Users thank the REVES project leads and team for their work on sequential revision that boosts Qwen models on coding and math benchmarks.
Based on 1 visible X reactions from 1 accounts; directional sample.
Ask a question below.
Published answers will appear here.
Huge thanks to @Yuanxin02 for leading this projects, and the whole team @ZhouRuida, Xinyan Zhao, @amr_n_sharaf, @hongzhou__lin, Arijit Biswas, Mohammad Ghavamzadeh, and @MingyiHong01!
Does it work? Consistent improvements on both coding and math, across Qwen2.5-3B, Qwen2.5-7B, and Qwen3-4B: ▸ LiveCodeBench 29.5 (REVES) vs 23.0 (RL) ▸ AIME24 45.7 vs 33.5 ▸ AIME25 40.5 vs 22.9 The attached table has the complete picture — every model, every test-time protocol (OneShot, Oracle-B, SelfConf-B, execution-verified TC-B). In the sequential-revision columns REVES is on top almost across the board. [5/9]
Where does that clean split come from? The sequential-revision objective factorizes exactly into one-step recovery probabilities, one per visited state. Improve recovery anywhere → the whole objective improves, and horizon-length credit assignment disappears. In practice it's a loop that feeds itself: 🔧 Stage I: roll out revisions with the current policy, keep the recovering ones, and turn every near-miss into a "fix this" prompt plus a "does this check out?" verification prompt. ⚡ Stage II: ordinary single-turn RL over that augmented prompt set. Since Stage II is off-policy and single-turn, the whole thing is also cheaper to run than multi-turn RL. [4/9]
Here's the part that makes this a capability rather than a trick: every harness in this family, at some point, asks the model to improve a previous attempt. So training that one skill pays off everywhere. One checkpoint, trained purely on sequential revision, comes out ahead under MCTS, three AB-MCTS variants, and Mind Evolution — none of which it ever saw in training. There's a conditional transfer bound in the paper, and the experiments back it up. [7/9]
Revision is a generalized capability — not a per-domain trick. Take checkpoints trained only on math and code, and drop them zero-shot onto n_queens and mini_sudoku. REVES gives the biggest jump on both — no puzzle data, no puzzle-specific tuning. Learning to look at your own attempt and repair it is a skill in its own right, and it travels to tasks the model has never trained on. [8/9]
Revision is a generalized capability — not a per-domain trick. Take checkpoints trained only on math and code, and drop them zero-shot onto n_queens and mini_sudoku. REVES gives the biggest jump on both — no puzzle data, no puzzle-specific tuning. Learning to look at your own attempt and repair it is a skill in its own right, and it travels to tasks the model has never trained on. [8/9]
The takeaway: we deploy LLMs inside loops — revise, search, evolve — and then train them as if a single forward pass were the whole job. REVES flips that: make the training objective the deployment objective. Do it once, and the gain shows up across the entire family of revision-style harnesses. [9/9] 📄 https://arxiv.org/abs/2606.18910 💻 https://github.com/yxliu02/REVES
More interesting result in the paper: 26-circle packing (maximize the summed radii). With REVES training, a Qwen3-4B hits 2.635983 — the best value reported anywhere, previously held by evolutionary-search systems running Gemini-2.0 Pro/Flash or 8B backbones. Half the parameters, fewer rollouts, same record. [6/9]
Users thank the REVES project leads and team for their work on sequential revision that boosts Qwen models on coding and math benchmarks.
Based on 1 visible X reactions from 1 accounts; directional sample.
Ask a question below.
Published answers will appear here.
The takeaway: we deploy LLMs inside loops — revise, search, evolve — and then train them as if a single forward pass were the whole job. REVES flips that: make the training objective the deployment objective. Do it once, and the gain shows up across the entire family of revision-style harnesses. [9/9] 📄 https://arxiv.org/abs/2606.18910 💻 https://github.com/yxliu02/REVES