Julian Salazar

Zejiang Hou, Julian Salazar, George Polovets · amazon-science

Large pretrained language models (PLMs) are often domain- or task-adapted via fine-tuning or prompting. Finetuning requires modifying all of the parameters and having enough data to avoid overfitting while prompting requires no training and few examples but limits performance. Instead, we prepare PLMs for data- and parameter-efficient adaptation by learning to learn the difference between general and adapted PLMs. This difference is expressed in terms of model weights and sublayer structure through our proposed dynamic low-rank reparameterization and learned architecture controller. Experiments on few-shot dialogue completion, low-resource abstractive summarization, and multi-domain language modeling show improvements in adaptation time and performance over direct finetuning or preparation via domain-adaptive pretraining. Ablations show our task-adaptive reparameterization (TARP) and model search (TAMS) components individually improve on other parameter-efficient transfer like adapters and structure-learning methods like learned sparsification.

Dimitri Kanevsky, Julian Salazar, Matt Harvey

Let $V$ be a smooth cubic surface over a $p$-adic field $k$ with good reduction. Swinnerton-Dyer (1981) proved that $R$-equivalence is trivial on $V(k)$ except perhaps if $V$ is one of three special types--those whose $R$-equivalence he could not bound by proving the universal (admissible) equivalence is trivial. We consider all surfaces $V$ currently known to have non-trivial universal equivalence. Beyond being intractable to Swinnerton-Dyer's approach, we observe that if these surfaces also had non-trivial $R$-equivalence, they would contradict Colliot-Thélène and Sansuc's conjecture regarding the $k$-rationality of universal torsors for geometrically rational surfaces. By devising new methods to study $R$-equivalence, we prove that for 2-adic surfaces with all-Eckardt reductions (the third special type, which contains every existing case of non-trivial universal equivalence), $R$-equivalence is trivial or of exponent 2. For the explicit cases, we confirm triviality: the diagonal cubic $X^3+Y^3+Z^3+ζ_3 T^3=0$ over $\mathbb{Q}_2(ζ_3)$--answering a long-standing question of Manin's (Cubic Forms, 1972)--and the cubic with universal equivalence of exponent 2 (Kanevsky, 1982). This is the first in a series of works derived from a year of interactions with generative AI models such as AlphaEvolve and Gemini 3 Deep Think, with the latter proving many of our lemmas. We disclose the timeline and nature of their use towards this paper, and describe our broader AI-assisted research program in a companion report (in preparation).

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Long Phan, Alice Gatti, Ziwen Han et al. · amazon-science, apple-ml

Benchmarks are important tools for tracking the rapid advancements in large language model (LLM) capabilities. However, benchmarks are not keeping pace in difficulty: LLMs now achieve over 90\% accuracy on popular benchmarks like MMLU, limiting informed measurement of state-of-the-art LLM capabilities. In response, we introduce Humanity's Last Exam (HLE), a multi-modal benchmark at the frontier of human knowledge, designed to be the final closed-ended academic benchmark of its kind with broad subject coverage. HLE consists of 2,500 questions across dozens of subjects, including mathematics, humanities, and the natural sciences. HLE is developed globally by subject-matter experts and consists of multiple-choice and short-answer questions suitable for automated grading. Each question has a known solution that is unambiguous and easily verifiable, but cannot be quickly answered via internet retrieval. State-of-the-art LLMs demonstrate low accuracy and calibration on HLE, highlighting a significant gap between current LLM capabilities and the expert human frontier on closed-ended academic questions. To inform research and policymaking upon a clear understanding of model capabilities, we publicly release HLE at https://lastexam.ai.

RJ Skerry-Ryan, Julian Salazar, Soroosh Mariooryad et al.

We introduce a neural network layer API and library for sequence modeling, designed for easy creation of sequence models that can be executed both layer-by-layer (e.g., teacher-forced training) and step-by-step (e.g., autoregressive sampling). To achieve this, layers define an explicit representation of their state over time (e.g., a Transformer KV cache, a convolution buffer, an RNN hidden state), and a step method that evolves that state, tested to give identical results to a stateless layer-wise invocation. This and other aspects of the SequenceLayers contract enables complex models to be immediately streamable, mitigates a wide range of common bugs arising in both streaming and parallel sequence processing, and can be implemented in any deep learning library. A composable and declarative API, along with a comprehensive suite of layers and combinators, streamlines the construction of production-scale models from simple streamable components while preserving strong correctness guarantees. Our current implementations of SequenceLayers (JAX, TensorFlow 2) are available at https://github.com/google/sequence-layers.

Eliya Nachmani, Alon Levkovitch, Roy Hirsch et al.

We present Spectron, a novel approach to adapting pre-trained large language models (LLMs) to perform spoken question answering (QA) and speech continuation. By endowing the LLM with a pre-trained speech encoder, our model becomes able to take speech inputs and generate speech outputs. The entire system is trained end-to-end and operates directly on spectrograms, simplifying our architecture. Key to our approach is a training objective that jointly supervises speech recognition, text continuation, and speech synthesis using only paired speech-text pairs, enabling a `cross-modal' chain-of-thought within a single decoding pass. Our method surpasses existing spoken language models in speaker preservation and semantic coherence. Furthermore, the proposed model improves upon direct initialization in retaining the knowledge of the original LLM as demonstrated through spoken QA datasets. We release our audio samples (https://michelleramanovich.github.io/spectron/spectron) and spoken QA dataset (https://github.com/google-research-datasets/LLAMA1-Test-Set).

Jianfeng He, Julian Salazar, Kaisheng Yao et al.

End-to-end (E2E) spoken language understanding (SLU) is constrained by the cost of collecting speech-semantics pairs, especially when label domains change. Hence, we explore \textit{zero-shot} E2E SLU, which learns E2E SLU without speech-semantics pairs, instead using only speech-text and text-semantics pairs. Previous work achieved zero-shot by pseudolabeling all speech-text transcripts with a natural language understanding (NLU) model learned on text-semantics corpora. However, this method requires the domains of speech-text and text-semantics to match, which often mismatch due to separate collections. Furthermore, using the entire collected speech-text corpus from any domains leads to \textit{imbalance} and \textit{noise} issues. To address these, we propose \textit{cross-modal selective self-training} (CMSST). CMSST tackles imbalance by clustering in a joint space of the three modalities (speech, text, and semantics) and handles label noise with a selection network. We also introduce two benchmarks for zero-shot E2E SLU, covering matched and found speech (mismatched) settings. Experiments show that CMSST improves performance in both two settings, with significantly reduced sample sizes and training time. Our code and data are released in https://github.com/amazon-science/zero-shot-E2E-slu.

Julian Salazar, Davis Liang, Toan Q. Nguyen et al.

Pretrained masked language models (MLMs) require finetuning for most NLP tasks. Instead, we evaluate MLMs out of the box via their pseudo-log-likelihood scores (PLLs), which are computed by masking tokens one by one. We show that PLLs outperform scores from autoregressive language models like GPT-2 in a variety of tasks. By rescoring ASR and NMT hypotheses, RoBERTa reduces an end-to-end LibriSpeech model's WER by 30% relative and adds up to +1.7 BLEU on state-of-the-art baselines for low-resource translation pairs, with further gains from domain adaptation. We attribute this success to PLL's unsupervised expression of linguistic acceptability without a left-to-right bias, greatly improving on scores from GPT-2 (+10 points on island effects, NPI licensing in BLiMP). One can finetune MLMs to give scores without masking, enabling computation in a single inference pass. In all, PLLs and their associated pseudo-perplexities (PPPLs) enable plug-and-play use of the growing number of pretrained MLMs; e.g., we use a single cross-lingual model to rescore translations in multiple languages. We release our library for language model scoring at https://github.com/awslabs/mlm-scoring.

Shaoshi Ling, Julian Salazar, Yuzong Liu et al.

We introduce BERTphone, a Transformer encoder trained on large speech corpora that outputs phonetically-aware contextual representation vectors that can be used for both speaker and language recognition. This is accomplished by training on two objectives: the first, inspired by adapting BERT to the continuous domain, involves masking spans of input frames and reconstructing the whole sequence for acoustic representation learning; the second, inspired by the success of bottleneck features from ASR, is a sequence-level CTC loss applied to phoneme labels for phonetic representation learning. We pretrain two BERTphone models (one on Fisher and one on TED-LIUM) and use them as feature extractors into x-vector-style DNNs for both tasks. We attain a state-of-the-art $C_{\text{avg}}$ of 6.16 on the challenging LRE07 3sec closed-set language recognition task. On Fisher and VoxCeleb speaker recognition tasks, we see an 18% relative reduction in speaker EER when training on BERTphone vectors instead of MFCCs. In general, BERTphone outperforms previous phonetic pretraining approaches on the same data. We release our code and models at https://github.com/awslabs/speech-representations.

Se Jin Park, Julian Salazar, Aren Jansen et al.

We consider the generative modeling of speech over multiple minutes, a requirement for long-form multimedia generation and audio-native voice assistants. However, textless spoken language models struggle to generate plausible speech past tens of seconds, due to high temporal resolution of speech tokens causing loss of coherence, architectural issues with long-sequence training or extrapolation, and memory costs at inference time. From these considerations we derive SpeechSSM, the first speech language model family to learn from and sample long-form spoken audio (e.g., 16 minutes of read or extemporaneous speech) in a single decoding session without text intermediates. SpeechSSMs leverage recent advances in linear-time sequence modeling to greatly surpass current Transformer spoken LMs in coherence and efficiency on multi-minute generations while still matching them at the utterance level. As we found current spoken language evaluations uninformative, especially in this new long-form setting, we also introduce: LibriSpeech-Long, a benchmark for long-form speech evaluation; new embedding-based and LLM-judged metrics; and quality measurements over length and time. Speech samples, the LibriSpeech-Long dataset, and any future code or model releases can be found at https://google.github.io/tacotron/publications/speechssm/.

Alon Levkovitch, Julian Salazar, Soroosh Mariooryad et al. · meta-ai

We present ZeroBAS, a neural method to synthesize binaural audio from monaural audio recordings and positional information without training on any binaural data. To our knowledge, this is the first published zero-shot neural approach to mono-to-binaural audio synthesis. Specifically, we show that a parameter-free geometric time warping and amplitude scaling based on source location suffices to get an initial binaural synthesis that can be refined by iteratively applying a pretrained denoising vocoder. Furthermore, we find this leads to generalization across room conditions, which we measure by introducing a new dataset, TUT Mono-to-Binaural, to evaluate state-of-the-art monaural-to-binaural synthesis methods on unseen conditions. Our zero-shot method is perceptually on-par with the performance of supervised methods on the standard mono-to-binaural dataset, and even surpasses them on our out-of-distribution TUT Mono-to-Binaural dataset. Our results highlight the potential of pretrained generative audio models and zero-shot learning to unlock robust binaural audio synthesis.

Hoang H Nguyen, Khyati Mahajan, Vikas Yadav et al.

Although multilingual LLMs have achieved remarkable performance across benchmarks, we find they continue to underperform on non-Latin script languages across contemporary LLM families. This discrepancy arises from the fact that LLMs are pretrained with orthographic scripts, which are dominated by Latin characters that obscure their shared phonology with non-Latin scripts. We propose leveraging phonemic transcriptions as complementary signals to induce script-invariant representations. Our study demonstrates that integrating phonemic signals improves performance across both non-Latin and Latin script languages, with a particularly significant impact on closing the performance gap between the two. Through detailed experiments, we show that phonemic and orthographic scripts retrieve distinct examples for in-context learning (ICL). This motivates our proposed Mixed-ICL retrieval strategy, where further aggregation from both leads to our significant performance improvements for both Latin script languages (up to 12.6%) and non-Latin script languages (up to 15.1%) compared to randomized ICL retrieval.

Eric Battenberg, RJ Skerry-Ryan, Daisy Stanton et al.

Autoregressive (AR) Transformer-based sequence models are known to have difficulty generalizing to sequences longer than those seen during training. When applied to text-to-speech (TTS), these models tend to drop or repeat words or produce erratic output, especially for longer utterances. In this paper, we introduce enhancements aimed at AR Transformer-based encoder-decoder TTS systems that address these robustness and length generalization issues. Our approach uses an alignment mechanism to provide cross-attention operations with relative location information. The associated alignment position is learned as a latent property of the model via backpropagation and requires no external alignment information during training. While the approach is tailored to the monotonic nature of TTS input-output alignment, it is still able to benefit from the flexible modeling power of interleaved multi-head self- and cross-attention operations. A system incorporating these improvements, which we call Very Attentive Tacotron, matches the naturalness and expressiveness of a baseline T5-based TTS system, while eliminating problems with repeated or dropped words and enabling generalization to any practical utterance length.

Ethan A. Chi, Julian Salazar, Katrin Kirchhoff

Non-autoregressive models greatly improve decoding speed over typical sequence-to-sequence models, but suffer from degraded performance. Infilling and iterative refinement models make up some of this gap by editing the outputs of a non-autoregressive model, but are constrained in the edits that they can make. We propose iterative realignment, where refinements occur over latent alignments rather than output sequence space. We demonstrate this in speech recognition with Align-Refine, an end-to-end Transformer-based model which refines connectionist temporal classification (CTC) alignments to allow length-changing insertions and deletions. Align-Refine outperforms Imputer and Mask-CTC, matching an autoregressive baseline on WSJ at 1/14th the real-time factor and attaining a LibriSpeech test-other WER of 9.0% without an LM. Our model is strong even in one iteration with a shallower decoder.

Phillip Keung, Julian Salazar, Yichao Lu et al.

We describe an unsupervised method to create pseudo-parallel corpora for machine translation (MT) from unaligned text. We use multilingual BERT to create source and target sentence embeddings for nearest-neighbor search and adapt the model via self-training. We validate our technique by extracting parallel sentence pairs on the BUCC 2017 bitext mining task and observe up to a 24.5 point increase (absolute) in F1 scores over previous unsupervised methods. We then improve an XLM-based unsupervised neural MT system pre-trained on Wikipedia by supplementing it with pseudo-parallel text mined from the same corpus, boosting unsupervised translation performance by up to 3.5 BLEU on the WMT'14 French-English and WMT'16 German-English tasks and outperforming the previous state-of-the-art. Finally, we enrich the IWSLT'15 English-Vietnamese corpus with pseudo-parallel Wikipedia sentence pairs, yielding a 1.2 BLEU improvement on the low-resource MT task. We demonstrate that unsupervised bitext mining is an effective way of augmenting MT datasets and complements existing techniques like initializing with pre-trained contextual embeddings.

Phillip Keung, Yichao Lu, Julian Salazar et al.

Multilingual contextual embeddings have demonstrated state-of-the-art performance in zero-shot cross-lingual transfer learning, where multilingual BERT is fine-tuned on one source language and evaluated on a different target language. However, published results for mBERT zero-shot accuracy vary as much as 17 points on the MLDoc classification task across four papers. We show that the standard practice of using English dev accuracy for model selection in the zero-shot setting makes it difficult to obtain reproducible results on the MLDoc and XNLI tasks. English dev accuracy is often uncorrelated (or even anti-correlated) with target language accuracy, and zero-shot performance varies greatly at different points in the same fine-tuning run and between different fine-tuning runs. These reproducibility issues are also present for other tasks with different pre-trained embeddings (e.g., MLQA with XLM-R). We recommend providing oracle scores alongside zero-shot results: still fine-tune using English data, but choose a checkpoint with the target dev set. Reporting this upper bound makes results more consistent by avoiding arbitrarily bad checkpoints.

Phillip Keung, Wei Niu, Yichao Lu et al.

We discuss the problem of echographic transcription in autoregressive sequence-to-sequence attentional architectures for automatic speech recognition, where a model produces very long sequences of repetitive outputs when presented with out-of-domain utterances. We decode audio from the British National Corpus with an attentional encoder-decoder model trained solely on the LibriSpeech corpus. We observe that there are many 5-second recordings that produce more than 500 characters of decoding output (i.e. more than 100 characters per second). A frame-synchronous hybrid (DNN-HMM) model trained on the same data does not produce these unusually long transcripts. These decoding issues are reproducible in a speech transformer model from ESPnet, and to a lesser extent in a self-attention CTC model, suggesting that these issues are intrinsic to the use of the attention mechanism. We create a separate length prediction model to predict the correct number of wordpieces in the output, which allows us to identify and truncate problematic decoding results without increasing word error rates on the LibriSpeech task.

Shaoshi Ling, Yuzong Liu, Julian Salazar et al.

We propose a novel approach to semi-supervised automatic speech recognition (ASR). We first exploit a large amount of unlabeled audio data via representation learning, where we reconstruct a temporal slice of filterbank features from past and future context frames. The resulting deep contextualized acoustic representations (DeCoAR) are then used to train a CTC-based end-to-end ASR system using a smaller amount of labeled audio data. In our experiments, we show that systems trained on DeCoAR consistently outperform ones trained on conventional filterbank features, giving 42% and 19% relative improvement over the baseline on WSJ eval92 and LibriSpeech test-clean, respectively. Our approach can drastically reduce the amount of labeled data required; unsupervised training on LibriSpeech then supervision with 100 hours of labeled data achieves performance on par with training on all 960 hours directly. Pre-trained models and code will be released online.

Toan Q. Nguyen, Julian Salazar

We evaluate three simple, normalization-centric changes to improve Transformer training. First, we show that pre-norm residual connections (PreNorm) and smaller initializations enable warmup-free, validation-based training with large learning rates. Second, we propose $\ell_2$ normalization with a single scale parameter (ScaleNorm) for faster training and better performance. Finally, we reaffirm the effectiveness of normalizing word embeddings to a fixed length (FixNorm). On five low-resource translation pairs from TED Talks-based corpora, these changes always converge, giving an average +1.1 BLEU over state-of-the-art bilingual baselines and a new 32.8 BLEU on IWSLT'15 English-Vietnamese. We observe sharper performance curves, more consistent gradient norms, and a linear relationship between activation scaling and decoder depth. Surprisingly, in the high-resource setting (WMT'14 English-German), ScaleNorm and FixNorm remain competitive but PreNorm degrades performance.

Julian Salazar, Katrin Kirchhoff, Zhiheng Huang

The success of self-attention in NLP has led to recent applications in end-to-end encoder-decoder architectures for speech recognition. Separately, connectionist temporal classification (CTC) has matured as an alignment-free, non-autoregressive approach to sequence transduction, either by itself or in various multitask and decoding frameworks. We propose SAN-CTC, a deep, fully self-attentional network for CTC, and show it is tractable and competitive for end-to-end speech recognition. SAN-CTC trains quickly and outperforms existing CTC models and most encoder-decoder models, with character error rates (CERs) of 4.7% in 1 day on WSJ eval92 and 2.8% in 1 week on LibriSpeech test-clean, with a fixed architecture and one GPU. Similar improvements hold for WERs after LM decoding. We motivate the architecture for speech, evaluate position and downsampling approaches, and explore how label alphabets (character, phoneme, subword) affect attention heads and performance.