Pith. sign in

REVIEW 4 major objections 4 minor 20 references

A unified LLM pipeline that jointly predicts cancer survival status, survival months, and step-by-step clinical rationales can outperform stronger biomedical baselines, with GRPO reinforcement alignment yielding the best accuracy and the mo

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 ↗

On MSK-CHORD, CoT/GRPO fine-tuned LLaMa3-8B reaches F1 0.84 and MAE 22.0 months while producing teacher-like rationales.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection A serious but sloppy empirical paper: the new protocol and released artifacts are real, but the headline CoT/GRPO claims don't survive contact with the paper's own tables and a likely data-split leak. the 4 major comments →

arxiv 2510.17532 v2 pith:OKVLA4FO submitted 2025-10-20 cs.CL cs.LG

OncoReason: Structuring Clinical Reasoning in LLMs for Robust and Interpretable Survival Prediction

Raghu Vamshi Hemadri , Geetha Krishna Guruju , Kristi Topollai , Anna Ewa Choromanska This is my paper
classification cs.CL cs.LG
keywords cancer survival predictionchain-of-thought reasoningGRPOreinforcement learning alignmentlarge language modelsclinical interpretabilitymulti-task learningoncology
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.

The reading

The paper tries to show that an 8-billion-parameter language model, fine-tuned on structured oncology records, can do three things at once: classify whether a patient is living or deceased after treatment, estimate survival in months, and produce a step-by-step clinical rationale. The authors compare three training strategies and find that adding chain-of-thought distillation improves F1 by 6 points and cuts mean absolute error by 12 percent, and that a reinforcement-learning stage called GRPO improves further, reaching the best F1, the lowest error, and the highest overlap with reference rationales. A sympathetic reader would care because it suggests interpretability and accuracy are not in tension: the model that explains itself best is also the one that predicts best. If the results hold, this points toward auditable LLM decision support in oncology instead of black-box predictors.

Core claim

The paper's central claim is that supervising a language model to imitate teacher-generated reasoning traces, then optimizing it with GRPO under a reward that combines exact outcome correctness, integer validity, and output format compliance, produces a model that simultaneously improves survival classification (F1 0.84), survival-month regression (MAE 22.0 months), and rationale quality (BLEU 98.15, BERT-F1 0.443) on a 24,950-patient oncology dataset. The authors treat the reasoning trace not as decoration but as part of the training signal: chain-of-thought prompting alone lifts F1 by 6 points and reduces MAE by 12 percent, and GRPO adds further gains. They also report that common biomedic

What carries the argument

The central object is the GRPO-aligned autoregressive multi-task protocol. The model is fine-tuned to output, token by token, three things for one patient prompt: binary survival status, continuous survival months, and a chain-of-thought rationale. The GRPO stage replaces a learned value network with group-relative rewards, scoring each generated response on four axes: exact correctness of both predictions, integer validity of the month estimate, strict XML-like formatting, and soft formatting. The rationale traces themselves are distilled from a larger teacher model and used both as supervised fine-tuning supervision and as the policy's cold-start initialization via cluster-selected exempla

Load-bearing premise

The headline comparisons rest on the premise that the GRPO stage was trained only on the 80 percent training split while the 20 percent evaluation split was never seen; the paper's appendix lists the full dataset as GRPO training data without explicitly stating that the held-out split was excluded.

What would settle it

Inspect the released code and confirm that the dataset argument for GRPO excludes the held-out 20 percent; then rerun the GRPO experiments on the clean 80 percent split and check whether F1, MAE, and BLEU results reproduce. A second check: have oncologists rate a random sample of generated rationales for clinical soundness; if most are judged invalid, the interpretability claim collapses even if the metrics stay high.

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

If this is right

  • If these results hold, a roughly 8-billion-parameter model is sufficient for joint survival prediction and rationale generation, allowing clinicians to audit why a prediction was made in a way not possible with Cox models or gradient-boosted trees.
  • Chain-of-thought prompting appears to be a cheap, reliable performance boost: adding teacher-distilled reasoning traces to supervised fine-tuning raises F1 by 6 points and lowers mean absolute error by 12 percent.
  • The 100 percent response rate on rationale generation, where strong biomedical baselines fail thousands of times, suggests that explicit format rewards are what make reasoning traces reliably appear.
  • Because interpretability metrics (BLEU, ROUGE, BERTScore) correlate with prediction accuracy, the authors argue reasoning quality and generalization are linked rather than competing.
  • The same three-part protocol—classification, regression, and rationale generation—could transfer to other clinical endpoints, since the reward function only requires outcome correctness and output format.

Where Pith is reading between the lines

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

  • If the evaluation-split ambiguity is resolved and the results hold, this recipe (cold-start SFT on teacher rationales plus GRPO) may let smaller open-weight models approach or exceed larger closed clinical models on structured prediction tasks without a clinician in the loop during training.
  • The near-ceiling BLEU scores against the teacher's rationales suggest the student may be reproducing teacher phrasing rather than generating independent reasoning; a clinician study or a factual-consistency check of individual reasoning steps would be a stronger test of interpretability than lexical or semantic similarity.
  • The all-or-nothing correctness reward could be replaced with a graded reward based on survival-month distance, which would likely improve calibration at the long tail where the paper's own plots show baselines underestimating high-survival cases.
Share X Bluesky LinkedIn Reddit HN

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 / 4 minor

Summary. The paper proposes OncoReason, a multi-task framework that fine-tunes 8B LLMs (LLaMa3, Med42) on the MSK-CHORD dataset to jointly predict binary survival status, continuous survival months, and generate chain-of-thought (CoT) rationales. Three training strategies are compared: plain SFT, SFT with CoT, and GRPO after a cold-start SFT with teacher-generated CoT traces. The manuscript claims that CoT prompting improves F1 by +6.0 and reduces MAE by 12%, and that GRPO achieves state-of-the-art predictive performance and interpretability across BLEU, ROUGE, and BERTScore. The paper also releases code and a CoT dataset.

Significance. The problem — interpretable, structured outcome prediction in oncology — is timely and clinically relevant. The broad design (autoregressive LLM fine-tuning with CoT distillation and GRPO) is reasonable, and the public release of code and data is a positive step. If the reported results were valid, the work would be a useful benchmark for reasoning-aligned clinical LLMs. However, the manuscript as written does not establish these claims: the headline MAE improvement is contradicted by the paper's own Table 4; Table 3 contains F1 values that are arithmetically incompatible with the reported precision and recall; the GRPO description and appendix leave open the possibility that the held-out evaluation split was used for RL training; and the interpretability metrics measure proximity to the teacher traces that were themselves used as training targets. These are load-bearing issues for the central comparative claim.

major comments (4)
  1. [Abstract / §Experiments, Table 4] The abstract claims 'CoT prompting improves F1 by +6.0 and reduces MAE by 12%.' Table 4 shows the opposite for MAE: OURS(MED42) worsens from 22.91 to 23.69 with CoT, and OURS(LLAMA3) changes from 24.05 to 24.01. The only MAE reductions are in the GRPO rows (22.30 and 22.00), which are not a CoT-prompting effect; relative to the no-CoT SFT baselines these are 2.7% and 8.5% reductions, not 12%. No entry in Table 4 supports the stated 12% CoT-driven improvement.
  2. [Table 3] Several F1 values are incompatible with the reported precision and recall under the standard F1 = 2PR/(P+R). For example, MEDITRON with CoT (P=0.57, R=0.52) gives F1 ≈ 0.54, not 0.42; OPENBIO with CoT (P=0.75, R=0.61) gives F1 ≈ 0.67, not 0.52; MED42 with CoT (P=0.71, R=0.66) gives F1 ≈ 0.68, not 0.63. Because F1 is a headline metric, the table must be corrected and the comparisons re-evaluated.
  3. [§Experiments, GRPO Training; Appendix §4] The manuscript describes an 80:20 train/evaluation split for the dataset and restricts the cold-start SFT to 'approximately 80%' of MSK-CHORD. However, the GRPO section says the policy is fine-tuned 'on the full dataset' after cold-start selection, and Appendix §4 explicitly lists 'Datasets: Cold-start subset (5K) and full MSK-CHORD' for GRPO. No sentence states that the 20% evaluation split was excluded from the RL phase. If GRPO saw the test patients or their teacher-generated rationales, every GRPO result in Tables 3–5 is in-sample and the baselines comparisons are invalid. This must be clarified by an explicit statement; if the split was not honored, the experiments must be rerun.
  4. [§Chain of Thought for Interpretability, Table 5] The interpretability evaluation is circular with respect to the training objective. The CoT traces from DeepSeek-R1 are used as the training targets in the SFT+CoT and GRPO stages, and Table 5 scores the models' outputs against the same traces using BLEU, ROUGE, and BERTScore. BLEU values near 98 indicate near-verbatim reproduction of the training references; they do not measure clinical validity, factual correctness, or reasoning quality. The paper needs a non-circular evaluation — e.g., clinician ratings, factuality checks against the patient record, or robustness to input perturbations — before claims of 'interpretable' and 'coherent rationales' can be accepted.
minor comments (4)
  1. [Appendix §2, Tables 1/6] The number of patient records is inconsistent: the main text Table 1 and Dataset Description say 24,950, while Appendix Table 6 says 25,000. The text also says CoT explanations were generated for '80% samples' in the appendix and 'for all samples' in the main text; please reconcile.
  2. [Appendix §4, GRPO configuration] The GRPO configuration lists both a 'Discount Factor γ=0.99' and 'GAE λ=0.95', which are PPO-style components. GRPO as originally formulated does not use a critic or GAE; please clarify whether the implementation deviates from the cited GRPO method.
  3. [§Related Work] Several citations appear to be placeholder-like or inaccurate (e.g., 'Deep Patient' is attributed to Zhang et al. 2016, but the reference is to a materials-science paper; 'Doctor AI' is cited to Kazachkov 2015, which is about oscillation theory). The reference list needs a careful audit.
  4. [§Experiments, Training] The sentence 'we generate CoT trajectories on 20K held-out samples' is ambiguous: 'held-out' could mean held out from the 5K cold-start subset, or held out from the evaluation split. Please use unambiguous terminology (e.g., 'training-only', 'cold-start-excluded') throughout.

Circularity Check

1 steps flagged

The interpretability headline is teacher-fidelity measured against the same DeepSeek-R1 traces used as training targets; survival metrics are external but the GRPO data-split ambiguity is a separate risk.

specific steps
  1. fitted input called prediction [Method ('Supervised Fine-Tuning with Chain-of-Thought Distillation', Eq. LCoT; 'Cold-Start SFT with GRPO') and Experiments ('Chain of Thought for Interpretability', Table 5)]
    "In this setting, a teacher model generates intermediate reasoning steps, denoted as z⋆, for each input x drawn from the MSK-CHORD dataset... we incorporate an auxiliary loss defined as: LCoT(z, z⋆) = D(z, z⋆)... To assess the interpretability and reasoning capability of our models, we evaluate CoT generations using BLEU, ROUGE, and BERTScore F1... between the model-generated rationales and ground truth annotations."

    The 'ground truth annotations' in the interpretability evaluation are the DeepSeek-R1 CoT traces, which are also the supervised targets z⋆ in LCoT and the trajectories used for GRPO cold-start/alignment. BLEU=98.15, ROUGE-L=0.370, BERT-F1=0.443 therefore quantify how closely the model reproduces its own training target, not independent interpretability or clinical validity. The claim of 'state-of-the-art interpretability' thus reduces to teacher-copy fidelity by construction.

full rationale

The predictive half (F1, MAE/RMSE) is anchored to external MSK-CHORD survival labels and is not circular. The circularity is concentrated in the interpretability evaluation: the model is trained to output DeepSeek-R1 traces (via LCoT and GRPO alignment) and then scored with BLEU/ROUGE/BERTScore against those same traces, so the interpretability comparison to baselines is partly forced. Separately, the GRPO configuration lists 'Datasets: Cold-start subset (5K) and full MSK-CHORD' and the method says the policy is fine-tuned 'on the full dataset'; if that includes the held-out evaluation split, the F1/MAE comparisons would be invalidated, but that is a data-leakage/correctness risk rather than a definitional circularity. There are no load-bearing self-citations or imported uniqueness theorems; the external references (DeepSeek-R1, GRPO) are substantive. Overall: one central claim is partially circular, so score 6.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The empirical claims rest on the quality of MSK-CHORD labels, on treating DeepSeek-R1 traces as expert reasoning, and on treating similarity-to-teacher as interpretability. These are domain assumptions, not derived results. Several hyperparameters (beta, lambda_CoT, K, feature-selection heuristics) are chosen by hand with no sensitivity analysis.

free parameters (4)
  • lambda_CoT
    Weights the CoT divergence term in L_SFT+CoT; no value or sensitivity analysis is reported.
  • beta (GRPO KL scaling)
    Scales reward against KL regularization in Eq. (1); value not reported.
  • K (cold-start cluster count) = 5000
    Number of K-Means clusters used to select GRPO cold-start exemplars; chosen by hand with no ablation.
  • Cancer-specific attribute selection heuristics
    Hand-built priority scores determine which clinical attributes enter each patient summary; no ablation or validation.
axioms (4)
  • domain assumption MSK-CHORD survival labels (status and months) are reliable ground truth.
    The predictive claims assume the public dataset's outcome labels are correct; no label audit is performed.
  • domain assumption DeepSeek-R1-generated CoT traces are valid expert-level clinical reasoning.
    Traces used as training targets and as evaluation 'ground truth' were produced by an LLM with no clinician validation; the abstract calls them 'expert-derived' without support.
  • domain assumption BLEU/ROUGE/BERTScore against teacher traces measure interpretability.
    Interpretability is assessed only as similarity to the teacher model's text, not by clinical correctness, human rating, or causal fidelity.
  • standard math GRPO/PPO objective and KL form from Shao et al. are applicable here.
    The RL objective is standard, but the sparse exact-match correctness reward may poorly shape continuous survival regression.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of OncoReason: Structuring Clinical Reasoning in LLMs for Robust and Interpretable Survival Prediction." pith.science (2026). https://pith.science/paper/OKVLA4FO

@misc{pith2026251017532,
  author       = {Pith},
  title        = {Pith review of: OncoReason: Structuring Clinical Reasoning in LLMs for Robust and Interpretable Survival Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OKVLA4FO}},
  note         = {Machine review of arXiv:2510.17532}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Predicting cancer treatment outcomes requires models that are both accurate and interpretable, particularly in the presence of heterogeneous clinical data. While large language models (LLMs) have shown strong performance in biomedical NLP, they often lack structured reasoning capabilities critical for high-stakes decision support. We present a unified, multi-task learning framework that aligns autoregressive LLMs with clinical reasoning for outcome prediction on the MSK-CHORD dataset. Our models are trained to jointly perform binary survival classification, continuous survival time regression, and natural language rationale generation. We evaluate three alignment strategies: (1) standard supervised fine-tuning (SFT), (2) SFT with Chain-of-Thought (CoT) prompting to elicit step-by-step reasoning, and (3) Group Relative Policy Optimization (GRPO), a reinforcement learning method that aligns model outputs to expert-derived reasoning trajectories. Experiments with LLaMa3-8B and Med42-8B backbones demonstrate that CoT prompting improves F1 by +6.0 and reduces MAE by 12%, while GRPO achieves state-of-the-art interpretability and predictive performance across BLEU, ROUGE, and BERTScore. We further show that existing biomedical LLMs often fail to produce valid reasoning traces due to architectural constraints. Our findings underscore the importance of reasoning-aware alignment in multi-task clinical modeling and set a new benchmark for interpretable, trustworthy LLMs in precision oncology.

Figures

Figures reproduced from arXiv: 2510.17532 by Anna Ewa Choromanska, Geetha Krishna Guruju, Kristi Topollai, Raghu Vamshi Hemadri.

Figure 1
Figure 1. Figure 1: From Black-Box Models to Reasoning￾Augmented LLMs for Cancer Outcome Prediction. Ex￾isting LLMs excel at general biomedical tasks but lack struc￾tured clinical reasoning and explainability. Our proposed framework augments LLMs with Chain-of-Thought (CoT) prompting and GRPO-based alignment to perform inter￾pretable, multi-task survival prediction on MSK-CHORD. to incorporate long-range dependencies and sema… view at source ↗
Figure 2
Figure 2. Figure 2: This diagram illustrates the methodology applied [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Absolute error distribution (in months) for all mod [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Predicted vs. actual survival months for O [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Correlation between reasoning quality (BLEU) [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 5
Figure 5. Figure 5: Generation quality under CoT prompting, evalu [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: ROC Curves for survival status prediction across [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗

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

20 extracted references · 5 linked inside Pith

  1. [1]

    Input Prompt 1### Instruction: 2You are a cancer clinical outcome prediction model

    Input and Model Outputs Here, we present an input prompt example and the corre- sponding responses generated by different models, including reasoning, prognosis comment, and final survival prediction. Input Prompt 1### Instruction: 2You are a cancer clinical outcome prediction model. Based on the patient data provided, predict the treatment outcome. Focus...

  2. [2]

    al 2024; Fong et al

    Dataset Description We utilize theMSK-CHORD dataset(Jee and et. al 2024; Fong et al. 2024b), a comprehensive clinical oncology dataset curated by Memorial Sloan Kettering Cancer Center (MSKCC). The dataset is publicly accessible via the cBio- Portal and includes structured clinical records of patients diagnosed with various cancer types, including lung ca...

    2024

  3. [3]

    These gaps limit precision in outcome attribution

    Performance status during treatments, and 4) Biomarker status ( e.g., PD-L1, EGFR) not provided. These gaps limit precision in outcome attribution. 12</comment> 13 14<prediction> 15Overall Survival Status: 1:DECEASED 16Estimated Overall Survival (months): 27.9 17</prediction> Our Med42 Base (GRPO) Output 1<reasoning> 2Step 1: Assessed patient age (60) and...

  4. [4]

    Training used bfloat16 precision with FlashAttention 2 enabled (Dao 2024)

    Experiment Hardware and Optimization Details All experiments were conducted using 4× NVIDIA A100 GPUs (40GB, PCIe) with PyTorch 2.1 and Hugging Face Transformers (v4.39). Training used bfloat16 precision with FlashAttention 2 enabled (Dao 2024). Input sequences were truncated or packed to a maximum length of 1024 tokens. Gradient checkpointing was used to...

    2024

  5. [5]

    Code and CoT Dataset The code used for our experiments is available at: https://github.com/OncoReason/Clinical-Reasoning-LLMs The Chain-of-Thought dataset we cu- rated is available on Hugging Face: https://huggingface.co/datasets/oncollm/cancer-reasoning- traces

  6. [6]

    Large language models in cancer: potentials, risks, and safeguards.BJR—Artificial Intelligence, 2(1): ubae019

  7. [18]

    Method Generalized Reward Policy Optimization (GRPO) Generalized Reward Policy Optimization (GRPO) (Shao et al. 2024) is a reinforcement learning framework that op- timizes policy models by directly incorporating reward sig- nals into the policy update process without requiring a sep- arate critic network. Instead of estimating a value function for each a...

    2024

  8. [31]

    The patient has advanced Non-Small Cell Lung Cancer (NSCLC) with multiple metastatic sites (Adrenal Glands, Bone, CNS/Brain, Intra Abdominal, Lung, Lymph Node), indicating an aggressive disease course

  9. [42]

    The patient’s age (60) is within the average range for NSCLC patients, but the presence of multiple metastases and a high clinical group (3A) suggest a poorer prognosis

  10. [51]

    Step-by-step reasoning using relevant clinical knowledge

  11. [53]

    The patient is a former/current smoker, which increases the risk of treatment-related complications and reduces overall survival

  12. [62]

    A concise comment on the patient’s prognosis

  13. [64]

    The treatment regimen includes multiple lines of chemotherapy ( CISPLATIN, ETOPOSIDE, CARBOPLATIN, PEMETREXED, GEMCITABINE) and immunotherapy (NIVOLUMAB), indicating a comprehensive and aggressive treatment approach

  14. [73]

    Final prediction in the specified format. 8 9Format your response as follows: 10<reasoning> 11[Step-by-step explanation] 12</reasoning> 13 14<comment> 15[Prognosis summary] 16</comment> 17 18<prediction> 19Overall Survival Status: ’0:LIVING’ or ’1:DECEASED’ 20Estimated Overall Survival (months): [ float value] 21</prediction> 22 23### Input: 24Clinical At...

  15. [75]

    8</reasoning> 9 10<comment> 11The patient’s prognosis is guarded due to the advanced stage of NSCLC, multiple metastatic sites, and smoking history

    The absence of radiation therapy may limit local control, but the inclusion of immunotherapy may enhance systemic control. 8</reasoning> 9 10<comment> 11The patient’s prognosis is guarded due to the advanced stage of NSCLC, multiple metastatic sites, and smoking history. 12</comment> 13 14<prediction> 15Overall Survival Status: 1:DECEASED 16Estimated Over...

  16. [78]

    Ankit Pal, M

    Minneapolis, Minnesota, USA: Association for Compu- tational Linguistics. Ankit Pal, M. S. 2024. OpenBioLLMs: Advancing Open- Source Large Language Models for Healthcare and Life Sciences. https://huggingface.co/aaditya/OpenBioLLM- Llama3-70B. Arasteh, H.; and Lotfinia, M. 2023. Large Language Models Streamline Automated Machine Learning for Clinical Stud...

    Pith/arXiv arXiv 2024

  17. [541]

    ISBN 978-1-4612- 4380-9

    New York, NY: Springer New York. ISBN 978-1-4612- 4380-9. Dao, T. 2024. FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning. InThe Twelfth Interna- tional Conference on Learning Representations. DeepSeek-AI; and et. al., D. G. 2025. DeepSeek-R1: Incen- tivizing Reasoning Capability in LLMs via Reinforcement Learning. arXiv:250...

    Pith/arXiv arXiv 2024

  18. [2021]

    Zhu, Y .; Wang, Z.; Gao, J.; Tong, Y .; An, J.; Liao, W.; Harrison, E

    Predicting the therapeutic efficacy of immunotherapy using multi-modal data: A deep learning approach.Nature Communications, 12(1): 6957. Zhu, Y .; Wang, Z.; Gao, J.; Tong, Y .; An, J.; Liao, W.; Harrison, E. M.; Ma, L.; and Pan, C. 2024. Prompt- ing Large Language Models for Zero-Shot Clinical Predic- tion with Structured Longitudinal Electronic Health R...

    Pith/arXiv arXiv 2024

  19. [2023]

    Multimodal fine-tuning of clinical language models for predicting COVID-19 outcomes.Artificial Intelligence in Medicine, 146: 102695. Hou, Y .; Bert, C.; Gomaa, A.; Lahmer, G.; Hoefler, D.; Weissmann, T.; V oigt, R.; Schubert, P.; Schmitter, C.; Depardon, A.; Semrau, S.; Maier, A.; Fietkau, R.; Huang, Y .; and Putz, F. 2024. Fine-Tuning a Local LLaMA-3 La...

    Pith/arXiv arXiv 2024

  20. [2024]

    Shao, Z.; Wang, P.; Zhu, Q.; Xu, R.; Song, J.; Bi, X.; Zhang, H.; Zhang, M.; Li, Y

    DeepSeekMath: Pushing the Limits of Mathemati- cal Reasoning in Open Language Models.arXiv e-prints, arXiv:2402.03300. Shao, Z.; Wang, P.; Zhu, Q.; Xu, R.; Song, J.; Bi, X.; Zhang, H.; Zhang, M.; Li, Y . K.; Wu, Y .; and Guo, D. 2024. DeepSeekMath: Pushing the Limits of Mathematical Rea- soning in Open Language Models. arXiv:2402.03300. Shoham, R. 2023. C...

    Pith/arXiv arXiv 2024

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.