Retrieval-Constrained Policy Optimization for Attack Technique Extraction from Cyber Threat Intelligence

Jiayun Zhang; Junshen Xu; Yi Fan; Zejun Xie

REVIEW 3 major objections 4 minor 36 references

Retrieval-Constrained Policy Optimization for Attack Technique Extraction from Cyber Threat Intelligence

T0 review · 3 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read A two-stage retrieval-then-select pipeline with a decomposed reinforcement-learning reward achieves the best average F1 yet reported for multi-label ATT&CK technique extraction on four benchmarks, using an 8B model that runs 28x faster…

desk verdict Solid, checkable empirical work on RLVR for multi-label ATT&CK extraction; the benchmark claim needs a leakage sanity check and the PlugX artifact must be cleaned up. read the letter →

arxiv 2608.06778 v1 pith:OYVD2QXR submitted 2026-08-07 cs.CR cs.CL

Jiayun Zhang , Junshen Xu , Zejun Xie , Yi Fan This is my paper · ORCID

classification cs.CRcs.CL

keywords cyberthreatintelligenceATT&CKtaxonomyattacktechniqueextractionretrieval-augmentedgenerationreinforcementlearningwithverifiablerewardsGRPOmulti-labelclassificationLLMfine-tuning

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that extracting ATT&CK attack techniques from cyber threat intelligence text should be modeled as retrieval-then-select, not as open-ended generation. It proposes TTP-R1, which first shrinks the taxonomy to a short candidate list with a hybrid lexical and semantic retriever and then fine-tunes an LLM to pick the correct subset from that list. The central move is a second training stage with reinforcement learning using verifiable rewards, where precision, recall, and output format are scored as separate channels normalized independently before combination. On four benchmarks, TTP-R1 achieves the highest average F1 at both parent and sub-technique granularity, including a 7.4-point gain in sub-technique F1 over Claude Sonnet 4.5 with the same candidates, while running 28 times faster as an 8B model on a single GPU. If the paper is right, high-quality multi-label TTP mapping can run locally at small-model cost instead of depending on expensive frontier APIs.

What carries the argument

The mechanism that carries the argument is the decomposed verifiable reward with decoupled normalization. Precision and recall are each computed as the average of a parent-technique-level overlap and a sub-technique-level overlap against the ground-truth set, the format channel verifies that the answer block is well-formed, and each channel is z-normalized inside its GRPO group before the weighted sum, so no single reward can dominate the policy gradient. This gives the optimizer separate signals for over-prediction and under-prediction, the two failure modes the paper attributes to prompted frontier models and to SFT-only models respectively. The supporting constraint is the hybrid retriever—BM25 plus ATT&CK-BERT, merged by minimum rank with a top-25 candidate cap—which turns a several-hundred-label space into a selection problem and, the paper notes, sets an upper bound on attainable recall.

What would settle it

An experiment that trains TTP-R1 with an oracle retriever guaranteed to include every true technique, keeping all other details identical, would settle whether the selection stage is actually the bottleneck: if Expert sub-technique F1 stays near the reported 44.2 percent instead of tracking the higher recall bound, the reward design or the SFT initialization, not retrieval, limits performance.

Watch

Extended reading notes

Core claim

The central claim is that direct, set-level supervision can fix the multi-label recall gap that token-level objectives leave behind. After supervised fine-tuning, the model is trained with Group Relative Policy Optimization using a decomposed verifiable reward: one channel scores precision, one scores recall, and one checks output format; precision and recall are each averaged over parent and sub-technique granularity, and every channel is normalized within its group before weighted aggregation. TTP-R1 reaches the highest average F1 on TRAM, Procedures, Derived Procedures, and Expert at both granularities. At the sub-technique level it beats Claude Sonnet 4.5 with retrieval augmentation by 7.4 F1 points, and because it emits short index lists rather than long justifications it answers in about 0.34 seconds per query, 28 times faster. The paper also reports that the RL stage selectively raises recall on multi-label benchmarks while keeping the precision gains of SFT, and that the pipeline transfers to a second 8B base model family.

Load-bearing premise

The retrieval step must place every true technique in the candidate list, because the model can only select from that list; on the Expert benchmark roughly 21 percent of true sub-techniques are absent after retrieval, so no policy trained inside this framework can recover them.

Editorial extensions

If this is right

At an average sub-technique F1 of 63.9% across the four splits, TTP-R1 is the strongest published method on these benchmarks, with its largest edge on the paragraph-level multi-label Expert set.
An 8B model with retrieval-augmented SFT plus RLVR can beat frontier APIs at this task, so accurate TTP mapping is feasible on a single GPU for teams that cannot call expensive hosted models.
Decoupling precision and recall into separately normalized reward channels changes the direction of the RL gradient; the paper reports technique-level recall gains of 18 percent on Derived Procedures and 24 percent on Expert with only a modest precision cost.
Because retrieval plus SFT supplies most of the performance, the candidate-list quality is the primary determinant of end-to-end accuracy; the RL stage is a targeted correction for multi-label completeness rather than a large accuracy jump.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Inference: since retrieval recall caps what any trained policy can output (78.6 percent at the sub-technique level on Expert), the next substantial accuracy gain on that benchmark will come from better retrieval, not from further policy optimization.
Inference: the same retrieve-then-select-plus-RLVR recipe should transfer to other hierarchical low-resource label spaces, such as ATLAS for AI-system threats or vulnerability taxonomies, whenever a retriever can be built with high recall; the paper names ATLAS as future work but does not test it.
Inference: the 28x latency comparison is conservative in a favorable direction, because TTP-R1 was timed with unoptimized local inference while the baselines were timed through API serving; production serving of the small model would likely widen the gap.
Inference: a testable extension is to replace the fixed top-25 hybrid retrieval with iterative or query-expanded retrieval and check whether Expert sub-technique F1 rises toward the 78.6 percent recall ceiling.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Colleague,

Quick read on arXiv:2608.06778. The paper proposes TTP-R1: retrieval-augmented SFT followed by GRPO with decomposed precision/recall rewards, for multi-label ATT&CK technique extraction. The new bit is the combination — prior work did retrieve-then-generate with token-level training, and the closest RLVR baseline (Minerva) only handles single-label extraction. The decomposed reward with decoupled normalization is a sensible adaptation of GDPO. The empirical work is honest: four benchmarks, ablations over two base models, and the RL gains are small but consistent. The method description is clear enough to reimplement, and the tables are internally consistent.

What I like: the paper acknowledges the retriever upper bound and modest RL gains in its own limitations. The comparison against TechniqueRAG is meaningful because the same splits and base model are used. The efficiency claim is directionally plausible, though the latency comparison mixes API serving and local GPU inference, so I'd treat the 28x factor as approximate.

The soft spots are real but not fatal. No code or checkpoints, no error bars or significance tests. No comparison with Minerva, which is cited but never evaluated. The most important concern is split overlap: candidate lists are retrieved from the training corpus, and TRAM and Procedures are short, templated sentences with retrieval recall above 96%. If the splits are random at sentence level, many test items are near-duplicates of training items, and the SFT stage could internalize the retriever's output distribution. The paper gives no overlap statistics and no source-level split to rule this out. That doesn't kill the relative comparison against Claude RAG on these splits, but it undercuts the generalization claim.

There's also an unexplained Aro.dat/PlugX paragraph sitting in Section 1 that has nothing to do with the paper. That is a manuscript hygiene issue a referee should flag.

Bottom line: the paper deserves a serious referee. The method is coherent, the contribution is a useful incremental step, and the benchmark-validity question is exactly what peer review should push on. Recommend conditional acceptance after the authors add overlap statistics and ideally a source-level split robustness check.

Referee Report

3 major / 4 minor

Summary. The paper proposes TTP-R1, a two-stage framework for multi-label MITRE ATT&CK technique extraction from CTI text. Stage 1 retrieves a small candidate set from a training corpus using a hybrid BM25 + ATT&CK-BERT retriever and fine-tunes an 8B LLM (Ministral-8B-Instruct, with an additional Qwen3-8B ablation) to select technique indices from that set. Stage 2 applies GRPO with a decomposed verifiable reward that separately scores precision, recall, and output format, using decoupled normalization and a selective threshold that restricts RL updates to examples that remain difficult after SFT. On four CTI benchmarks (TRAM, Procedures, Derived Procedures, Expert), the paper reports the best average macro-F1 at both technique and sub-technique granularity, a +7.4 pp sub-technique F1 gain over Claude Sonnet 4.5 with the same retrieval-augmented prompt, and a 28x latency reduction when served as an 8B model on a single GPU. The paper also includes ablations across two base models and an honest limitations section acknowledging the retriever upper bound and the modest size of RL gains.

Significance. If the empirical results hold under overlap-controlled evaluation, the contribution is significant: it shows that a retrieval-constrained 8B model trained with RLVR can match or beat frontier-LLM RAG pipelines on multi-label ATT&CK extraction while being far cheaper to serve. The retrieve-then-select formulation, the decomposed precision/recall/format rewards with decoupled normalization, and the selective RL strategy are coherent and clearly described. The paper also gives credit where due: the RAG baselines receive the same candidate list, so the retriever recall cap does not by itself invalidate the relative comparison; the ablation across two base model families is a useful robustness check; and the limitations section correctly identifies the retriever upper bound. However, the headline benchmark claim rests on single-run numbers on splits whose train/test overlap is not analyzed, and the most directly related RLVR baseline (Minerva) is cited but not compared. No code is released, which limits reproducibility. These issues are load-bearing for the central claim and require revision.

major comments (3)

[Section 4.1 / 4.3] The main benchmarking claim may be inflated by train/test near-duplication. The candidate list for every test sample is built from the training corpus C (Section 3.1), and the paper adopts Lekssays et al.'s splits without stating whether they are random at the sentence or paragraph level. On TRAM and Procedures, sub-technique retrieval recall is 97.3% and 96.9%, respectively; if the splits are random, many test sentences will have near-identical training sentences carrying the same gold labels. TTP-R1 is fine-tuned on prompts produced by exactly this retriever, so it can internalize the training-side retrieval distribution, whereas inference-only baselines such as Claude RAG cannot. The retriever recall cap itself does not undermine the relative comparison, because all RAG baselines receive the same candidate list; what needs to be addressed is the overlap. Please report train/test near-duplicate statistics, and where possible report results on overlap-controlled splits (for example, grouping near-duplicates into the same split or deduplicating the corpus) to show that the 7.4 pp sub-technique gain over Claude RAG is not a template-memorization artifact.
[Tables 2 and 3] The results are reported for a single run with no error bars, no multiple seeds, and no significance tests. The average F1 advantage over TechniqueRAG is 3.9 points at technique level and 4.2 points at sub-technique level, and the advantage over TTP-R1 w/o RL is 2.3 and 2.2 points; these margins could easily lie within run-to-run variance, particularly on Expert, which has only 157 test samples and an average of 3.32 labels per sample. Please report variance across at least three seeds and, if feasible, paired significance tests (for instance, bootstrap resampling or McNemar-type tests over per-sample F1) for the key comparisons in Tables 2 and 3.
[Section 2.3 / 4.2] Minerva [2] is discussed in Related Work as applying RLVR to a broad suite of CTI subtasks including technique extraction, but it does not appear in the Compared Methods of Section 4.2. Because the paper's central contribution is an RLVR-based technique extraction method, the absence of any empirical comparison to Minerva leaves the state-of-the-art claim open with respect to the most closely related RLVR approach. Please add Minerva as a baseline on the four benchmarks, or if its published results do not cover these benchmarks, state that explicitly and justify why a direct comparison is not possible.

minor comments (4)

[Section 1] The paragraph beginning 'Aro.dat is, in fact, an encrypted and compressed PlugX payload' appears between the author affiliations and Figure 1, is unrelated to the surrounding introduction, and looks like a leftover from another document. It should be removed or rewritten as a properly introduced motivating example.
[Section 3.2] Equations (3) and (4) do not define precision and recall when the predicted set is empty. Since the format reward only requires non-empty <answer> content and the answer could still be empty or malformed, please specify the convention used for zero-size predicted sets in the reward computation.
[Reproducibility] The paper does not mention a code or model release, and several implementation details are underspecified (for example, the deduplication criterion for retrieved entries, tie-breaking in the rank(.) merge, and the number of RL optimization steps). Releasing the training and evaluation code, or at least the exact retrieval and decoding configurations, would substantially strengthen the reproducibility of the benchmark claims.
[Section 4.6 / Figure 5] The latency comparison in Figure 5 measures TTP-R1 locally with unoptimized HuggingFace inference while measuring the frontier baselines through Bedrock API calls. API latency includes network overhead and serving infrastructure effects, so the 28x speedup is not an apples-to-apples comparison; please report local latency for all models or state explicitly which portions of the measured latency are attributable to serving versus generation.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: TTP-R1's claims are empirical held-out comparisons; self-citations are peripheral and no prediction reduces to a fitted input.

full rationale

The paper's central derivation is not circular. Stage 1 trains on labelled training sentences by minimizing causal language-model loss over the correct candidate indices (Eq. 2), and Stage 2 computes verifiable precision, recall, and format rewards (Eqs. 3-5) against the same training ground-truth labels. All headline results in Tables 2 and 3 are macro-F1 computed on held-out test splits taken from Lekssays et al., not on the training distribution, so the reported gains are an empirical comparison rather than a quantity forced by construction. The retrieval recall cap (Section 4.3: 97.3%, 96.9%, 89.7%, 78.6% on the four test sets) is explicitly acknowledged and applies equally to TTP-R1 and to every RAG baseline, since the same candidate list is provided to all RAG models; therefore the cap does not make the relative F1 comparison circular. The only self-citations are [28] and [32], used respectively for downstream threat-detection applications and few-shot adaptation in the future-work section; neither supports a load-bearing premise of the extraction method. The reward weights (0.45/0.45/0.1) and selective-training threshold (tau=0.99) are manually set hyperparameters, not values fitted to the test set, so they are a tuning concern rather than evidence of circularity. The benchmark-validity concern about possible train/test template overlap is a generalization and correctness risk, not a circularity of the derivation chain.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical or conceptual entities, but it relies on a set of hand-chosen hyperparameters (reward weights, selective threshold, candidate size, sampling temperature) and on benchmark label quality and retrieval coverage assumptions. These are the main items that the reader, not the paper, pays for.

free parameters (5)

Reward weights wP, wR, wF = 0.45, 0.45, 0.1
Chosen by hand for the weighted reward in Eq. (5); directly controls the precision/recall/format tradeoff and is not derived from data.
Selective RL threshold tau = 0.99
Chosen by hand to decide which prompts receive RL gradient updates; affects which examples are optimized in Stage 2.
Candidate list size k and cap m = 25
Number of retrieved technique candidates presented to the LLM; sets the upper bound on achievable recall, with retrieval recall between 78.6% and 97.3% across datasets.
RL sampling temperature = 1.4
Used for group generation in GRPO; affects diversity of completions and the resulting advantage estimates.
LoRA ranks and alphas = SFT r=32 alpha=64; RL r=8 alpha=16
Rank and alpha for LoRA adapters, chosen as standard hyperparameters and not justified by analysis in the paper.

assumptions (4)

domain assumption Ground-truth technique labels in TRAM, Procedures, Derived Procedures, and Expert are correct and complete.
All rewards and metrics are computed against these labels; if labels are noisy or incomplete (the paper notes TRAM label noise), measured F1 may not reflect true extraction quality. See Section 4.1 and Section 4.4.
domain assumption The retrieved candidate set is a faithful restriction of the taxonomy for the evaluation data.
The LLM can only select from the BM25+ATT&CK-BERT candidate list, so any true technique outside it is unreachable; retrieval recall is 78.6% on Expert and 89.7% on Derived Procedures. See Section 3.1 and Section 4.3.
domain assumption The benchmark splits are non-leaky and representative of real-world CTI text.
The retrieval corpus is built from training sentences and evaluation uses published test splits; if test and training CTI reports overlap, candidate recall and F1 would be inflated. See Section 4.1 and Section 4.3.
standard math GRPO with the decomposed reward improves set-level F1 in the intended direction beyond SFT.
The paper relies on prior results that RLVR/GRPO improves reasoning and that decoupled normalization stabilizes multi-reward training; these are treated as background methods. See Section 2.3 and Eqs. (6)-(7).

how reviews work

0 comments

Cite this review

Pith. "Pith review of Retrieval-Constrained Policy Optimization for Attack Technique Extraction from Cyber Threat Intelligence." pith.science (2026). https://pith.science/paper/OYVD2QXR

@misc{pith2026260806778,
  author       = {Pith},
  title        = {Pith review of: Retrieval-Constrained Policy Optimization for Attack Technique Extraction from Cyber Threat Intelligence},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OYVD2QXR}},
  note         = {Machine review of arXiv:2608.06778}
}

read the original abstract

Mapping cyber threat intelligence (CTI) text to MITRE ATT&CK techniques is essential for structured threat analysis, yet manual annotation is costly and does not scale. The ATT&CK taxonomy comprises several hundred attack techniques, and a single CTI passage may describe multiple techniques, making accurate and complete extraction challenging. Existing automated approaches fall short in different ways: multi-label classifiers struggle with severe class imbalance and the large label space, while LLM-based methods--retrieval pipelines and fine-tuned generators--optimize token-level objectives that treat technique annotation as sequence generation rather than set prediction, lacking direct supervision on whether the predicted technique set is correct and complete. We propose TTP-R1, a two-stage framework that combines retrieval-augmented supervised fine-tuning (SFT) with reinforcement learning using verifiable rewards (RLVR). A hybrid retriever first narrows the large label space to a candidate set, and a fine-tuned LLM learns to select the correct techniques. We then apply Group Relative Policy Optimization with a decomposed reward that directly supervises the precision, recall, and output format of the predicted technique set. Across four CTI benchmarks, TTP-R1 achieves the best average F1, improving sub-technique-level F1 by 7.4 percentage points over Claude Sonnet 4.5 with retrieval augmentation, while running 28x faster when served as an 8B-parameter model on a single GPU.

Figures

Figures reproduced from arXiv: 2608.06778 by the authors.

**Figure 2.** Overview of the TTP-R1 framework. Stage 1 uses a hybrid BM25 + embedding retriever to narrow the 500+ ATT&CK taxonomy to a candidate set, then fine-tunes an LLM to select correct techniques from the candidates. Stage 2 applies GRPO with decomposed verifiable rewards that separately supervise precision, recall, and output format. Applied selectively on examples that remain difficult after SFT. extended this paradigm … view at source ↗

**Figure 4.** RL prompt template used in Stage 2 training. Com [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗

Figures from the paper (1 more)

**Figure 6.** Figure 6: Ablation study across two base models (Ministral-8B [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 21 canonical work pages

[2]

Md Tanvirul Alam, Dipkamal Bhusal, Shaswata Ahmad, Nidhi Rastogi, and Philip Worth. 2026. Minerva: Reinforcement Learning with Verifiable Rewards for Cyber Threat Intelligence LLMs.arXiv preprint arXiv:2602.00513(2026)

work page Pith review arXiv 2026
[1]

Basel Abdeen, Ehab Al-Shaer, Anoop Singhal, Latifur Khan, and Kevin Hamlen

work page
[3]

Md Tanvirul Alam, Dipkamal Bhusal, Youngja Park, and Nidhi Rastogi. 2023. Looking beyond iocs: Automatically extracting attack patterns from external cti. InProceedings of the 26th international symposium on research in attacks, intrusions and defenses. 92–108

work page 2023
[4]

Anthropic. 2025. Introducing Claude Sonnet 4.5. https://www.anthropic.com/ news/claude-sonnet-4-5. Accessed: 2026

work page 2025
[5]

2025.{SoK}: Automated{TTP} Extraction from{CTI} Reports–Are We There Yet?

Marvin Büchel, Tommaso Paladini, Stefano Longari, Michele Carminati, Stefano Zanero, Hodaya Binyamini, Gal Engelberg, Dan Klein, Giancarlo Guizzardi, Marco Caselli, et al. 2025.{SoK}: Automated{TTP} Extraction from{CTI} Reports–Are We There Yet?. In34th USENIX security symposium (USENIX Security 25). 4621–4641

work page 2025
[6]

Center for Threat-Informed Defense. 2023. Threat Report ATT&CK Mapper (TRAM). https://github.com/center-for-threat-informed-defense/tram/. Ac- cessed: 2025

work page 2023
[7]

Andrew Crossman, Jonah Dodd, Viralam Ramamurthy Chaithanya Kumar, Riyaz Mohammed, Andrew R Plummer, Chandra Sekharudu, Deepak Warrier, and Mohammad Yekrangian. 2026. Constructing Multi-label Hierarchical Classifica- tion Models for MITRE ATT&CK Text Tagging.arXiv preprint arXiv:2601.14556 (2026)

work page arXiv 2026
[8]

Mohamed Amine Ferrag, Fatima Alwahedi, Ammar Battah, Bilel Cherif, Ab- dechakour Mechri, and Norbert Tihanyi. 2024. Generative ai and large language models for cyber security: All insights you need.A vailable at SSRN 4853709 (2024)

work page 2024

Show all 36 references

[9]

Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, et al. 2025. Deepseek-r1: Incen- tivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948(2025)

2025 arXiv
[10]

Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-Rank Adaptation of Large Language Models. InInternational Conference on Learning Representations. https://openreview.net/forum?id=nZeVKeeFYf9

2022
[11]

Ghaith Husari, Ehab Al-Shaer, Mohiuddin Ahmed, Bill Chu, and Xi Niu. 2017. Ttpdrill: Automatic and accurate extraction of threat actions from unstructured text of cti sources. InProceedings of the 33rd annual computer security applications conference. 103–115

2017
[12]

Amazon Artificial General Intelligence. 2025. Amazon Nova 2: Mul- timodal reasoning and generation models.Amazon Technical Reports (2025). https://www.amazon.science/publications/amazon-nova-2-multimodal- reasoning-and-generation-models

2025
[13]

Udesh Kumarasinghe, Ahmed Lekssays, Husrev Taha Sencar, Sabri Boughorbel, Charitha Elvitigala, and Preslav Nakov. 2024. Semantic Ranking for Automated Adversarial Technique Annotation in Security Text. InProceedings of the 19th ACM Asia Conference on Computer and Communication...

2024
[14]

Ahmed Lekssays, Husrev Taha Sencar, and Tong Yu. 2025. From Text to Action- able Intelligence: Automating STIX Entity and Relationship Extraction.arXiv preprint arXiv:2507.16576(2025)

2025 arXiv
[15]

Ahmed Lekssays, Utsav Shukla, Husrev Taha Sencar, and Md Rizwan Parvez
[16]

Zhenyuan Li, Jun Zeng, Yan Chen, and Zhenkai Liang. 2022. AttacKG: Con- structing technique knowledge graph from cyber threat intelligence reports. In European symposium on research in computer security. Springer, 589–609

2022
[17]

Shih-Yang Liu, Xin Dong, Ximing Lu, Shizhe Diao, Peter Belcak, Mingjie Liu, Min-Hung Chen, Hongxu Yin, Yu-Chiang Frank Wang, Kwang-Ting Cheng, et al
[18]

Mistral AI. 2024. Ministral 8B. https://mistral.ai/news/ministral/. Accessed: 2025

2024
[19]

MITRE Corporation. 2021. MITRE ATLAS: Adversarial Threat Landscape for Artificial-Intelligence Systems. https://atlas.mitre.org/ Accessed: 2026-05-12

2021
[20]

Tu Nguyen, Nedim Šrndić, and Alexander Neth. 2024. Noise contrastive estimation-based matching framework for low-resource security attack pat- tern recognition. InFindings of the Association for Computational Linguistics: EACL 2024. 355–373

2024
[21]

Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback.Advances in neural information processing systems35 (...

2022
[22]

Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. InProceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCN...

2019
[23]

2009.The probabilistic relevance frame- work: BM25 and beyond

Stephen Robertson and Hugo Zaragoza. 2009.The probabilistic relevance frame- work: BM25 and beyond. Vol. 4. Now Publishers Inc

2009
[24]

Asaf Shabtai et al . 2024. LLMCloudHunter: Harnessing LLMs for Auto- mated Extraction of Detection Rules from Cloud-Based CTI.arXiv preprint arXiv:2407.05194(2024)

2024 arXiv
[25]

Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300(2024)

2024 arXiv
[26]

Blake E Strom, Andy Applebaum, Doug P Miller, Kathryn C Nickels, Adam G Pennington, and Cody B Thomas. 2018. Mitre att&ck: Design and philosophy. (2018)

2018
[27]

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

2022
[28]

Junshen Xu, Jiayun Zhang, and Yi Fan. 2026. Multi-Domain Marker Aggregation for Threat Detection in Cloud Environments. InProceedings of the ACM Web Conference 2026. 7809–7820

2026
[29]

Ming Xu, Hongtai Wang, Jiahao Liu, Yun Lin, Chenyang Xu Yingshi Liu, Hoon Wei Lim, and Jin Song Dong. 2024. Intelex: A llm-driven attack-level threat intelligence extraction framework.arXiv e-prints(2024), arXiv–2412

2024
[30]

An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jia...

2025 arXiv
[31]

Yizhe You, Jun Jiang, Zhengwei Jiang, Peian Yang, Baoxu Liu, Huamin Feng, Xuren Wang, and Ning Li. 2022. TIM: threat context-enhanced TTP intelligence mining on unstructured threat data.Cybersecurity5, 1 (2022), 3

2022
[32]

Jiayun Zhang, Junshen Xu, Bugra Can, and Yi Fan. 2025. React: Residual-adaptive contextual tuning for fast model adaptation in threat detection. InProceedings of the ACM on Web Conference 2025. 2488–2499

2025
[33]

Shengyao Zhuang, Xueguang Ma, Bevan Koopman, Jimmy Lin, and Guido Zuc- con. 2025. Rank-r1: Enhancing reasoning in llm-based document rerankers via reinforcement learning.arXiv preprint arXiv:2503.06034(2025)

2025 arXiv
[2023]

InIFIP annual conference on data and applications security and privacy

Smet: Semantic mapping of cve to att&ck and its application to cyberse- curity. InIFIP annual conference on data and applications security and privacy. Springer, 243–260
[2025]

InFindings of the Association for Computational Linguistics: ACL 2025

Techniquerag: Retrieval augmented generation for adversarial technique annotation in cyber threat intelligence text. InFindings of the Association for Computational Linguistics: ACL 2025. 20913–20926

2025
[2026]

Gdpo: Group reward-decoupled normalization policy optimization for multi-reward rl optimization.arXiv preprint arXiv:2601.05242(2026)

2026 arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [2]

Md Tanvirul Alam, Dipkamal Bhusal, Shaswata Ahmad, Nidhi Rastogi, and Philip Worth. 2026. Minerva: Reinforcement Learning with Verifiable Rewards for Cyber Threat Intelligence LLMs.arXiv preprint arXiv:2602.00513(2026)

work page Pith review arXiv 2026

[2] [1]

Basel Abdeen, Ehab Al-Shaer, Anoop Singhal, Latifur Khan, and Kevin Hamlen

work page

[3] [3]

Md Tanvirul Alam, Dipkamal Bhusal, Youngja Park, and Nidhi Rastogi. 2023. Looking beyond iocs: Automatically extracting attack patterns from external cti. InProceedings of the 26th international symposium on research in attacks, intrusions and defenses. 92–108

work page 2023

[4] [4]

Anthropic. 2025. Introducing Claude Sonnet 4.5. https://www.anthropic.com/ news/claude-sonnet-4-5. Accessed: 2026

work page 2025

[5] [5]

2025.{SoK}: Automated{TTP} Extraction from{CTI} Reports–Are We There Yet?

Marvin Büchel, Tommaso Paladini, Stefano Longari, Michele Carminati, Stefano Zanero, Hodaya Binyamini, Gal Engelberg, Dan Klein, Giancarlo Guizzardi, Marco Caselli, et al. 2025.{SoK}: Automated{TTP} Extraction from{CTI} Reports–Are We There Yet?. In34th USENIX security symposium (USENIX Security 25). 4621–4641

work page 2025

[6] [6]

Center for Threat-Informed Defense. 2023. Threat Report ATT&CK Mapper (TRAM). https://github.com/center-for-threat-informed-defense/tram/. Ac- cessed: 2025

work page 2023

[7] [7]

Andrew Crossman, Jonah Dodd, Viralam Ramamurthy Chaithanya Kumar, Riyaz Mohammed, Andrew R Plummer, Chandra Sekharudu, Deepak Warrier, and Mohammad Yekrangian. 2026. Constructing Multi-label Hierarchical Classifica- tion Models for MITRE ATT&CK Text Tagging.arXiv preprint arXiv:2601.14556 (2026)

work page arXiv 2026

[8] [8]

Mohamed Amine Ferrag, Fatima Alwahedi, Ammar Battah, Bilel Cherif, Ab- dechakour Mechri, and Norbert Tihanyi. 2024. Generative ai and large language models for cyber security: All insights you need.A vailable at SSRN 4853709 (2024)

work page 2024