arXiv subjects

Pashmina Cameron

Publications and source records attributed to Pashmina Cameron.

At least 19 recordsLinked to original sources

arXiv2026

There Is No Free Loop: When Does Looping a Trained Model Pay for Its Inference Compute?

Looping re-applies a block of layers so that each token passes through the block more than once. Adding a loop to a trained model is attractive because it leaves the parameter count, and hence the disk size, the memory footprint and the training cost unchanged. Unfortunately, every generated token pays for the loop. Re-applying m of N layers makes each token about (N+m)/N times as expensive. Loops that a model learns during large-scale pretraining justify this price, but whether loops added after pretraining justify this cost remains unclear. We compare a looped model with two simple baselines. We first compare it with the same model trained identically without the loop (equal parameters). Secondly, we compare it with the best unlooped use of the same inference floating-point operations (equal FLOPs). Across Qwen3 and Qwen3.5 models from 0.8B to 9B parameters, only the looped 0.8B model beats both baselines. At 4B and 9B, the looped model does not beat the first baseline, i.e. the same model trained without the loop. This is not because the benchmarks are too easy, since on the same benchmarks the unlooped 9B model still scores up to 13 points higher than the 4B model, showing that there is ample room for improvement. Introducing loops in an already pretrained model therefore only pays for small models with a fixed parameter budget.

cs.LG↗

arXiv2026

Scaling Reasoning Efficiently via Relaxed On-Policy Distillation

On-policy distillation (OPD) is a promising approach for transferring reasoning capabilities to capacity-constrained models, yet sampled-token OPD often suffers from entropy collapse and negative transfer. We identify uniform credit allocation as a key bottleneck: existing objectives obtain dense teacher feedback on student-generated trajectories, but apply it uniformly across sampled tokens despite highly heterogeneous token-level signals. Most tokens carry redundant near-zero credit, while rare heavy-tailed negative credits can dominate updates and prematurely suppress plausible reasoning trajectories. We introduce REOPOLD (Relaxed On-Policy Distillation), a framework that relaxes this uniform assignment by controlling where teacher feedback is applied, how strongly it affects the update, and when the allocation rule changes during training. Across diverse reasoning tasks, REOPOLD improves sampled-token OPD over recent post-training baselines with up to 12x higher sample efficiency, and extends to cross-vocabulary distillation, self-distillation, and test-time scaling.

cs.LG↗

arXiv2026

ThinkBit: Search Codewords, Recalibrate Scales, Select by Reasoning

Reasoning language models generate thousands of tokens per answer, and generation speed is limited by memory bandwidth, so storing weights in two bits instead of bfloat16 can potentially make inference 8$\times$ cheaper. Two-bit methods are usually judged by perplexity and multiple-choice question-answering accuracy, both of which say little about whether a model can still reason through more difficult problems. We present ThinkBit, a strictly scalar two-bit framework built to preserve reasoning, in which every weight is an integer codeword times a bfloat16 scale shared by a group of weights. We first use a sensitivity metric to assign each layer a ternary, 2-bit, or 3-bit tier to allocate bitwidths optimally among layers. This is followed by a layerwise Gumbel-softmax-based scalar quantization algorithm that searches for optimal codewords block by block. Since a layerwise scheme does not capture how quantized layers interact with each other, we recalibrate scales, with the codewords frozen, using a knowledge-distillation-based loss against the teacher model. We then perform a short quantization-aware training (QAT) stage that lets the model mimic the base model's distribution by reassigning codewords. Finally, a paired reasoning test on training problems selects the final checkpoint between the recalibration stage and the QAT stage. On Qwen3-1.7B at matched size, our recalibrated checkpoint scores 46.7 on MATH-500 against 30.0 for EfficientQAT trained on the same data, and 73.0 after QAT on reasoning traces; at 0.52 GB it is smaller than every published 2-bit GGUF of the model yet more accurate on MATH-500. We show that this framework scales and up to 14B, a 3.74 GB model scores 94.8 on MATH-500, against 97.5 for the 29.5 GB bfloat16 model.

cs.AI↗

arXiv2026

Beyond 4 bits: Data-Centric Post-training Quantization and Recovery

Post-training quantization (PTQ) compresses the weights of a trained model with limited compute and data. Down to 4 bits it is near-lossless, but at 2 bits scalar PTQ configurations fall to chance-level accuracy. We study three choices usually treated as implementation details that matter little at 4 bits but decide 2-bit quality: the quantization representation, the calibration data, and the recovery fine-tuning data, across six dense and mixture-of experts models (ranging from $3.8$ to $32$B parameters). Representation is the largest factor and vector codebooks restore usable models where scalar PTQ pipelines collapse. Most of the remaining loss in accuracy is on \emph{generative} tasks, and the amount of performance drop is model-dependent. Furthermore, at a short recovery budget, calibration and recovery data are coupled. The best recovery data depends on the calibration data, so the two must be chosen jointly. The resulting recipe involves a vector codebook representation, a small task-aligned calibration mixture, and an inexpensive codebook-only recovery stage. Including codebooks, it uses about $2.3$ effective bits per weight and retains $89$ to $97\%$ of full-precision question-answering (QA) accuracy where scalar PTQ is at chance, and, including recovery, uses $1.4$ to $30\times$ less compute than the strongest quantization-aware-training (QAT) baseline.

cs.AI↗

arXiv2026

Eliciting Weak-to-Strong Generalization with On-Policy Reverse Distillation

Weak-to-strong generalization asks whether stronger models can learn from weaker supervisors and surpass them. This is particularly important for successive model generations and multi-domain consolidation, where frontier-scale post-training from scratch can be prohibitively expensive. Yet conventional distillation treats the weak teacher as an optimization target, potentially imposing its capacity ceiling on the student. We introduce On-Policy Reverse Distillation (OPRD), which evaluates the teacher's policy shift relative to its reference policy on student rollouts and amplifies the component of the student's verifier-driven policy gradient along that direction. By rescaling only verifier-supported updates, OPRD preserves the stationary points of policy optimization while accelerating learning beyond the teacher. Our evaluations cover mathematical reasoning, logical reasoning, code generation, and visual tool use. In both successive model transfer and multi-teacher distillation, OPRD achieves higher performance with fewer student updates than existing RL and distillation approaches. Response-style analysis shows that OPRD students remain closer to verifier-only RL models than to their weak teachers, suggesting that guidance accelerates rather than redirects student optimization. Strong-to-weak results further show that OPRD combines verifier-driven optimization with teacher guidance regardless of capacity ordering.

cs.LG↗

arXiv2026

Sliding-window beats linear attention

Due to the nature of quadratic attention, Large Language Models (LLMs) consume a lot of memory and energy: every new token costs more than the previous one, and its keys and values must be stored in memory indefinitely, which is unsustainable. Two main lines of work address this: compressing the KV cache, e.g., by evicting or quantizing keys and values, and retrofitting LLMs to use Linear Attention, which replaces the KV cache with a fixed-size state. Retrofitting has attracted a lot of attention, given its promise to solve the quadratic scaling problem with state-of-the-art performance at low cost. However, it has not been properly compared to the simplest form of KV-cache eviction: Sliding Window Attention (SWA) with attention sinks. In this work, we show that SWA with sinks performs as well or better than most retrofitted Linear Attention models across multiple LLMs and downstream tasks, with the largest gains on long-context and generative tasks. On long-context reasoning tasks (Needle-in-a-Haystack and BABILong), SWA achieves massively higher performance (2 to 10 times higher than linear attention). SWA requires no additional training, is extremely fast, and requires little memory, making it an extremely cheap and reliable solution. When the training budget is limited, switching to SWA is a much more effective way to reduce inference memory cost than retrofitting linear attention. Linear attention models have shown promise, but they require training from scratch or extensive retrofitting to reap their architectural benefits and come close to SWA.

cs.CL↗

arXiv2026

JIVEAdapter: A Multi-Task Additive Low-Rank Adapter via Joint and Individual Variation Explained (JIVE)

Parameter-efficient fine-tuning adapts pretrained models at a fraction of the cost of full fine-tuning, yet most low-rank adapters are single-task and represent each weight update multiplicatively, leaving no explicit account of what is shared across tasks and what is task-specific. We introduce JIVEAdapter, a multi-task "additive" low-rank adapter inspired by statistical Joint and Individual Variation Explained (JIVE). JIVEAdapter decomposes every weight update into a Joint structure shared across all tasks plus a per-task Individual structure, penalizes the Individual structures to be near-orthogonal to the Joint so shared and task-specific signal stay "interpretable" and separated, and allocates rank adaptively across a shared Joint pool and a per-task Individual pool. The Joint is learned once, jointly over a task group or incrementally, one task at a time, then frozen and reused as a prior for new tasks without retraining the shared part. On GLUE and SuperGLUE with DeBERTaV3-base, JIVEAdapter is competitive with strong single-task and multi-task low-rank baselines at a matched per-task effective rank, without extra modules such as MoE, and when a related held-in task exists its frozen Joint serves a held-out task by reusing that task's Individual with only a cheap per-direction scale, otherwise training a small new one.

cs.AI↗

arXiv2026

AttSVD:Prompt-Adaptive Low-Rank KV Cache Compression via Attention-Guided SVD

The key-value (KV) cache of autoregressive transformers grows linearly with context length and dominates memory at long context. Most training-free remedies evict low-importance tokens, an irreversible choice along the sequence axis. We instead keep every token and store it more cheaply along the "feature" axis. We therefore propose AttSVD, a new "interpretable" low-rank compression whose basis is derived from each prompt's own attention geometry: an online, per-prompt truncated SVD that keeps only the directions attention actually reads, cutting persistent per-head KV memory in proportion to the retained rank. We propose two decode-time caching strategies, accumulating and streaming, for short and long generation regimes. Furthermore, we propose two refinements that make compression adaptive. A per-matrix energy rule sizes the logit space and the attention mass independently. An attention-aware basis truncates only in the spaces attention actually reads, preserving both the attention logits and the attention output. The same factors also provide free, per-head interpretability insights into the effective rank and the geometry attention consumes. Across multiple models, on both an agentic benchmark and the full LongBench suite AttSVD stays on par with the dense cache while using up to 50% of the KV-cache memory.

cs.LG↗

arXiv2026

Self-reflecting Large Language Models: A Hegelian Dialectical Approach

In this paper, we introduce a self-reflection framework for Large Language Models (LLMs) grounded in the Hegelian Dialectic, a philosophical method in which an initial proposition is challenged by a generated opposition, and both are reconciled into a unified, more comprehensive idea. We formalize this process as an iterative operator over the space of consistent theories and apply it to two complementary tasks:(i) generating novel scientific ideas across domains such as mathematics, physics, economics, and philosophy, and (ii)improving reasoning by enabling LLMs to identify and correct their own errors through structured self-critique. We study generation temperature through two configurations (a dynamic annealing schedule that shifts from creative exploration to refinement, and a constant temperature), to examine the effect of fixed versus dynamic temperature rather than advocate either. To evaluate ideas without domain experts, we introduce Multi-Agent Majority Voting (MAMV), in which multiple LLMs independently assess the validity and novelty of each synthesis. Our experiments show significant gains over baselines on mathematical (GSM-8k, GSM-hard), symbolic (GSM-Symbolic), and knowledge-intensive (MMLU Pro) reasoning, with promising qualitative results in open-ended scientific ideation.

cs.CL↗

arXiv2026

WINA: Weight Informed Neuron Activation for Accelerating Large Language Model Inference

The growing computational demands of large language models (LLMs) make efficient inference and activation strategies increasingly critical. While recent approaches, such as Mixture-of-Experts (MoE), leverage selective activation but require specialized training, training-free sparse activation methods offer broader applicability and superior resource efficiency through their plug-and-play design. However, many existing methods rely solely on hidden state magnitudes to determine activation, resulting in high approximation errors and suboptimal inference accuracy. To address these limitations, we propose WINA (Weight Informed Neuron Activation), a novel, simple, and training-free sparse activation framework that jointly considers hidden state magnitudes and the column-wise $\ell_2$-norms of weight matrices. We show that this leads to a sparsification strategy that obtains optimal approximation error bounds with theoretical guarantees tighter than existing techniques. Empirically, WINA also outperforms state-of-the-art methods (e.g., TEAL) by up to $2.94\%$ in average performance at the same sparsity levels, across a diverse set of LLM architectures and datasets. These results position WINA as a new performance frontier for training-free sparse activation in LLM inference, advancing training-free sparse activation methods and setting a robust baseline for efficient inference. The source code is available at https://github.com/microsoft/wina.

cs.LG↗

arXiv2026

StableQAT: Stable Quantization-Aware Training at Ultra-Low Bitwidths

Quantization-aware training (QAT) is essential for deploying large models under strict memory and latency constraints, yet achieving stable and robust optimization at ultra-low bitwidths remains challenging. Common approaches based on the straight-through estimator (STE) or soft quantizers often suffer from gradient mismatch, instability, or high computational overhead. As such, we propose StableQAT, a unified and efficient QAT framework that stabilizes training in ultra low-bit settings via a novel, lightweight, and theoretically grounded surrogate for backpropagation derived from a discrete Fourier analysis of the rounding operator. StableQAT strictly generalizes STE as the latter arises as a special case of our more expressive surrogate family, yielding smooth, bounded, and inexpensive gradients that improve QAT training performance and stability across various hyperparameter choices. In experiments, StableQAT exhibits stable and efficient QAT at 2-4 bit regimes, demonstrating improved training stability, robustness, and superior performance with negligible training overhead against standard QAT techniques. Our code is available at https://github.com/microsoft/StableQAT.

cs.LG↗

arXiv2026

CUA-Skill: Develop Skills for Computer Using Agent

Computer-Using Agents (CUAs) aim to autonomously operate computer systems to complete real-world tasks. However, existing agentic systems remain difficult to scale and lag behind human performance. A key limitation is the absence of reusable and structured skill abstractions that capture how humans interact with graphical user interfaces and how to leverage these skills. We introduce CUA-Skill, a computer-using agentic skill base that encodes human computer-use knowledge as skills coupled with parameterized execution and composition graphs. CUA-Skill is a large-scale library of carefully engineered skills spanning common Windows applications, serving as a practical infrastructure and tool substrate for scalable, reliable agent development. Built upon this skill base, we construct CUA-Skill Agent, an end-to-end computer-using agent that supports dynamic skill retrieval, argument instantiation, and memory-aware failure recovery. Our results demonstrate that CUA-Skill substantially improves execution success rates and robustness on challenging end-to-end agent benchmarks, establishing a strong foundation for future computer-using agent development. On WindowsAgentArena, CUA-Skill Agent achieves state-of-the-art 57.5% (best of three) successful rate while being significantly more efficient than prior and concurrent approaches. The project page is available at https://microsoft.github.io/cua_skill/.

cs.AI↗

arXiv2025

AppSelectBench: Application-Level Tool Selection Benchmark

Computer Using Agents (CUAs) are increasingly equipped with external tools, enabling them to perform complex and realistic tasks. For CUAs to operate effectively, application selection, which refers to deciding which application to use before invoking fine-grained tools such as APIs, is a fundamental capability. It determines whether the agent initializes the correct environment, avoids orchestration confusion, and efficiently focuses on relevant context. However, existing benchmarks primarily assess fine-grained API selection, offering limited insight into whether models can reason across and choose between different applications. To fill this gap, we introduce AppSelectBench, a comprehensive benchmark for evaluating application selection in CUAs. AppSelectBench contains a novel user task generation pipeline that produces realistic, diverse, and semantically grounded user intents at scale, together with unified evaluation protocols covering random, heuristic, zero-shot, few-shot, and retrieval-augmented-settings. AppSelectBench covers one hundred widely used desktop applications and includes more than one hundred thousand realistic, diverse, and semantically grounded user tasks. Extensive experiments across both closed-source and open-source large language models reveal systematic strengths and weaknesses in inter-application reasoning, showing that even the most capable models still struggle to make consistent application choices. Together, these results establish AppSelectBench as a foundation for studying and advancing application level reasoning, an essential yet underexplored capability of intelligent CUAs. The source is available at https://microsoft.github.io/appselectbench/.

cs.CL↗

arXiv2024

QuaRot: Outlier-Free 4-Bit Inference in Rotated LLMs

We introduce QuaRot, a new Quantization scheme based on Rotations, which is able to quantize LLMs end-to-end, including all weights, activations, and KV cache in 4 bits. QuaRot rotates LLMs in a way that removes outliers from the hidden state without changing the output, making quantization easier. This computational invariance is applied to the hidden state (residual) of the LLM, as well as to the activations of the feed-forward components, aspects of the attention mechanism, and to the KV cache. The result is a quantized model where all matrix multiplications are performed in 4 bits, without any channels identified for retention in higher precision. Our 4-bit quantized LLaMa2-70B model has losses of at most 0.47 WikiText-2 perplexity and retains 99% of the zero-shot performance. We also show that QuaRot can provide lossless 6 and 8 bit LLaMa2 models without any calibration data using round-to-nearest quantization. Code is available at: https://github.com/spcl/QuaRot.

cs.LG↗

arXiv2024

Learning to Extend Molecular Scaffolds with Structural Motifs

Recent advancements in deep learning-based modeling of molecules promise to accelerate in silico drug discovery. A plethora of generative models is available, building molecules either atom-by-atom and bond-by-bond or fragment-by-fragment. However, many drug discovery projects require a fixed scaffold to be present in the generated molecule, and incorporating that constraint has only recently been explored. Here, we propose MoLeR, a graph-based model that naturally supports scaffolds as initial seed of the generative procedure, which is possible because it is not conditioned on the generation history. Our experiments show that MoLeR performs comparably to state-of-the-art methods on unconstrained molecular optimization tasks, and outperforms them on scaffold-based tasks, while being an order of magnitude faster to train and sample from than existing approaches. Furthermore, we show the influence of a number of seemingly minor design choices on the overall performance.

cs.LG↗

arXiv2023

HMD-NeMo: Online 3D Avatar Motion Generation From Sparse Observations

Generating both plausible and accurate full body avatar motion is the key to the quality of immersive experiences in mixed reality scenarios. Head-Mounted Devices (HMDs) typically only provide a few input signals, such as head and hands 6-DoF. Recently, different approaches achieved impressive performance in generating full body motion given only head and hands signal. However, to the best of our knowledge, all existing approaches rely on full hand visibility. While this is the case when, e.g., using motion controllers, a considerable proportion of mixed reality experiences do not involve motion controllers and instead rely on egocentric hand tracking. This introduces the challenge of partial hand visibility owing to the restricted field of view of the HMD. In this paper, we propose the first unified approach, HMD-NeMo, that addresses plausible and accurate full body motion generation even when the hands may be only partially visible. HMD-NeMo is a lightweight neural network that predicts the full body motion in an online and real-time fashion. At the heart of HMD-NeMo is the spatio-temporal encoder with novel temporally adaptable mask tokens that encourage plausible motion in the absence of hand observations. We perform extensive analysis of the impact of different components in HMD-NeMo and introduce a new state-of-the-art on AMASS dataset through our evaluation.

cs.CV↗

arXiv2022

FLAG: Flow-based 3D Avatar Generation from Sparse Observations

To represent people in mixed reality applications for collaboration and communication, we need to generate realistic and faithful avatar poses. However, the signal streams that can be applied for this task from head-mounted devices (HMDs) are typically limited to head pose and hand pose estimates. While these signals are valuable, they are an incomplete representation of the human body, making it challenging to generate a faithful full-body avatar. We address this challenge by developing a flow-based generative model of the 3D human body from sparse observations, wherein we learn not only a conditional distribution of 3D human pose, but also a probabilistic mapping from observations to the latent space from which we can generate a plausible pose along with uncertainty estimates for the joints. We show that our approach is not only a strong predictive model, but can also act as an efficient pose prior in different optimization settings where a good initial latent code plays a major role.

cs.CV↗

arXiv2021

Instructions and Guide for Diagnostic Questions: The NeurIPS 2020 Education Challenge

Digital technologies are becoming increasingly prevalent in education, enabling personalized, high quality education resources to be accessible by students across the world. Importantly, among these resources are diagnostic questions: the answers that the students give to these questions reveal key information about the specific nature of misconceptions that the students may hold. Analyzing the massive quantities of data stemming from students' interactions with these diagnostic questions can help us more accurately understand the students' learning status and thus allow us to automate learning curriculum recommendations. In this competition, participants will focus on the students' answer records to these multiple-choice diagnostic questions, with the aim of 1) accurately predicting which answers the students provide; 2) accurately predicting which questions have high quality; and 3) determining a personalized sequence of questions for each student that best predicts the student's answers. These tasks closely mimic the goals of a real-world educational platform and are highly representative of the educational challenges faced today. We provide over 20 million examples of students' answers to mathematics questions from Eedi, a leading educational platform which thousands of students interact with daily around the globe. Participants to this competition have a chance to make a lasting, real-world impact on the quality of personalized education for millions of students across the world.

cs.CY↗