VibeThinker

Weibo · 2026-06-15 · analysed 2026-08-10 · skill v0.3.5

dense full-attention text-only undisclosed medium reasoning-specialised

Design overview

VibeThinker-3B is a post-training-only project: a fixed 3B dense base (Qwen2.5-Coder-3B, named but never described internally) is driven to frontier-level verifiable-reasoning performance through a five-stage pipeline built on the authors' Spectrum-to-Signal Principle. Stage 1 is a two-stage curriculum SFT: broad multi-domain coverage (math, code, STEM, chat, instruction following) over 5 epochs, then 2 epochs on a hard subset of traces ≥5K tokens filtered by VibeThinker-1.5B rollouts, with Diversity-Exploring Distillation merging domain-specialist checkpoints at the parameter level to preserve a wide solution 'spectrum'. Stage 2 runs MGPO — a GRPO-style clipped policy objective reweighting prompts by their proximity to the maximum-entropy point p=0.5, keeping training at the capability boundary — sequentially over Math, Code, and STEM RL at a single 64K context window (progressive window expansion is rejected as harmful at this scale), followed by Long2Short Math RL, which redistributes reward among correct trajectories toward shorter ones to compress reasoning length. Stage 3 is offline self-distillation: verified trajectories from the Math/Code/STEM checkpoints, selected by a learning-potential score, are distilled back into one student. Stage 4 is Instruct RL with rule-based and rubric rewards to restore strict instruction controllability. At inference, CLR scales test-time compute at the claim level: 32 trajectories, five decision-relevant claims each, self-verified and aggregated by a reliability-weighted vote. The report's thesis — the Parametric Compression-Coverage Hypothesis — is that verifiable reasoning compresses into a small 'reasoning core' while knowledge needs broad parameter coverage, which the GPQA-Diamond gap (70.2) versus frontier peers illustrates.

What the report claims is novel

Atlas assessment: This is a post-training recipe report on a fixed Qwen2.5-Coder-3B base: the block is entirely undisclosed, so there is no architectural novelty relative to atlas entries. The post-training stack is a GRPO-family variant (MGPO adds maximum-entropy prompt weighting to the clipped objective — a modest, well-motivated twist on deepseek-r1's verifiable-RL recipe), plus Long2Short length-reward redistribution (echoing length-bias control seen in the deepseek-r1/post-training literature), offline self-distillation from own checkpoints (similar to deepseek-r1's R1-distill consolidation), and CLR, a claim-level test-time scaling scheme that is a genuine contribution relative to trace-level self-verification (deepseek-r1's 'verify' mode). The headline claims (matching 671B-class models on AIME26) are internally consistent but rest on 2026 competition sets and majority of evaluation is Pass@1 over 64 samples with LLM-as-judge assistance on complex answers; the knowledge gap (GPQA-D 70.2 vs 80+ for frontier peers) is honestly reported and matches the compression-coverage thesis. Nothing here contradicts atlas peers; the report's own framing admits reasoning-knowledge decoupling rather than general frontier parity.

Relation to the atlas

Lineage in

Influence out

No descendants recorded yet.

Notable omissions

Closest relatives in the atlas

Spec sheet

Identity

FamilyVibeThinker[1]
OrganisationWeibo
Release date2026-06-15[2]
Report URLhttps://arxiv.org/abs/2606.16140[3]
Licensenot disclosed
Open weightsnot disclosed

Variants

Scale

Total parameters3B[4]
Active parameters / token3B
Layersnot disclosed
Hidden dimensionnot disclosed
FFN inner dimensionnot disclosed
FFN rationot disclosed
Query headsnot disclosed
KV headsnot disclosed
Head dimensionnot disclosed
Vocabulary sizenot disclosed
Embedding tyingnot disclosed

Core block

Block typedense[5]
Attention variantnot disclosed
Attention layer patternnot disclosed
Depth mixingsequential-residual
Activationnot disclosed
Expert countn/a[6]
Experts per tokenn/a[6]
Shared expertsn/a[6]
Routingn/a[6]
Load balancingn/a[6]
Expert granularityn/a[6]
Methodnot disclosed
RoPE base frequencynot disclosed
Partial RoPEnot disclosed
Typenot disclosed
Placementnot disclosed
QK-normnot disclosed
Attention sinksno
Softcappingno

Context

Trained context length65,536[7]
Deployed context lengthnot disclosed
Extension methodnot disclosed

Tokenizer

Algorithmnot disclosed
Notesnot disclosed

Training

Training tokensnot disclosed
Data compositionSFT: multi-domain mixed dataset covering math, code, STEM reasoning, general chat, and instruction following; query synthesis/expansion from high-confidence seeds (explicit final answers, unit tests), multi-path teacher distillation with majority-voting pseudo-labels; quality control via n-gram filtering (incl. benchmark-contamination removal), LLM query-quality filtering, and trace-correctness filtering (answer verification, code sandbox, LLM majority voting). RL: math (final-answer verification), code (sandbox execution + test cases), STEM (answer matching + option verification), all under strict decontamination; Instruct RL: format-sensitive prompts, long-context instructions, general alignment examples[8]
Curriculum / stagingtwo-stage curriculum SFT (stage 1: broad coverage, 5 epochs, full filtered dataset, sequence packing; stage 2: hard long-horizon subset, 2 epochs, traces ≥5K tokens, easy problems filtered via VibeThinker-1.5B rollouts with error rate <0.75); sequential multi-domain RL: Math RL → Code RL → STEM RL at a single 64K window; Long2Short Math RL (accuracy-first MGPO, then token-efficiency reward redistribution among correct trajectories, λ=0.2); offline self-distillation; Instruct RL[9]
Optimizernot disclosed
LR schedulecosine annealing from 5×10⁻⁵ decaying to 8×10⁻⁸ with 5% linear warmup (both SFT stages)[10]
Batch scheduleglobal batch size 128 (both SFT stages)[10]
Precisionnot disclosed
Parallelismnot disclosed
Hardwarenot disclosed
Disclosed computenot disclosed

Post-training

SFTyes[11]
Preference optimisationother[12]
Reasoning trainingmulti-domain verifiable reasoning RL (Math → Code → STEM) with MGPO at a single 64K long-context window (progressive window expansion rejected); Long2Short Math RL redistributing rewards toward shorter correct trajectories; CLR (Claim-Level Reliability Assessment) claim-level test-time scaling — K=32 candidate trajectories, M=5 decision-relevant claims extracted per trajectory, self-verification verdicts mapped through a nonlinear reliability score (r_k = (mean verdict)^M), reliability-weighted answer aggregation — evaluated over 8 independent runs; curriculum CoT SFT with a 5K-token trace threshold[13]
Distillationoffline self-distillation: verified high-learning-potential trajectories (S_LP = length-normalized student NLL, bucketed per domain/length) from the model's own Math/Code/STEM RL checkpoints are distilled back into a unified student via SFT; note: the SFT-stage 'multi-path distillation' uses external strong-teacher models to generate training data, which the schema treats as teacher-generated data (see prose), not distillation[14]

Modality

Typetext-only
Attachmentn/a[15]

Inference efficiency

KV-cache designnot disclosed
Quantisation shippednone disclosed[16]
Speculative decodingnone disclosed[16]
Serving optimisationsnone disclosed[17]

Evaluation

Benchmarks (report's own numbers only)

BenchmarkValueRef
AIME26 (Pass@1, 64 samples)94.3Table 1
AIME25 (Pass@1, 64 samples)91.4Table 1
HMMT25 (Pass@1, 64 samples)89.3Table 1
BruMO25 (Pass@1, 64 samples)93.8Table 1
IMO-AnswerBench (Pass@1, 16 samples)76.4Table 1
LiveCodeBench v6 (Pass@1)80.2Table 1
OJBench (Pass@1)38.6Table 1
GPQA-Diamond70.2Table 1
IFEval93.4Table 1
IFBench74.5Table 1
LeetCode weekly & biweekly contests (Apr 25–May 31 2026, acceptance rate)96.1 (123/128)Table 3
AIME26 + CLR (Pass@1)97.1Table 2
AIME25 + CLR (Pass@1)96.7Table 2
HMMT25 + CLR (Pass@1)95.4Table 2
BruMO25 + CLR (Pass@1)99.2Table 2
IMO-AnswerBench + CLR (Pass@1)80.6Table 2
GPQA-Diamond + CLR72.9Table 2

Provenance references

  1. stated · Title; §1 ('Extending the technical lineage of the VibeThinker series'); prior work VibeThinker-1.5B [42]
  2. stated · arXiv header line ('arXiv:2606.16140v1 [cs.AI] 15 Jun 2026')
  3. stated · pipeline
  4. stated · Abstract ('a compact dense model with 3B parameters'); §4 ('comprising only 3 billion parameters')
  5. stated · Abstract ('a compact dense model with 3B parameters'); §2 ('Qwen2.5-Coder-3B base, a compact 3B dense foundation model')
  6. n/a · Abstract (dense model)
  7. stated · §2.2.2 ('we directly conduct RL with a single 64K long-context window')
  8. stated · §2.1.1, §2.2.2, §2.4
  9. stated · §2.1.2, §2.2.2, §2.3, §2.4
  10. stated · §2.1.2
  11. stated · §2.1 (two-stage curriculum SFT with Diversity-Exploring Distillation and domain-specialist checkpoint merging)
  12. stated · §2.2.1, §2.4
  13. stated · §2.2.2, §3.1
  14. stated · §2.3, §2.1.1
  15. n/a · §1–§3 (text-only model)
  16. stated · §2–§3
  17. stated · §3.1