures, and at scale on RoBERTa-large, rank one is competitive with the existing prescription $r = 12$; on multi-class MNLI the optimal rank shifts above one, also as predicted. The binary-regime guarantees are conditional on standard NTK assumptions; the multi-class extension is left to future work.

LCSB: Layer-Cyclic Selective Backpropagation for Memory-Efficient On-Device LLM Fine-Tuning

Memory-efficient backpropagation (MeBP) has enabled first-order fine-tuning of large language models (LLMs) on mobile devices with less than 1GB memory. However, MeBP requires backward computation through all transformer layers at every step, where weight decompression alone accounts for 32--42% of backward time. We propose Layer-Cyclic Selective Backpropagation (LCSB), which computes gradients for only a subset of layers per step. Our key insight is that residual connections guarantee gradient flow through identity paths, while AdamW momentum provides implicit updates for non-selected layers. We interpret LCSB as Block Coordinate Descent on the LoRA parameter space, providing theoretical justification for convergence. LCSB achieves up to 1.40$\times$ speedup with less than 2\% quality degradation across five models and three tasks. Surprisingly, in 4-bit quantized settings, LCSB exhibits superior stability: a 3B model that completely diverges under full backpropagation converges smoothly with LCSB, suggesting an implicit regularization effect from selective gradient computation.

Memory-Efficient Structured Backpropagation for On-Device LLM Fine-Tuning

On-device fine-tuning enables privacy-preserving personalization of large language models, but mobile devices impose severe memory constraints, typically 6--12GB shared across all workloads. Existing approaches force a trade-off between exact gradients with high memory (MeBP) and low memory with noisy estimates (MeZO). We propose Memory-efficient Structured Backpropagation (MeSP), which bridges this gap by manually deriving backward passes that exploit LoRA's low-rank structure. Our key insight is that the intermediate projection $h = xA$ can be recomputed during backward at minimal cost since rank $r \ll d_{in}$, eliminating the need to store it. MeSP achieves 49\% average memory reduction compared to MeBP on Qwen2.5 models (0.5B--3B) while computing mathematically identical gradients. Our analysis also reveals that MeZO's gradient estimates show near-zero correlation with true gradients (cosine similarity $\approx$0.001), explaining its slow convergence. MeSP reduces peak memory from 361MB to 136MB for Qwen2.5-0.5B, enabling fine-tuning scenarios previously infeasible on memory-constrained devices.

Riemannian Optimization for LoRA on the Stiefel Manifold

While powerful, large language models (LLMs) present significant fine-tuning challenges due to their size. Parameter-efficient fine-tuning (PEFT) methods like LoRA provide solutions, yet suffer from critical optimizer inefficiencies; notably basis redundancy in LoRA's $B$ matrix when using AdamW, which fundamentally limits performance. We address this by optimizing the $B$ matrix on the Stiefel manifold, imposing explicit orthogonality constraints that achieve near-perfect orthogonality and full effective rank. This geometric approach dramatically enhances parameter efficiency and representational capacity. Our Stiefel optimizer consistently outperforms AdamW across benchmarks with both LoRA and DoRA, demonstrating that geometric constraints are the key to unlocking LoRA's full potential for effective LLM fine-tuning.

Towards Zero-Shot Functional Compositionality of Language Models

Large Pre-trained Language Models (PLM) have become the most desirable starting point in the field of NLP, as they have become remarkably good at solving many individual tasks. Despite such success, in this paper, we argue that current paradigms of working with PLMs are neglecting a critical aspect of modeling human intelligence: functional compositionality. Functional compositionality - the ability to compose learned tasks - has been a long-standing challenge in the field of AI (and many other fields) as it is considered one of the hallmarks of human intelligence. An illustrative example of such is cross-lingual summarization, where a bilingual person (English-French) could directly summarize an English document into French sentences without having to translate the English document or summary into French explicitly. We discuss why this matter is an important open problem that requires further attention from the field. Then, we show that current PLMs (e.g., GPT-2 and T5) don't have functional compositionality yet and it is far from human-level generalizability. Finally, we suggest several research directions that could push the field towards zero-shot functional compositionality of language models.

Evaluating the Knowledge Dependency of Questions

The automatic generation of Multiple Choice Questions (MCQ) has the potential to reduce the time educators spend on student assessment significantly. However, existing evaluation metrics for MCQ generation, such as BLEU, ROUGE, and METEOR, focus on the n-gram based similarity of the generated MCQ to the gold sample in the dataset and disregard their educational value. They fail to evaluate the MCQ's ability to assess the student's knowledge of the corresponding target fact. To tackle this issue, we propose a novel automatic evaluation metric, coined Knowledge Dependent Answerability (KDA), which measures the MCQ's answerability given knowledge of the target fact. Specifically, we first show how to measure KDA based on student responses from a human survey. Then, we propose two automatic evaluation metrics, KDA_disc and KDA_cont, that approximate KDA by leveraging pre-trained language models to imitate students' problem-solving behavior. Through our human studies, we show that KDA_disc and KDA_soft have strong correlations with both (1) KDA and (2) usability in an actual classroom setting, labeled by experts. Furthermore, when combined with n-gram based similarity metrics, KDA_disc and KDA_cont are shown to have a strong predictive power for various expert-labeled MCQ quality measures.

Automated Evaluation for Student Argumentative Writing: A Survey

This paper surveys and organizes research works in an under-studied area, which we call automated evaluation for student argumentative writing. Unlike traditional automated writing evaluation that focuses on holistic essay scoring, this field is more specific: it focuses on evaluating argumentative essays and offers specific feedback, including argumentation structures, argument strength trait score, etc. The focused and detailed evaluation is useful for helping students acquire important argumentation skill. In this paper we organize existing works around tasks, data and methods. We further experiment with BERT on representative datasets, aiming to provide up-to-date baselines for this field.

GRAM: Fast Fine-tuning of Pre-trained Language Models for Content-based Collaborative Filtering

Content-based collaborative filtering (CCF) provides personalized item recommendations based on both users' interaction history and items' content information. Recently, pre-trained language models (PLM) have been used to extract high-quality item encodings for CCF. However, it is resource-intensive to finetune PLM in an end-to-end (E2E) manner in CCF due to its multi-modal nature: optimization involves redundant content encoding for interactions from users. For this, we propose GRAM (GRadient Accumulation for Multi-modality): (1) Single-step GRAM which aggregates gradients for each item while maintaining theoretical equivalence with E2E, and (2) Multi-step GRAM which further accumulates gradients across multiple training steps, with less than 40\% GPU memory footprint of E2E. We empirically confirm that GRAM achieves a remarkable boost in training efficiency based on five datasets from two task domains of Knowledge Tracing and News Recommendation, where single-step and multi-step GRAM achieve 4x and 45x training speedup on average, respectively.