cs.LGMar 17, 2026

PRISM: Demystifying Retention and Interaction in Mid-Training

Authors: Bharat Runwal, Ashish Agrawal, Anurag Roy, Rameswar Panda

Organizations: IBM Research, MIT-IBM Watson AI Lab

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

We present PRISM, a comprehensive empirical study of mid-training design choices for large language models. Through controlled experiments across seven base models spanning four families (Granite, LLaMA, Mistral, Nemotron-H), two architecture types (dense Transformer and attention-Mamba hybrid), and scales from 3B to 24B parameters, we show that mid-training on approximately 27B high-quality tokens yields consistent gains of +15 to +40 points on math, +5 to +12 points on code, and +6 to +13 points on science benchmarks while preserving general performance. The full PRISM to RL pipeline improves macro-average across six reasoning benchmarks from under 12 to 29-42 (a 3-4x improvement), whereas RL applied directly to most of the base models remains substantially less effective, with AIME scores near zero. Data composition matters most at mid-training, not RL: including science data during mid-training unlocks +17 to +28 point GPQA-Diamond gains during RL, while changing the RL mix produces less than 2 point differences. Mechanistically, mid-training densely restructures over 90% of model weights, while RL makes sparse, front-loaded refinements to approximately 5% of parameters. Representation analysis (CKA) confirms that RL consistently preserves mid-training's representational geometry (over 0.998 CKA) across architectures. Crucially, RL applies identical weight changes regardless of starting point, yet only succeeds on mid-trained models, consistent with mid-training placing the model in a configuration from which RL can effectively improve performance. Our results demonstrate that retention-aware mid-training is highly effective for reliable reasoning enhancement and provide practical guidance for designing robust mid-training pipelines.

Figures & tables

Figure 1: PRISM overview. Mid-training decisions are decomposed into their principal design axes, including retention of general and long-context abilities, domain interaction (math, code, science), benchmark selection, reinforcement learning compatibility, and scaling behavior. PRISM enables holistic evaluation of mid-training choices across model families at scale.
DatasetTypeTokens (B)
DCLM-EDU Allal et al. (2025)General web data111.46
Open-R1 (MoT) Lozhkov et al. (2025)Math reasoning0.60
Nemotron Post-Training v1 Nathawani et al. (2025)Math35.93
Megamath-Web-Pro Zhou et al. (2025)Math web14.73
Open-R1 (MoT) Penedo et al. (2025)Code reasoning1.18
OpenCodeReasoning-2 Ahmad et al. (2025)Code reasoning1.12
Table 1: Datasets used in mid-training mixtures. Token counts are reported in billions (Granite 3.3, 8B).
Figure 2: Mid-training data mixture configurations and per-source sampling percentages. The outer ring shows individual data sources; the inner ring groups them by domain category.
CategoryBenchmarksWhat it measuresWhy it matters
General abilityLeaderboard-V1 (LB-V1) (ARC, HellaSwag, MMLU, TruthfulQA, Winogrande, GSM8K), Leaderboard-V2 (LB-V2) (IFEval, BBH, MATH, GPQA, MUSR, MMLU-Pro)Broad multitask knowledge and robustnessDetects generalization regressions hidden by domain-specific gains.
Long-contextRULERLong-context retrievalEnsures mid-training does not degrade long-context retrieval capabilities.
CodeLiveCodeBench Jain et al. (2024) , Codeforces Penedo et al. (2025)Executable program synthesis and reasoningCaptures real-world coding ability.
MathAIME AIME () , MATH500 Lightman et al. (2023)Mathematical reasoningHighly sensitive to data quality and mid-training composition.
ScienceGPQA-Diamond Rein et al. (2023)Expert-level scientific reasoningProbes scientific reasoning capabilities
Table 2: Benchmark categories recommended for evaluating mid-training design choices.
 Leaderbds.CodeSci.Math
StageV1V2LCBCFGPQAAI24AI25M500
Phase 363.3019.447.058.6119.539.3816.0965.88
Phase 462.8420.857.897.9517.8510.0014.0661.70
After LC62.9120.5310.396.1825.9323.5920.9477.44
Table 3: Effect of when mid-training is applied on Granite-4 Micro (3B). Phase 3/4 = intermediate/late pretraining; After LC = after long-context extension.
 Leaderboard V1Leaderboard V2
MixtureARCHellaSwagMMLUTruthfulQAWinograndeGSM8KOpenLLM V1 AvgIFEvalBBHMATHGPQAMUSRMMLU-ProOpenLLM V2 Avg
Base61.9583.4662.5652.2480.3556.3366.1546.6224.6810.206.388.8823.8220.10
Math only62.5478.7264.2946.0475.3071.9566.4746.4625.5717.755.599.0829.8622.39
Math + Code61.0178.0962.6547.3674.7473.4666.2245.5626.8718.435.9310.6028.4022.63
Math + Code + Science61.6978.1262.9846.9674.9074.2266.4846.4426.3220.027.278.6029.5523.03
Table 4: Leaderboard V1 and V2 results for Granite-3.3-8B mid-trained with the mixtures in Fig. 2 .
MixtureCodeMathGPQA
Base2.078.9522.56
Math2.8136.4317.34
Math+Code10.7144.9919.02
Math+Code+Sci10.5848.7529.12
Table 5: Domain-specific results for Granite-3.3 (8B). Code/Math are averages; full results in Appendix Table 19 .
  LeaderboardsCodeScienceMath
ModelVariantLB V1LB V2LCBCFCode AvgGPQA-DAIME24AIME25MATH500Math Avg
Granite-3.3 (8B)Base66.1520.102.151.992.0722.560.460.3126.098.95
 PRISM66.4823.0310.6310.5210.5829.1237.1827.9681.1148.75
Granite-4 Micro (3B)Base66.0121.820.242.281.2621.5516.0912.3450.4226.28
 PRISM62.9120.5310.876.258.5634.3427.1922.2979.4042.96
Granite-4-H Micro (3B)Base64.4918.990.600.880.7420.887.082.7030.1713.32
Table 6: Base versus PRISM (Math+Code+Science) mid-training results across model families. Code Avg is the mean of LiveCodeBench (LCB) and Codeforces (CF). Math Avg is the mean of AIME24, AIME25, and MATH500. All values are reported to two decimal places.
 RULERCode / ScienceMath
Model Variant8k16k32k64k128kLCBCFCode AvgGPQA-DAIME24AIME25MATH500Math Avg
Granite-3.3 Base85.8182.4075.5364.9159.092.151.992.0722.560.460.3126.098.95
Mid-Train (Math+Code)89.0260.4421.5211.716.4611.1110.3010.7119.0232.4428.3374.2244.99
Mid-Train + LC (Attention)90.0482.5671.4754.6336.3223.7815.5319.6517.8536.5632.5567.2045.44
Mid-Train + LC (Full)89.2980.7470.8656.0238.4129.9921.0425.5214.4835.2130.3662.3042.62
Merge (15% Base + 85% Mid-Train)89.1269.7632.6315.4411.3210.7510.9610.8622.2228.3924.9072.9742.09
Table 7: Restoring long-context capability after mid-training for Granite-3.3 (8B). RULER is evaluated from 8k to 128k input lengths. Downstream performance includes Code (LiveCodeBench, Codeforces), Science (GPQA-Diamond), and Math (AIME24, AIME25, MATH500).
Figure 3: Long-context restoration pipeline. After PRISM mid-training degrades RULER@128k from 59.09 to 6.46, a linear merge (15% base + 85% mid-trained) followed by long-context extension recovers performance to 42.16 (full params) or 37.75 (attention-only).
ContextLB-V1LB-V2LCBCFGPQAAIME24AIME25M500
Base66.0121.820.242.2821.5516.0912.3450.42
8k62.9120.5310.876.2534.3427.1922.2979.40
16k64.2320.3712.198.9038.8931.8225.2682.47
32k64.4821.0514.937.5039.8930.9821.8782.70
Table 8: Mid-training context length ablation on Granite-4 Micro (3B) with Math+Code+Science mix. V1/V2 = Leaderboard V1/V2.
Tok. (B)LB-V1LB-V2CodeGPQAMath
Base66.0121.821.2621.5526.28
10.4963.4519.509.5919.1940.21
15.7363.2419.799.0223.0642.07
26.2163.2819.638.6919.1942.22
31.4663.1620.057.6221.3842.42
Table 9: Token budget ablation on Granite-4 Micro (3B), Math+Code mix. Full table in Appendix 20 .
DomainSourcesCount
MathDeepScaleR-Preview294K
 INTELLECT-2-RL 
 Skywork-OR1-RL-Data 
ScienceNemotron-PT-v1-stem100K
CodeDeepCoder-Preview142K
 Skywork-OR1-RL-Data 
Table 10: RL datasets and prompt counts.
Figure 4: \textscPRISM→RL\textsc{PRISM}\to\text{RL} : Granite-3.3-8B. RL training curves on the PRISM -mid-trained checkpoint using the unbalanced MCS mix. All benchmarks show consistent, monotonic improvements.
Figure 5: \textscPRISM→RL\textsc{PRISM}\to\text{RL} : Mistral-Small 24B. The largest model tested shows the strongest GPQA-Diamond gains ( +27.95 ) and non-saturating code improvements.
Figure 6: \textscPRISM→RL\textsc{PRISM}\to\text{RL} : Nemotron-H 8B (Hybrid). RL yields stable gains on the hybrid attention-Mamba architecture, confirming that mid-training benefits extend beyond dense Transformers.
Figure 7: RL on Granite-3.3-8B base (no mid-training). AIME24/25 remain near zero throughout training, and overall gains are substantially smaller than the \textscPRISM→RL\textsc{PRISM}\to\text{RL} pipeline (Fig. 4 ).
Figure 8: RL on Nemotron-H 8B base (no mid-training). Even for hybrid architectures, RL on the base model shows limited progress on harder benchmarks compared to \textscPRISM→RL\textsc{PRISM}\to\text{RL} (Fig. 6 ).
Figure 9: \textscPRISM→RL\textsc{PRISM}\to\text{RL} with balanced mix: Granite-3.3-8B. Domain-equalized RL produces comparable math and code gains to the unbalanced mix (Fig. 4 ), with stable training throughout.
Figure 10: Pass rates on initially unsolved (code, score = 0) and hardest (math, score = 1) prompts during RL training of Granite-3.3 (8B) with the balanced mix.
ModelMTRLLCBCFCode AvgAIME24AIME25MATH500Math AvgGPQAAVG
LLaMA-3.1––0.000.070.040.050.156.512.2420.207.49
MC–6.936.036.4820.6719.5873.7037.9819.5321.33
MCS–6.095.455.7716.4519.3273.4736.4121.0421.07
MCMC12.3111.8512.0825.4723.2378.9942.5623.0625.90
MCMCS11.8312.8012.3224.4323.1278.6242.0624.7526.38
MCSMC13.6211.4112.5120.4721.6777.1039.7534.0128.76
Table 11: Full Base→Mid-training→RL\text{Base}\to\text{Mid-training}\to\text{RL} pipeline results across LLaMA-3.1-8B, Granite-3.3-8B, and Mistral-7B. MC = math + code mix; MCS = math + code + science mix. MT = mid-training mix; RL = RL mix. Highlighted rows show the best configuration per model.
Figure 11: Mid-training densely restructures the network; RL makes sparse, surgical refinements. Top row: layer-wise normalized L2 divergence for Granite-3.3 (8B, left) and Nemotron-H (8B, right). Mid-training (blue) changes weights 370–580 ×\times more than RL (red, orange), broadly across all layers with some layer-wise variation. For Nemotron-H, the repeating pattern reflects its hybrid architecture where Mamba-2, self-attention and FFN are separate sequential layers with independent residual connections ( NVIDIA et al., 2025 ) . Bottom row: update sparsity by component type. Mid-training modifies >> 90% of all parameters (attention, MLP, and Mamba alike), while RL leaves >> 93% unchanged.
ModelMTTransitionAttnMLPMambaTotalSparsity
Granite-3.3 (8B)MCSBase →\to MT0.1750.329–0.1759.3%
MT →\to RL0.00030.0006–0.000395.9%
Base →\to RL (no MT)0.00040.0007–0.000496.0%
MCBase →\to MT0.1770.333–0.1779.3%
MT →\to RL0.00030.0006–0.000395.8%
Nemotron-H (8B, Hybrid)MCSBase →\to MT0.2300.2890.1380.1122.7%
Table 12: Weight divergence summary across models and architectures. Normalized L2 = ∥wnew−wold∥2/∥wold∥2\|w_{\text{new}}-w_{\text{old}}\|_{2}/\|w_{\text{old}}\|_{2} . Nemotron-H reports all three component types (Attention, MLP, Mamba). Sparsity = fraction of parameters with << 1% relative change. The dense/sparse asymmetry is consistent across all component types and architectures.
ModelMixAttnMLPMambaTotal
Granite-3.3 (8B)MC0.1770.333–0.177
MCS0.1750.329–0.175
Nemotron-H (8B)MC0.2320.2920.1400.113
MCS0.2300.2890.1380.112
Table 13: MC vs. MCS weight divergence (Base →\to MT normalized L2). Both models show nearly identical per-component L2 norms across data compositions, confirming that the training intensity is matched between MC and MCS despite their different downstream capabilities.
Figure 12: Data composition redirects weight updates across all sub-components. Cosine similarity between MC and MCS weight update vectors ( ΔW=WMT−Wbase\Delta W=W_{MT}-W_{base} ) for Granite-3.3 (left) and Nemotron-H (right). Overall cosine similarity of 0.52 and 0.62 respectively confirms that different data compositions steer weights in substantially different directions despite nearly identical magnitudes. The embedding/LM-head layers are most aligned (0.82–0.88), while attention, MLP, and Mamba layers all show low directional similarity (0.48–0.64).
Figure 13: Pass rate landscape on held-out MATH500 problems. (a) Math pass rate at linearly interpolated weight checkpoints along the Base →\to MT →\to RL path for Granite-3.3 and LLaMA-3.1, evaluated on 200 held-out MATH500 problems (7680 generation tokens). Pass rate increases monotonically from Base to MT (16.9% →\to 75.5% for G33, 2.6% →\to 43.1% for LLaMA) and continues increasing through RL. (b) 2D pass rate landscape for Granite-3.3 centered at MT, with axes toward RL ( α\alpha ) and toward Base ( β\beta ). The RL direction consistently improves performance while moving toward Base degrades it.
Figure 14: RL preserves representational geometry; mid-training reshapes it in model-specific ways. Layer-wise linear CKA Kornblith et al. (2019) on Wikipedia (top) and GSM8K math prompts (bottom) for Granite-3.3 (left) and Nemotron-H (right), evaluated on 200 prompts per input type with batch-size-1 encoding. MT vs. RL (green) is ≈\approx 1.0 at every layer across both models and both input types, confirming RL preserves mid-training’s representational geometry. Base vs. MT and Base vs. RL (blue, pink) are nearly identical, confirming all representational change comes from mid-training. The magnitude and layer pattern of mid-training’s representational shift is model- and input-specific. See Table 14 for the full summary.
ModelArch.WikiC4GSM8K
Granite-3.3 (8B)Dense0.9999 ±\pm 0.00000.9999 ±\pm 0.00000.9997 ±\pm 0.0000
LLaMA-3.1 (8B)Dense0.9999 ±\pm 0.00000.9999 ±\pm 0.00000.9996 ±\pm 0.0001
Nemotron-H (8B)Hybrid0.9999 ±\pm 0.00000.9998 ±\pm 0.00000.9993 ±\pm 0.0001
Table 14: MT vs. RL representational similarity (minimum linear CKA ±\pm bootstrap std) across input distributions. Values are the minimum layer-wise CKA across 20 bootstrap resamples of 100 from 200 inputs. RL consistently preserves mid-training’s representational geometry ( >> 0.998) across all three models and all three input types, spanning both dense Transformers and hybrid attention-Mamba architectures.
ModelStagePassMed. LenNeg-LPCorr.Incorr.
Granite-3.3 (8B)Base16.9%1200.382–0.383
MT75.5%2,2540.1380.1280.153
RL79.5%1,7000.1410.1350.160
LLaMA-3.1 (8B)Base2.6%1580.758–0.780
MT43.1%1,0520.3770.1460.469
RL64.6%1,1880.2670.1490.320
Table 15: Correctness, response length, and prediction confidence across pipeline stages on 200 held-out MATH500 problems (8 samples/prompt, 7680 max generation tokens, step-by-step reasoning prompt). Pass = mean pass rate across 8 samples per prompt (%). Med. Len = median response length (tokens). Neg-LP = mean negative log-probability. Corr./Incorr. = mean neg-LP for correct/incorrect responses; – indicates too few correct samples. The \textscPRISM→RL\textsc{PRISM}\to\text{RL} pipeline consistently achieves the highest pass rates across all three model families.
Figure 15: Mid-training transforms prediction behavior: models learn to reason longer with calibrated confidence. Evaluated on 200 held-out MATH500 problems. Top row: response length distributions shift from short outputs (Base, gray) to extended reasoning chains (MT, blue), with RL (red) adjusting length. Bottom row: mean negative log-probability at each stage.
Figure 16: RL optimization is front-loaded and starting-point-invariant. Top row: cumulative L2 divergence from the initial checkpoint over RL steps for Granite-3.3 (left) and Nemotron-H (right). Solid lines: MT →\to RL; dashed lines: Base →\to RL. Most weight change occurs in the first ∼\sim 200–400 steps, then plateaus. MT →\to RL and Base →\to RL produce nearly identical divergence profiles, confirming that RL’s weight footprint is independent of the starting point. Bottom row: sparsity evolution showing the fraction of parameters within 1% of their initial values. The active parameter set grows progressively from ∼\sim 1.5% at step 20 to ∼\sim 5–6% by step 960, with all component types following the same trajectory.
Table 32
DatasetTypeTokens (B)
DCLM-EDUGeneral Web Data2.51
Nemotron Post-Training v1Math (QA/Reasoning)5.08
Megamath-Web-ProMath (web)4.33
StarCoder2Code (web)37.52
xenArcAI-codexCode (QA/Reasoning)3.860
Table 18: Datasets used in the long-context extension phase. Token counts in billions (Granite 3.3, 8B).
 CodeScienceMath
Model / MixtureLiveCodeBenchCodeforcesCode AvgGPQA-DiamondAIME 24AIME 25MATH 500Math Avg
Base2.151.992.0722.560.460.3126.098.95
Math only2.153.462.8117.3426.7222.0860.5036.43
Math + Code11.1110.3010.7119.0232.4428.3374.2244.33
Math + Code + Science10.6310.5210.5829.1237.1827.9681.1148.75
Table 19: Domain-specific evaluation results for Granite-3.3-8B (full breakdown). Code Avg is the mean of LiveCodeBench and Codeforces; Math Avg is the mean of AIME 24, AIME 25, and MATH 500.
Token Budget (B)LB V1LB V2LCBCFCode AvgGPQA-DAIME24AIME25MATH500Math Avg
Base66.0121.820.242.281.2621.5516.0912.3450.4226.28
10.4963.4519.5010.518.689.5919.1923.9519.1777.5240.21
15.7363.2419.7910.757.289.0223.0626.1421.3078.7642.07
26.2163.2819.639.807.588.6919.1928.4920.1078.0842.22
31.4663.1620.058.246.997.6221.3828.0222.0877.1542.42
Table 20: Effect of increasing mid-training token budget on Granite-4 Micro (3B) using the Math+Code mixture with fixed 8k context length. LCB denotes LiveCodeBench, CF denotes Codeforces, and GPQA-D denotes GPQA-Diamond. Code Avg averages LCB and CF, while Math Avg averages AIME24, AIME25, and MATH500.
HyperparameterValue
Sampling Parameters 
Responses per Prompt16
Temperature1.0
Top- pp1.0
Training Dynamics 
Total Batch Size1024
Table 21: RL training hyperparameters and configuration.
Figure 17: RL training correctness curves for Granite-4 Micro Dense (3B) after PRISM mid-training. All four domains show stable, monotonically improving training dynamics over 2k steps.
Figure 18: \textscPRISM→RL\textsc{PRISM}\to\text{RL} : Mistral-7B. Consistent improvements across code and math benchmarks using the unbalanced MCS mix.
Figure 19: \textscPRISM→RL\textsc{PRISM}\to\text{RL} : LLaMA-3.1-8B. Stable, monotonic gains across all reasoning benchmarks.
Figure 20: \textscPRISM→RL\textsc{PRISM}\to\text{RL} : Granite-4 Micro Dense (3B). Consistent but smaller absolute gains compared to 8B models.
Figure 21: RL on LLaMA-3.1-8B base (no mid-training). AIME24/25 remain near zero; GPQA-Diamond regresses below the base model.
Figure 22: RL on Mistral-7B base (no mid-training). Limited and unstable gains; AIME24/25 remain near zero throughout.
Figure 23: \textscPRISM→RL\textsc{PRISM}\to\text{RL} with balanced mix: Mistral-Small 24B. GPQA-Diamond gain (+25.93) slightly lower than unbalanced mix.
Figure 24: \textscPRISM→RL\textsc{PRISM}\to\text{RL} with balanced mix: Granite-4 Micro Dense (3B). Code and GPQA-Diamond gains are notable (+4.30 LCB, +6.06 GPQA).
Figure 25: \textscPRISM→RL\textsc{PRISM}\to\text{RL} with balanced mix: Granite-4-H Micro (Hybrid, 3B). Largest balanced-mix gains among small models, with +8.09 on Codeforces and +11.95 on GPQA-Diamond.
Threshold (%)Base →\to MT unchangedMT →\to RL unchanged
0.12.1%82.3%
0.52.4%87.3%
1.03.5%95.0%
2.05.3%97.5%
5.011.1%99.0%
10.020.1%99.5%
Table 22: Per-element sparsity at different relative change thresholds (Granite-3.3, 8B). The dense/sparse asymmetry between mid-training and RL is robust across all tested thresholds. Bold row shows the 1% threshold used in the main text.
 Granite-3.3 (8B)Nemotron-H (8B)
Sub-componentBase →\to MTMT →\to RLBase →\to RLΔ\DeltaBase →\to MTMT →\to RLBase →\to RLΔ\Delta
Attn: Q83.5%4.4%5.2%+0.897.5%5.5%4.9%−- 0.6
Attn: K83.2%4.6%5.2%+0.696.9%5.5%4.8%−- 0.7
Attn: V97.9%5.7%7.5%+1.897.1%8.5%7.6%−- 0.9
Attn: O98.0%5.6%6.7%+1.197.2%7.2%6.4%−- 0.8
MLP: gate98.3%5.4%6.1%+0.7––––
Table 23: Sub-component weight analysis: % of parameters changed ( >> 1% relative change) across pipeline transitions. Δ\Delta = difference between Base →\to RL and MT →\to RL. The near-zero Δ\Delta values confirm that RL targets the same sub-components in the same proportions regardless of starting point. Value and output projections are consistently the most affected by RL across both architectures, while Mamba parameters (A, dt) and norms remain frozen.
Figure 26: RL targets the same sub-components regardless of starting point. Fraction of parameters changed ( >> 1% relative change) during RL for Granite-3.3 (left) and Nemotron-H (right), comparing MT →\to RL (blue) vs. Base →\to RL (pink). The near-identical distributions confirm that RL’s sub-component targeting is intrinsic to the optimization, not a consequence of mid-training.
Figure 27: CKA across input distributions: Granite-3.3 (8B). MT vs. RL (green) remains ≈\approx 1.0 on all three inputs. Base vs. MT divergence is input-dependent: strongest on GSM8K math prompts (min 0.55), weakest on C4 general text (min 0.94).
Figure 28: CKA across input distributions: LLaMA-3.1 (8B). Same pattern as Granite-3.3: MT vs. RL ≈\approx 1.0 everywhere, confirming the finding generalizes across dense Transformer families.
Figure 29: CKA across input distributions: Nemotron-H (8B, Hybrid). MT vs. RL >> 0.998 on all three inputs, consistent with all other models. Base vs. MT divergence is most pronounced on GSM8K math prompts (min ≈\approx 0.41), reflecting mid-training’s targeted restructuring of reasoning-relevant representations.
Figure 30: AIME 2026 accuracy over RL training steps. Both Granite-3.3 (8B) and Mistral-Small (24B) show steady gains on this recently released benchmark, confirming generalization of the \textscPRISM→RL\textsc{PRISM}\to\text{RL} pipeline.
More in this topic

LLM Training

Large language model training studies the data, objectives, optimizers, and systems used to build and adapt language models.

See all 342 papers

Explore similar work

CardsList

  1. Mid-Training with Self-Generated Data Improves Reinforcement Learning in Language Models

    May 8, 2026Aswin RRV, Jacob Dineen, Divij Handa +4Synthetic Data GenerationRL for Language Model Reasoning

    Arizona State University · Google Cloud AI Research · Google DeepMind

  2. ExpRL: Exploratory RL for LLM Mid-Training

    Jun 15, 2026Violet Xiang, Amrith Setlur, Chase Blagden +2Reinforcement LearningRL for Language Model Reasoning

    1Stanford University · 2Carnegie Mellon University · 3OpenAI +1

  3. Understanding Reasoning from Pretraining to Post-Training

    Jul 17, 2026Jingyan Shen, Ang Li, Salman Rahman +4Language Model PretrainingRL for Language Model Reasoning

    New York University · Columbia University · University of California, Los Angeles +2