Pith. sign in

REVIEW 4 major objections 6 minor 28 references

DIRECT trains LLMs with preference optimization then forces label-only decoding so sequence labeling is both more accurate and up to nine times faster.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-30 17:57 UTC pith:LCPTCVHO

load-bearing objection Solid engineering recipe for few-shot sequence labeling, but the SOTA F1 claim is not cleanly isolated from newer backbones. the 4 major comments →

arxiv 2607.26891 v1 pith:LCPTCVHO submitted 2026-07-29 cs.CL

DIRECT: Direct Decoding for Efficient and Aligned Sequence Labeling with Large Language Models

Yilei Wang , Jiaxin Gan , Kexuan Zhang , Ling Li , Wentao Zhang , Peichao Lai This is my paper
classification cs.CL
keywords Large Language ModelsFew-shot LearningSequence LabelingDirect Preference OptimizationNamed Entity RecognitionPart-of-Speech TaggingControlled DecodingKV Cache
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Sequence labeling asks a model to tag every word (named entities, parts of speech, and the like). Existing large-language-model methods often drift from the required format and run slowly because they regenerate whole sentences. DIRECT first fine-tunes the model, then runs Direct Preference Optimization on pairs that contrast high-BLEU but low-F1 outputs against the true labels, so the model learns what a correct tag sequence looks like. At inference it locks the output to a fixed template and only lets the model emit label tokens, reusing the rest of the text through the key-value cache. On eight NER and POS datasets in low-resource settings the method reaches best or second-best F1 while cutting inference time dramatically. A sympathetic reader cares because the same recipe turns ordinary chat models into reliable, fast taggers without rewriting the whole generation stack.

Core claim

After supervised fine-tuning, Direct Preference Optimization on carefully chosen preference pairs plus a constrained template-filling decoder that generates only label tokens yields both stronger domain alignment and up to 9× faster inference than prior LLM sequence-labeling methods on eight NER and POS datasets.

What carries the argument

Template-filling controlled decoding: the model is forced to emit only standardized label tokens inside a fixed word(label) skeleton while known prefixes are prefilled and reused via the KV cache, combined with DPO after SFT on preference pairs that favor ground-truth labels over high-BLEU/low-F1 rollouts.

Load-bearing premise

The reported gains come from the DPO-plus-rectification recipe itself rather than from simply using stronger, newer backbone models than the baselines.

What would settle it

Re-run the same eight-dataset, K=250/500/1000 protocol with every baseline also using LLaMA-3.1-8B-Instruct or GLM-4-9B-Chat; if the F1 and speed gaps shrink to noise, the method claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Low-resource NER and POS can be served from ordinary chat LLMs with strict format guarantees and far lower latency.
  • Preference pairs built from BLEU-versus-F1 disagreement become a practical signal for aligning generative taggers.
  • Any token-classification task whose labels fit a short fixed vocabulary can reuse the same label-only KV-cache decoding pattern.
  • Deployments that previously avoided generative IE for speed reasons become practical once only label tokens are generated.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same constrained label-only decode could be dropped onto larger open models without further DPO if the candidate set is small enough to mask logits at every step.
  • Preference construction that pits surface fluency against span F1 may transfer to other structured-generation tasks where format drift is the main failure mode.
  • If backbone strength is later shown to dominate, the lasting contribution is still the decoding recipe rather than the training recipe.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes DIRECT, a two-stage framework for low-resource sequence labeling (NER and POS) with LLMs. After supervised fine-tuning, it applies Direct Preference Optimization on preference pairs built by offline sampling: ground truth is preferred, while high-BLEU/low-F1 rollouts are treated as less-preferred (Eqs. 1–3). At inference it enforces a fixed word(label) template, restricts each label token to a candidate vocabulary, and prefills known prefixes (words, parentheses) so that only label tokens are generated while reusing the KV cache (Eq. 4, Fig. 1). Experiments on six NER and two POS datasets with K ∈ {250, 500, 1000} report best or second-best F1 using LLaMA-3.1-8B-Instruct and GLM-4-9B-Chat, plus substantially lower wall-clock latency than InstructUIE, GoLLIE, and GNER (Table 1, Fig. 2). Ablations remove DPO or keep only SFT (Table 2); case studies illustrate format failures of baselines (Fig. 3).

Significance. If the gains are truly attributable to DPO-after-SFT plus constrained template-filling decoding, the work is a useful systems contribution: it jointly targets domain alignment and decoding efficiency for a practically important IE setting, and the KV-cache template-filling idea is concrete and deployable. The efficiency numbers (roughly an order of magnitude vs. the reported baselines under batch size 1) would matter for interactive or high-throughput labeling. Credit is due for multi-dataset coverage (Chinese and English NER/POS), explicit ablations of DPO vs. SFT-only, and qualitative cases that show format control. Significance is currently limited by unmatched backbone comparisons, so the method-specific advance over prior LLM IE recipes is not yet cleanly established.

major comments (4)
  1. [§3.1, Table 1, Table 2] The central SOTA claim (abstract, §1, §3.2, Table 1) attributes large F1 lifts—e.g. +13–15 on MIT-Movie—to DIRECT’s DPO-plus-rectification recipe, but baselines remain on Flan-T5/mT5-xxl, Code-LLaMA-7B, and LLaMA-7B while DIRECT uses LLaMA-3.1-8B-Instruct and GLM-4-9B-Chat (§3.1). Table 2 “w/ SFT” rows on the newer backbones already exceed several baseline scores (e.g. Taobao, CoNLL03, Resume), so backbone strength is a plausible primary driver. Without re-running InstructUIE/GoLLIE/GNER (or equivalent SFT+generative baselines) on the same LLaMA-3.1/GLM-4 checkpoints, or running DIRECT’s full pipeline on the baseline backbones, the method contribution is not isolated. This control is load-bearing for the comparative claim.
  2. [§3.4, Fig. 2] Inference efficiency (Fig. 2, §3.4) reports ~9× speedup vs. GNER and larger gaps vs. InstructUIE/GoLLIE, but timing is cross-model (different architectures, sizes, and decoding stacks) on 10 CTB6 sentences, batch size 1, single L40. Template-filling + KV prefilling is a plausible source of speedup, yet the figure cannot support a method-level efficiency claim until the same backbone is timed with vs. without the constrained template decoder (and, ideally, against a same-model unconstrained autoregressive baseline). As written, model choice and decoding recipe are confounded.
  3. [§3.3, Table 2] Ablations (Table 2) contrast full DIRECT vs. “w/o DPO” vs. “w/ SFT” but do not factorially isolate inference rectification (format lock + candidate restriction + template prefilling) from DPO. “w/ SFT” removes both DPO and inference intervention, so the large drops (e.g. MIT-Movie 88.03→67.82 on LLaMA) cannot be attributed to either component alone. A same-backbone 2×2 (SFT vs. SFT+DPO × free decoding vs. DIRECT decoding), with F1 and latency, is needed to support the dual training-time / inference-time story in §2.2–2.3.
  4. [§3.2, Table 1] No multiple random seeds, error bars, or significance tests are reported for Table 1 despite small-K sampling (K=250/500/1000). Several margins over the second-best system are <1 F1 (e.g. CoNLL03, UD, Youku under some K). Without variance estimates it is unclear which “best/second-best” rankings are stable. At minimum, report mean±std over 3 seeds for the main backbone and note how K-subsets are drawn.
minor comments (6)
  1. [Table 1, Table 2] Table 1 vs Table 2 inconsistency: Resume K=1000 DIRECT(LLaMA) is 95.13 in Table 1 but 93.13 in Table 2; please reconcile all shared cells.
  2. [§2.2] Preference construction (§2.2): selecting highest-BLEU / lowest-F1 rollouts as O− is underspecified (how many samples per X; ties; what if F1 is already high). A short sensitivity check or example pair would help.
  3. [§2.3] Label standardization to fixed length T and padding tokens such as O__, PER_ (§2.3, Fig. 3) should be stated explicitly (padding character, handling of multi-token labels, effect on candidate set V).
  4. [Fig. 2] Fig. 2 y-axis is total seconds on 10 sentences; also report tokens/s or ms/sentence and confirm whether baseline implementations use comparable KV-cache settings.
  5. [§1, §2.3] Minor writing issues: spacing in “DPO-basedInferenceRECTification” (§1); “we first to-kenize” line break (§2.3); Index Terms lists Few-shot Learning though the method is fine-tuning with K-shot data rather than pure ICL.
  6. [§3.1] Baselines omit stronger recent same-era open models; even after backbone-matched runs, briefly situating against other constrained-decoding or IE-tuned LLMs would clarify novelty of the template-filling design.

Circularity Check

0 steps flagged

No circularity: empirical F1 and wall-clock gains are external measurements, not forced by definition or self-citation.

full rationale

DIRECT is a standard empirical NLP methods paper. Its load-bearing claims are (i) higher entity/POS F1 on eight held-out datasets after SFT+DPO plus constrained decoding, and (ii) lower wall-clock inference via template-filling and KV-cache reuse. Both quantities are measured against external gold labels and a stopwatch; neither is algebraically identical to a fitted training input. Preference-pair construction (high-BLEU/low-F1 rollouts as O−, gold as O+) and the DPO loss (Eqs. 1–3) are ordinary training objectives; test F1 is not a rewrite of those objectives. Controlled decoding restricts the token vocabulary to label candidates and enforces a fixed word(label) template, which can legitimately raise format-valid F1, but the metric remains comparison to human annotation, not a tautology. There are no self-citations that supply a uniqueness theorem, no ansatz imported from the authors’ prior work as external fact, and no renaming of a known closed-form result. Backbone mismatch vs. baselines is a validity/confound concern, not circularity. Derivation chain is self-contained against external benchmarks; score 0.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 2 invented entities

Central claims rest on standard LLM fine-tuning assumptions, the DPO preference model, and the engineering premise that forcing label tokens inside a fixed template with KV reuse preserves accuracy while cutting compute. Free choices include β, K-shot sizes, the BLEU/F1 less-preferred selection rule, and backbone selection. No new physical entities; the ‘DIRECT’ stack is a methodological bundle.

free parameters (4)
  • DPO beta = 0.1
    Scaling hyperparameter in the DPO loss (Eq. 3); set to 0.1 with no sweep reported.
  • few-shot K = 250, 500, 1000
    Training subset sizes chosen by authors for low-resource protocol.
  • preference pair selection rule = max BLEU & min F1 → O-
    Offline rule: ground truth as preferred; among rollouts, highest BLEU and lowest F1 as less-preferred. Hand-designed, not derived.
  • fixed label token length T
    Labels are standardized to fixed-length token sequences for constrained decoding; T is an implementation choice affecting candidate sets.
axioms (4)
  • domain assumption DPO with preference pairs improves alignment of generative LLMs to human-preferred sequence-labeling outputs beyond SFT alone.
    Invoked in §2.2 as the training-time fix for insufficient domain alignment; supported by ablation but assumed transferable from general DPO literature.
  • domain assumption Restricting next-token distribution to a predefined label candidate set and fixed ‘word(label)’ template does not harm (and can improve) task F1 relative to free generation.
    Core of inference rectification §2.3; accuracy claims depend on this not introducing systematic label errors.
  • standard math KV-cache prefilling of known template tokens yields correct conditional probabilities identical to full autoregressive generation of the same string.
    Standard transformer caching identity used to justify efficiency without quality loss (§2.3).
  • domain assumption F1 on entity/span extraction and token tags is the appropriate primary measure of sequence-labeling success under the stated output format.
    Used throughout §3; format enforcement makes string-level metrics secondary.
invented entities (2)
  • DIRECT framework (SFT→DPO + controlled template-filling decoding) no independent evidence
    purpose: Name the combined training and inference recipe claimed to fix alignment and efficiency for LLM sequence labeling.
    Bundle of known components; no independent existence outside this paper’s method definition.
  • BLEU/F1 offline less-preference sampler no independent evidence
    purpose: Construct O- responses that maximize entropy gain for DPO on structured outputs.
    Paper-specific data-selection heuristic in §2.2 without external validation.

pith-pipeline@v1.2.0-daily-grok45 · 14291 in / 3273 out tokens · 65742 ms · 2026-07-30T17:57:18.872172+00:00 · methodology

0 comments
read the original abstract

Sequence labeling is a fine-grained information extraction task, yet existing large language model-based approaches suffer from insufficient domain alignment and low inference efficiency. To address these issues, we propose DIRECT, a framework that addresses these issues through training-time optimization and inference-time rectification. Specifically, DIRECT performs Direct Preference Optimization (DPO) after supervised fine-tuning to strengthen task alignment with human preferences, and introduces a controlled decoding process that enforces fixed output formats and restricts predictions to candidate sets. To further improve efficiency, a template-filling mechanism requires the model to generate only label tokens while reusing prefixed content through the KV Cache, thus reducing redundant computation. Experimental results on eight datasets demonstrate that DIRECT achieves significant improvements in both performance and efficiency compared to existing methods.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

28 extracted references · 2 canonical work pages

  1. [1]

    It is crucial for other downstream natural language processing (NLP) tasks, such as co-reference resolution [1], knowledge graph [2], and question gen- eration [3]

    INTRODUCTION Sequence labeling is a fine-grained information extraction (IE) task that covers sub-tasks such as named entity recognition (NER), word segmentation, and part-of-speech (POS) tagging. It is crucial for other downstream natural language processing (NLP) tasks, such as co-reference resolution [1], knowledge graph [2], and question gen- eration ...

  2. [2]

    Task Definition Given a token sequence of lengthn,X={x 1, x2,

    METHODOLOGY 2.1. Task Definition Given a token sequence of lengthn,X={x 1, x2, . . . , xn}, the objective of sequence labeling is to map it into a label se- quence of the same length,Y={y 1, y2, . . . , yn}. In this study, arXiv:2607.26891v1 [cs.CL] 29 Jul 2026 Prompt Instruction: Please assign the appropriate label to each word in the given input sentenc...

    Pith/arXiv arXiv 2026

  3. [3]

    Experimental Setup DatasetsWe evaluate the performance of our model on six NER datasets and two POS datasets

    EXPERIMENTS 3.1. Experimental Setup DatasetsWe evaluate the performance of our model on six NER datasets and two POS datasets. The NER datasets include four Chi- nese datasets Weibo [14], Youku [15], Taobao [15], and Resume [16], as well as two English datasets MIT-Movie [17] and CoNLL03 [18]. The POS datasets include two Chinese datasets, UD [19] and CTB...

  4. [4]

    CONCLUSION In this paper, we propose DIRECT, a DPO-based inference recti- fication method. To enhance the alignment capability of LLMs in domain adaptation, we perform DPO optimization with preference pairs during training and intervene in the generation process dur- ing inference, ensuring that the outputs of LLMs are better aligned with human preference...

  5. [5]

    Chain-of- thought prompting elicits reasoning in large language models,

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al., “Chain-of- thought prompting elicits reasoning in large language models,” Advances in neural information processing systems, vol. 35, pp. 24824–24837, 2022

    2022

  6. [6]

    Improving coref- erence resolution by learning entity-level distributed represen- tations,

    Kevin Clark and Christopher D Manning, “Improving coref- erence resolution by learning entity-level distributed represen- tations,” inProceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2016, pp. 643–653, https://doi.org/10.18653/v1/P16- 1061

    doi:10.18653/v1/p16- 2016

  7. [7]

    A survey on knowledge graphs: Representation, acquisition, and applications,

    Shaoxiong Ji, Shirui Pan, Erik Cambria, Pekka Marttinen, and S Yu Philip, “A survey on knowledge graphs: Representation, acquisition, and applications,”IEEE transactions on neural networks and learning systems, vol. 33, no. 2, pp. 494–514, 2021

    2021

  8. [8]

    Boosting low-resource biomedical qa via entity- aware masking strategies,

    Gabriele Pergola, Elena Kochkina, Lin Gui, Maria Liakata, and Yulan He, “Boosting low-resource biomedical qa via entity- aware masking strategies,” inProceedings of the 16th Con- ference of the European Chapter of the Association for Com- putational Linguistics: Main Volume, 2021, pp. 1977–1985, https://doi.org/10.18653/v1/2021.eacl-main.169

    work page doi:10.18653/v1/2021.eacl-main.169 2021

  9. [9]

    Training language models to follow instructions with human feedback,

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agar- wal, Katarina Slama, Alex Ray, et al., “Training language models to follow instructions with human feedback,”Advances in neural information processing systems, vol. 35, pp. 27730– 27744, 2022

    2022

  10. [10]

    Rethinking negative instances for generative named entity recognition,

    Yuyang Ding, Juntao Li, Pinzheng Wang, Zecheng Tang, Bowen Yan, and Min Zhang, “Rethinking negative instances for generative named entity recognition,”arXiv preprint arXiv:2402.16602, 2024

    Pith/arXiv arXiv 2024

  11. [11]

    Graphusion: Leveraging large language models for scientific knowledge graph fusion and construction in nlp education,

    Rui Yang, Boming Yang, Sixun Ouyang, Tianwei She, Aosong Feng, Yuang Jiang, Freddy Lecue, Jinghui Lu, and Irene Li, “Graphusion: Leveraging large language models for scientific knowledge graph fusion and construction in nlp education,” arXiv preprint arXiv:2407.10794, 2024

    Pith/arXiv arXiv 2024

  12. [12]

    Unified structure gen- eration for universal information extraction,

    Yaojie Lu, Qing Liu, Dai Dai, Xinyan Xiao, Hongyu Lin, Xi- anpei Han, Le Sun, and Hua Wu, “Unified structure gen- eration for universal information extraction,”arXiv preprint arXiv:2203.12277, 2022

    Pith/arXiv arXiv 2022

  13. [13]

    Universal information extraction as unified semantic matching,

    Jie Lou, Yaojie Lu, Dai Dai, Wei Jia, Hongyu Lin, Xianpei Han, Le Sun, and Hua Wu, “Universal information extraction as unified semantic matching,” inProceedings of the AAAI conference on Artificial Intelligence, 2023, vol. 37, pp. 13318– 13326

    2023

  14. [14]

    Instructuie: Multi-task instruction tuning for unified in- formation extraction,

    Xiao Wang, Weikang Zhou, Can Zu, Han Xia, Tianze Chen, Yuansen Zhang, Rui Zheng, Junjie Ye, Qi Zhang, Tao Gui, et al., “Instructuie: Multi-task instruction tuning for unified in- formation extraction,”arXiv preprint arXiv:2304.08085, 2023

    Pith/arXiv arXiv 2023

  15. [15]

    Better modeling of incomplete annotations for named en- tity recognition,

    Zhanming Jie, Pengjun Xie, Wei Lu, Ruixue Ding, and Linlin Li, “Better modeling of incomplete annotations for named en- tity recognition,” inProceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), 2019, pp. 729–734

    2019

  16. [16]

    Gollie: An- notation guidelines improve zero-shot information-extraction,

    Oscar Sainz, Iker Garc ´ıa-Ferrero, Rodrigo Agerri, Oier Lopez de Lacalle, German Rigau, and Eneko Agirre, “Gollie: An- notation guidelines improve zero-shot information-extraction,” arXiv preprint arXiv:2310.03668, 2023

    Pith/arXiv arXiv 2023

  17. [17]

    C-icl: contrastive in- context learning for information extraction,

    Ying Mo, Jiahao Liu, Jian Yang, Qifan Wang, Shun Zhang, Jingang Wang, and Zhoujun Li, “C-icl: contrastive in- context learning for information extraction,”arXiv preprint arXiv:2402.11254, 2024

    Pith/arXiv arXiv 2024

  18. [18]

    Bleu: a method for automatic evaluation of machine translation,

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu, “Bleu: a method for automatic evaluation of machine translation,” inProceedings of the 40th annual meeting of the Association for Computational Linguistics, 2002, pp. 311–318

    2002

  19. [19]

    Named entity recogni- tion for chinese social media with jointly trained embed- dings,

    Nanyun Peng and Mark Dredze, “Named entity recogni- tion for chinese social media with jointly trained embed- dings,” inProceedings of the 2015 conference on empirical methods in natural language processing, 2015, pp. 548–554, https://doi.org/10.18653/v1/D15-1064

    work page doi:10.18653/v1/d15-1064 2015

  20. [20]

    The penn chinese treebank: Phrase structure annotation of a large corpus,

    Naiwen Xue, Fei Xia, Fu-Dong Chiou, and Marta Palmer, “The penn chinese treebank: Phrase structure annotation of a large corpus,”Natural language engineering, vol. 11, no. 2, pp. 207– 238, 2005

    2005

  21. [21]

    Chinese ner using lattice lstm,

    Yue Zhang and Jie Yang, “Chinese ner using lattice lstm,” arXiv preprint arXiv:1805.02023, 2018

    Pith/arXiv arXiv 2018

  22. [22]

    Asgard: A portable architecture for multilingual dialogue sys- tems,

    Jingjing Liu, Panupong Pasupat, Scott Cyphers, and Jim Glass, “Asgard: A portable architecture for multilingual dialogue sys- tems,” in2013 IEEE International Conference on Acoustics, Speech and Signal Processing. IEEE, 2013, pp. 8386–8390

    2013

  23. [23]

    Introduction to the conll- 2003 shared task: Language-independent named entity recog- nition,

    Erik F Sang and Fien De Meulder, “Introduction to the conll- 2003 shared task: Language-independent named entity recog- nition,”arXiv preprint cs/0306050, 2003

    Pith/arXiv arXiv 2003

  24. [24]

    Universal dependencies v1: A multilingual treebank collec- tion,

    Joakim Nivre, Marie-Catherine De Marneffe, Filip Ginter, Yoav Goldberg, Jan Hajic, Christopher D Manning, Ryan Mc- Donald, Slav Petrov, Sampo Pyysalo, Natalia Silveira, et al., “Universal dependencies v1: A multilingual treebank collec- tion,” inProceedings of the Tenth International Conference on Language Resources and Evaluation (LREC’16), 2016, pp. 1659–1666

    2016

  25. [26]

    The llama 3 herd of models,

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Ab- hishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al., “The llama 3 herd of models,”arXiv e-prints, pp. arXiv–2407, 2024

    2024

  26. [27]

    Chatglm: A family of large language models from glm-130b to glm-4 all tools,

    Team GLM, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Dan Zhang, Diego Rojas, Guanyu Feng, Han- lin Zhao, et al., “Chatglm: A family of large language models from glm-130b to glm-4 all tools,”arXiv preprint arXiv:2406.12793, 2024

    Pith/arXiv arXiv 2024

  27. [28]

    mt5: A massively multilingual pre-trained text-to-text trans- former,

    Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel, “mt5: A massively multilingual pre-trained text-to-text trans- former,”arXiv preprint arXiv:2010.11934, 2020

    Pith/arXiv arXiv 2010

  28. [1000]

    PLOT” entity or the “MISC

    The best results are highlighted inbold. Youku Taobao Weibo Resume DIRECT(LLaMA-3.1-8B-Instruct) 75.72 78.68 70.52 93.13 w/o DPO 75.37 78.34 68.66 94.94 w/ SFT 74.50 67.68 65.51 92.49 DIRECT(GLM4-9B-Chat) 79.31 78.58 70.31 95.94 w/o DPO 78.72 78.27 68.94 95.78 w/ SFT 78.48 70.52 60.82 92.56 CoNLL03 MIT-Movie UD CTB6 DIRECT(LLaMA-3.1-8B-Instruct) 90.10 88....