Can LLMs consistently improve their previous outputs for better results? For this to be true, LLMs would need to be better at discriminating among previously-generated alternatives, than generating initial responses. We explore the validity of this hypothesis in practice. We first formulate a unified framework that allows us to compare the generative and discriminative capability of any model on any task. In our resulting experimental analysis of several open-source and industrial LLMs, we observe that model’s are not reliably better at discriminating among previously-generated alternatives than generating initial responses. This finding challenges the notion that LLMs may be able to enhance their performance only through their own judgment.

Dongwei Jiang, Guoxuan Wang, Yining Lu, Andrew Wang, Jingyu Zhang, Chuyu Liu, Benjamin Van Durme, Daniel Khashabi. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Chain-of-thought (CoT) via prompting is the de facto method for eliciting reasoning capabilities from large language models (LLMs). But for what kinds of tasks is this extra "thinking" really helpful? To analyze this, we conducted a quantitative meta-analysis covering over 100 papers using CoT and ran our own evaluations of 20 datasets across 14 models. Our results show that CoT gives strong performance benefits primarily on tasks involving math or logic, with much smaller gains on other types of tasks. On MMLU, directly generating the answer without CoT leads to almost identical accuracy as CoT unless the question or model's response contains an equals sign, indicating symbolic operations and reasoning. Following this finding, we analyze the behavior of CoT on these problems by separating planning and execution and comparing against tool-augmented LLMs. Much of CoT's gain comes from improving symbolic execution, but it underperforms relative to using a symbolic solver. Our results indicate that CoT can be applied selectively, maintaining performance while saving inference costs. Furthermore, they suggest a need to move beyond prompt-based CoT to new paradigms that better leverage intermediate computation across the whole range of LLM applications.

Yining Lu, Dixuan Wang, Tianjian Li, Dongwei Jiang, Sanjeev Khudanpur, Meng Jiang, Daniel Khashabi. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Recent studies have shown LLMs possess some ability to improve their responses when given external feedback. However, it remains unclear how effectively and thoroughly these models can incorporate extrinsic feedback. In an ideal scenario, if LLMs receive near-perfect and complete feedback, we would expect them to fully integrate the feedback and reach correct solutions. In this paper, we systematically investigate LLMs’ ability to incorporate feedback by designing a controlled experimental environment. For each problem, a solver model attempts a solution, then a feedback generator with access to near-complete ground-truth answers produces targeted feedback, after which the solver tries again. We evaluate this pipeline across a diverse range of tasks, including math reasoning, knowledge reasoning, scientific reasoning, and general multi-domain evaluations with state-of-the-art language models including Claude 3.7 with extended thinking. Surprisingly, even under these near-ideal conditions, solver models consistently show resistance to feedback, a limitation that we term FEEDBACK FRICTION. To mitigate this limitation, we experiment with sampling-based strategies like progressive temperature increases and explicit rejection of previously attempted incorrect answers, which yield improvements but still fail to help models achieve target performance. We analyze FEEDBACK FRICTION and find that models’ confidence on specific questions, measured by semantic entropy, predicts feedback resistance: high-confidence predictions remain resistant to external correction. We hope that highlighting this issue in LLMs will help future research in self-improvement.

Nathaniel Weir, Kate Sanders, Orion Weller, Shreya Sharma, Dongwei Jiang, Zhengping Jiang, Bhavana Dalvi Mishra, Oyvind Tafjord, Peter Jansen, Peter Clark, Benjamin Van Durme. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

Dongwei Jiang, Marcio Fonseca, Shay Cohen. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

This paper introduces a new open-sourced Mandarin speech corpus, called DiDiSpeech. It consists of about 800 hours of speech data at 48kHz sampling rate from 6000 speakers and the corresponding texts. All speech data in the corpus is recorded in quiet environment and is suitable for various speech processing tasks, such as voice conversion, multi-speaker text-to-speech and automatic speech recognition. We conduct experiments with multiple speech tasks and evaluate the performance, showing that it is promising to use the corpus for both academic research and practical application. The corpus is available at https://outreach.didichuxing.com/research/opendata/.

The construction of an effective good speech recognition system typically requires large amounts of transcribed data, which is expensive to collect. To overcome this problem, many unsupervised pretraining methods have been proposed. Among these methods, Masked Predictive Coding achieved significant improvements on various speech recognition datasets with BERT-like Masked Reconstruction loss and transformer backbone. However, many aspects of MPC have yet to be fully investigated. In this paper, we conduct a further study on MPC and focus on three important aspects: the effect of pretraining data speaking style, its extension on streaming model, and strategies for better transferring learned knowledge from pretraining stage to downstream tasks. The experimental results demonstrated that pretraining data with a matching speaking style is more useful on downstream recognition tasks. A unified training objective with APC and MPC provided an 8.46% relative error reduction on the streaming model trained on HKUST. Additionally, the combination of target data adaption and layerwise discriminative training facilitated the knowledge transfer of MPC, which realized 3.99% relative error reduction on AISHELL over a strong baseline.

Recently, a variety of acoustic tasks and related applications arised. For many acoustic tasks, the labeled data size may be limited. To handle this problem, we propose an unsupervised pre-training method using Transformer based encoder to learn a general and robust high-level representation for all acoustic tasks. Experiments have been conducted on three kinds of acoustic tasks: speech emotion recognition, sound event detection and speech translation. All the experiments have shown that pre-training using its own training data can significantly improve the performance. With a larger pre-training data combining MuST-C, Librispeech and ESC-US datasets, for speech emotion recognition, the UAR can further improve absolutely 4.3% on IEMOCAP dataset. For sound event detection, the F1 score can further improve absolutely 1.5% on DCASE2018 task5 development set and 2.1% on evaluation set. For speech translation, the BLEU score can further improve relatively 12.2% on En-De dataset and 8.4% on En-Fr dataset.

Self-supervised visual pretraining has shown significant progress recently.Among those methods, SimCLR greatly advanced the state of the art in self-supervised and semisupervised learning on ImageNet.The input feature representations for speech and visual tasks are both continuous, so it is natural to consider applying similar objective on speech representation learning.In this paper, we propose Speech SimCLR, a new self-supervised objective for speech representation learning.During training, Speech SimCLR applies augmentation on raw speech and its spectrogram.Its objective is the combination of contrastive loss that maximizes agreement between differently augmented samples in the latent space and reconstruction loss of input representation.The proposed method achieved competitive results on speech emotion recognition and speech recognition.

Audio Visual Scene-aware Dialog (AVSD) is a task to generate responses when discussing about a given video. The previous state-of-the-art model shows superior performance for this task using Transformer-based architecture. However, there remain some limitations in learning better representation of modalities. Inspired by Neural Machine Translation (NMT), we propose the Transformer-based Modal Translator (TMT) to learn the representations of the source modal sequence by translating the source modal sequence to the related target modal sequence in a supervised manner. Based on Multimodal Transformer Networks (MTN), we apply TMT to video and dialog, proposing MTN-TMT for the video-grounded dialog system. On the AVSD track of the Dialog System Technology Challenge 7, MTN-TMT outperforms the MTN and other submission models in both Video and Text task and Text Only task. Compared with MTN, MTN-TMT improves all metrics, especially, achieving relative improvement up to 14.1% on CIDEr. Index Terms: multimodal learning, audio-visual scene-aware dialog, neural machine translation, multi-task learning