Meta tags:
author=Dongwei Jiang;
description=A simple, whitespace theme for academics. Based on [*folio](htt????/github.com/bogoli/-folio) design. ;
keywords=jekyll, jekyll-theme, academic-website, portfolio-website;

Headings (most frequently used words):

previously, at, dongwei, jiang, research, interest, news, more, about, me, selected, publications, llm, agents, reasoning, and, self, improvement, focused, on, speech, currently, applied, scientist, amazon, master, student, jhu,

Text of the page (most frequently used words):
and (67), the (65), reasoning (30), with (24), #speech (23), this (22), for (21), that (21), training (19), 2025 (18), self (15), llms (15), cot (15), feedback (15), from (14), models (14), jiang (13), dongwei (12), pre (12), can (11), these (11), data (10), tasks (10), entailment (10), recognition (9), model (9), pdf (9), abstract (9), performance (8), logical (8), which (7), based (7), paper (7), datasets (7), like (7), also (7), our (7), language (7), decompositional (7), symbolic (7), 2024 (7), using (6), supervised (6), learning (6), 2021 (6), conference (6), previously (6), lean (6), research (6), improvement (6), worked (6), unsupervised (5), method (5), applied (5), end (5), more (5), nov (5), among (5), simclr (5), objective (5), new (5), its (5), results (5), better (5), their (5), inference (5), dataset (5), help (5), solver (5), across (5), math (5), rationalyst (5), about (5), agents (5), are (4), industrial (4), however (4), large (4), problem (4), masked (4), transformer (4), cer (4), over (4), relative (4), reduction (4), wei (4), has (4), methods (4), representation (4), combination (4), between (4), been (4), mpc (4), not (4), knowledge (4), processing (4), zhang (4), need (4), generated (4), than (4), responses (4), proceedings (4), benjamin (4), van (4), durme (4), daniel (4), khashabi (4), textual (4), long (4), problems (4), llm (4), theorem (4), them (4), near (4), jul (4), papers (4), but (4), process (4), external (4), friction (4), was (4), there (4), accepted (4), various (3), building (3), predictive (3), coding (3), proposed (3), show (3), further (3), wubo (3), zou (3), xiangang (3), improving (3), pretraining (3), significant (3), state (3), art (3), input (3), natural (3), during (3), loss (3), reconstruction (3), achieved (3), international (3), association (3), have (3), fully (3), study (3), how (3), consistently (3), improve (3), discriminating (3), generating (3), challenges (3), aaai (3), correct (3), struggle (3), without (3), logic (3), address (3), find (3), rdte (3), accuracy (3), complex (3), proving (3), ability (3), fine (3), 100 (3), each (3), human (3), thinking (3), really (3), analyze (3), iclr (3), supervision (3), scale (3), including (3), wang (3), acl (3), when (3), incorporate (3), complete (3), neurips (3), where (3), company (3), work (3), throughout (3), working (3), scientist (3), amazon (3), 2026 (2), april (2), folio (2), technologies (2), enormous (2), applications (2), good (2), system (2), usually (2), requires (2), amounts (2), transcribed (2), expensive (2), collect (2), tackle (2), experiments (2), hkust (2), achieve (2), best (2), reduce (2), baseline (2), corr (2), luo (2), visual (2), shown (2), progress (2), representations (2), both (2), contrastive (2), augmented (2), samples (2), annual (2), communication (2), interspeech (2), miao (2), cao (2), many (2), improvements (2), aspects (2), three (2), speaking (2), style (2), streaming (2), transfer (2), downstream (2), matching (2), unified (2), error (2), target (2), discriminative (2), strong (2), icassp (2), june (2), han (2), would (2), alternatives (2), initial (2), first (2), formulate (2), framework (2), allows (2), any (2), experimental (2), analysis (2), finding (2), may (2), through (2), own (2), jingyu (2), orion (2), weller (2), nathaniel (2), weir (2), recent (2), such (2), proof (2), relying (2), standing (2), clear (2), protocol (2), what (2), gains (2), consistent (2), theoretically (2), grounded (2), approach (2), annotating (2), evaluate (2), prior (2), via (2), significantly (2), improves (2), quality (2), peter (2), inconsistencies (2), theorems (2), within (2), risk (2), extensive (2), achieves (2), domain (2), computational (2), linguistics (2), volume (2), naacl (2), mexico (2), shay (2), cohen (2), chain (2), thought (2), prompting (2), evaluations (2), benefits (2), much (2), following (2), execution (2), they (2), range (2), diego (2), greg (2), durrett (2), mainly (2), rationales (2), rationale (2), web (2), unlabelled (2), diverse (2), scientific (2), tuned (2), andrew (2), ideal (2), environment (2), answers (2), general (2), even (2), resistance (2), confidence (2), high (2), publications (2), career (2), companies (2), systems (2), stock (2), massive (2), china (2), free (2), time (2), swimming (2), every (2), all (2), activities (2), current (2), foundation (2), master (2), jhu (2), professor (2), currently (2), dec (2), security (2), agentic (2), jan (2), lilian (2), weng (2), blog (2), attending (2), come (2), say (2), limits (2), copyright, powered, theme, last, updated, jekyll, gaining, popularity, called, same, exceeding, absolute, 2019, xiaoning, lei, yuxuan, arxiv, recently, those, greatly, advanced, semi, imagenet, feature, continuous, consider, applying, similar, propose, applies, augmentation, raw, spectrogram, maximizes, agreement, differently, latent, space, competitive, emotion, 22nd, brno, czechia, august, september, combining, bert, backbone, investigated, conduct, focus, important, effect, extension, learned, stage, reveled, useful, apc, provided, trained, adaption, layer, wise, helped, aishell, ieee, acoustics, signal, toronto, canada, ruixiong, yang, kun, previous, outputs, true, explore, validity, hypothesis, practice, compare, generative, capability, task, resulting, several, open, source, observe, reliably, notion, able, enhance, only, judgment, artificial, intelligence, enable, opportunities, structured, text, construction, intuitive, trees, brittle, formal, direction, hampered, lack, determining, _valid, entailment_, absence, causes, noisy, limited, modern, neuro, engines, impact, recognizing, substantially, higher, internal, consistency, suggesting, step, forward, forming, discerning, oriented, classifier, distillation, employing, tree, engine, illustrating, practical, benefit, advance, empirical, kate, sanders, shreya, sharma, zhengping, bhavana, dalvi, mishra, oyvind, tafjord, jansen, clark, enhancing, systematic, informal, emnlp, often, due, inherent, difficulty, use, formalizing, into, solve, disproving, corresponding, reduces, enhances, treat, library, proofs, level, proofwriter, notably, were, accomplished, tuning, fewer, north, american, chapter, city, marcio, fonseca, leanreasoner, boosting, facto, eliciting, capabilities, kinds, extra, helpful, conducted, quantitative, meta, covering, ran, gives, primarily, involving, smaller, other, types, mmlu, directly, answer, leads, almost, identical, unless, question, response, contains, equals, sign, indicating, operations, behavior, separating, planning, comparing, against, tool, gain, comes, underperforms, indicate, selectively, maintaining, while, saving, costs, furthermore, suggest, move, beyond, prompt, paradigms, leverage, intermediate, computation, whole, thirteenth, singapore, zayne, rea, sprague, fangcong, yin, juan, rodriguez, manya, wadhwa, prasann, singhal, xinyu, zhao, kyle, mahowald, helps, steps, might, incomplete, mimic, leaps, common, everyday, found, underlying, frequently, left, implicit, unstated, challenge, introduce, vast, collection, annotations, extracted, unlabeled, extract, 79k, pile, minimal, intervention, generalize, mathematical, commonsense, llama, average, representative, benchmarks, demonstrates, superior, compared, larger, verifiers, gpt, similarly, sized, sets, 63rd, meeting, guoxuan, yining, chuyu, liu, studies, possess, some, given, remains, unclear, effectively, thoroughly, extrinsic, scenario, receive, perfect, expect, integrate, reach, solutions, systematically, investigate, designing, controlled, attempts, solution, then, generator, access, ground, truth, produces, targeted, after, tries, again, pipeline, multi, claude, extended, surprisingly, under, conditions, limitation, term, mitigate, limita, tion, experiment, sampling, strategies, progressive, temperature, increases, explicit, rejection, attempted, incorrect, yield, still, fail, specific, questions, measured, semantic, entropy, predicts, predictions, remain, resistant, correction, hope, highlighting, issue, will, future, alvin, nicholas, andrews, selected, marked, series, devastating, events, fundamentally, disrupted, didi, developing, hit, chinese, government, forcing, delisting, york, exchange, creating, operational, uncertainty, yuanfudao, abruptly, affected, essentially, crippled, core, business, educational, technology, later, shopee, global, economic, downturn, tensions, triggered, price, collapse, successive, corporate, disruptions, necessitated, transitions, roles, faced, existential, made, continuing, technical, untenable, layoffs, double, policy, severe, regulatory, action, sometimes, play, civ, hearthstone, rotate, tennis, badminton, bouldering, day, well, four, days, who, counting, shows, racquet, sports, mortality, either, achieving, immortality, just, bad, actually, something, puzzle, probably, explains, why, enjoy, alongside, role, spent, six, years, industry, pioneered, approaches, deploy, asr, production, release, chatgpt, recognized, potential, wanted, deeply, distractions, returned, academia, degree, collaborated, edinburgh, nyu, berkeley, projects, continue, pursue, related, areas, had, lot, people, journey, believe, giving, back, service, feel, book, calendar, dawn, song, joined, featured, mar, vienna, san, news, probing, prevents, effective, understanding, reason, begin, superintelligence, agi, mind, won, get, teach, interact, area, formulates, uses, prover, discovering, jason, discussion, investigating, effectiveness, extraction, provide, mentioned, purpose, verifier, realm, autonomously, identify, validate, vulnerabilities, extensively, nowadays, increasingly, interested, autoregressive, prediction, architecture, truly, bridge, gap, you, love, chat, themselves, interest, student, focused, ctrl, toggle, navigation,

Text of the page (random words):
o interact with the environment and self improve within this area i ve worked on understanding the reason that prevents llm from effective self improvement 5 probing the limits of self improvement even with high quality feedback 6 news dec 2025 attending neurips 2025 in san diego come say hi jul 2025 attending acl 2025 in vienna come say hi may 2025 paper feedback friction accepted to neurips 2025 mar 2025 paper rationalyst accepted to acl 2025 featured in lilian weng s blog jan 2025 paper to cot or not to cot accepted to iclr 2025 jan 2025 joined amazon as applied scientist ii working on security agents and agentic rl dec 2024 paper self in correct accepted to aaai 2025 more about me prior to my current role i spent six years in industry working on speech processing where i pioneered self supervised learning approaches for speech like masked predictive coding 7 and speech simclr 8 and was among the first to deploy end to end asr systems at production scale following the release of chatgpt i recognized the enormous potential of foundation models and wanted to study them deeply without distractions so i returned to academia to complete my master s degree at jhu there i worked with professor daniel khashabi and benjamin van durme and also collaborated with professor shay cohen from edinburgh greg durrett from nyu and dawn song from uc berkeley on various research projects currently i m working as an applied scientist at amazon where i continue to pursue research in foundation models and related areas i ve had help from a lot of people throughout my journey and i believe in giving it back if i can be of service feel free to book time on my calendar in my free time i sometimes play civ 6 or hearthstone i also rotate between tennis badminton swimming and bouldering every day well more like every three or four days but who s counting research shows racquet sports can reduce mortality risk by 47 and swimming by 28 so between all these activities i m either achieving immortality or just really bad at math there s actually something puzzle like about all these activities which probably explains why i enjoy them alongside my research work more about my industrial career my industrial career has been marked by a series of devastating external events that fundamentally disrupted the companies where i worked at didi i was developing speech processing systems when the company was hit by severe regulatory action from the chinese government forcing its delisting from the new york stock exchange and creating massive operational uncertainty at yuanfudao my work was abruptly affected when the double reduction policy essentially crippled the core business model of educational technology companies throughout china later at shopee the combination of global economic downturn and us china tensions triggered an 80 stock price collapse and extensive layoffs throughout the company these successive corporate disruptions necessitated my transitions between roles as each company faced existential challenges that made continuing my technical work there untenable selected publications feedback friction llms struggle to fully incorporate external feedback neurips dongwei jiang alvin zhang andrew wang nicholas andrews and daniel khashabi corr 2025 abstract pdf 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 limita tion 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 rationalyst pre training process supervision for improving reasoning acl dongwei jiang guoxuan wang yining lu andrew wang jingyu zhang chuyu liu benjamin van durme and daniel khashabi in proceedings of the 63rd annual meeting of the association for computational linguistics volume 1 long papers jul 2025 abstract pdf the reasoning steps generated by llms might be incomplete as they mimic logical leaps common in everyday communication found in their pre training data underlying rationales are frequently left implicit unstated to address this challenge we introduce rationalyst a model for process supervision of reasoning based on pre training on a vast collection of rationale annotations extracted from unlabeled data we extract 79k rationales from web scale unlabelled dataset the pile and a combination of reasoning datasets with minimal human intervention this web scale pre training for reasoning allows rationalyst to consistently generalize across diverse reasoning tasks including mathematical commonsense scientific and logical reasoning fine tuned from llama 3 8b rationalyst improves the accuracy of reasoning by an average of 3 9 on 7 representative reasoning benchmarks it also demonstrates superior performance compared to significantly larger verifiers like gpt 4 and similarly sized models fine tuned on matching training sets to cot or not to cot chain of thought helps mainly on math and symbolic reasoning iclr zayne rea sprague fangcong yin juan diego rodriguez dongwei jiang manya wadhwa prasann singhal xinyu zhao xi ye kyle mahowald and greg durrett in the thirteenth international conference on learning representations iclr 2025 singapore april 24 28 2025 jul 2025 abstract pdf 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 leanreasoner boosting complex logical reasoning with lean naacl dongwei jiang marcio fonseca and shay b cohen in proceedings of the 2024 conference of the north american chapter of the association for computational linguistics human language technologies volume 1 long papers naacl 2024 mexico city mexico june 16 21 2024 jul 2024 abstract pdf large language models llms often struggle with complex logical reasoning due to logical inconsistencies and the inherent difficulty of such reasoning we use lean a theorem proving framework to address these challenges by formalizing logical reasoning problems into theorems within lean we can solve them by proving or disproving the corresponding theorems this method reduces the risk of logical inconsistencies with the help of lean s symbolic solver it also enhances our ability to treat complex reasoning tasks by using lean s extensive library of theorem proofs our method achieves state of the art performance on the folio dataset and achieves performance near this level on proofwriter notably these results were accomplished by fine tuning on fewer than 100 in domain samples for each dataset enhancing systematic decompositional natural language inference using informal logic emnlp nathaniel weir kate sanders orion weller shreya sharma dongwei jiang zhengping jiang bhavana dalvi mishra oyvind tafjord peter jansen peter clark and benjamin van durme in proceedings of the 2024 conference on empirical methods in natural language processing nov 2024 abstract pdf recent language models enable new opportunities for structured reasoning with text such as the construction of intuitive proof like textual entailment trees without relying on brittle formal logic however progress in this direction has been hampered by a long standing lack of a clear protocol for determining what _valid decompositional entailment_ is this absence causes noisy datasets and limited performance gains by modern neuro symbolic entailment engines to address these problems we formulate a consistent and theoretically grounded approach to annotating decompositional entailment and evaluate its impact on llm based textual inference we find that our new dataset rdte recognizing decompositional textual entailment has a substantially higher internal consistency than prior decompositional entailment datasets suggesting that rdte is a significant step forward in the long standing problem of forming a clear protocol for discerning entailment we also find that training an rdte oriented entailment classifier via knowledge distillation and employing it in an entailment tree reasoning engine significantly improves both accuracy and proof quality illustrating the practical benefit of this advance for textual inference self in correct llms struggle with discriminating self generated responses aaai dongwei jiang jingyu zhang orion weller nathaniel weir benjamin van durme and daniel khashabi in proceedings of the aaai conference on artificial intelligence nov 2025 abstract pdf 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 models 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 a further study of unsupervised pretraining for transformer based speech recognition icassp dongwei jiang wubo li ruixiong zhang miao cao ne luo yang han wei zou kun han and xiangang li in ieee international conference on acoustics speech and signal processing icassp 2021 toronto on canada june 6 11 2021 nov 2021 abstract pdf building a good speech recognition system usually requires large amounts of transcribed data which is expensive to collect to tackle this problem many unsupervised pre training 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 not been fully investigated in this paper we conduct a further study on mpc and focus on three important aspects the effect of pre training data speaking style its extension on streaming model and how to better transfer learned knowledge from pre training stage to downstream tasks experiments reveled that pre training data with matching speaking style is more useful on downstream recognition tasks a unified training objective with apc and mpc provided 8 46 relative error reduction on streaming model trained on hkust also the combination of target data adaption and layer wise discriminative training helped the knowledge transfer of mpc which achieved 3 99 relative error reduction on aishell over a strong baseline speech simclr combining contrastive and reconstruction objective for self supervised speech representation learning interspeech dongwei jiang wubo li miao cao wei zou and xiangang li in 22nd annual conference of the international speech communication association interspeech 2021 brno czechia august 30 september 3 2021 nov 2021 abstract pdf self supervised visual pretraining has shown significant progress recently among those methods simclr greatly advanced the state of the art in self supervised and semi supervised 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 improving transformer based speech recognition using unsupervised pre training arxiv dongwei jiang xiaoning lei wubo li ne luo yuxuan hu wei zou and xiangang li corr nov 2019 abstract pdf speech recognition technologies are gaining enormous popularity in various industrial applications however building a good speech recognition system usually requires large amounts of transcribed data which is expensive to collect to tackle this problem an unsupervised pre training method called masked predictive coding is proposed which can be applied for unsupervised pre training with transformer based model experiments on hkust show that using the same training data we can achieve cer 23 3 exceeding the be...