- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 3(+1)
Read review
Paper summary
DiPPO trains the prompt proposer itself rather than keeping it fixed. A masked diffusion LM fills an instruction slot placed immediately before the query, so bidirectional attention conditions the proposal on the query that follows; only its LoRA parameters are updated, and the sole learning signal is the scalar task score of a frozen target LM on the resulting response. No external proposer LLM and no natural-language critic are required. DiPPO-HF starts from a fully masked slot and optimizes clipped GRPO over the denoising trajectory likelihood following d2, with group-relative advantages replacing a value model. DiPPO-HS places the human-written instruction in the candidate set as a competing item and applies DPO over reward-derived preference pairs, scoring a candidate by remasking-and-reconstruction. One instruction is selected from an archive on a disjoint development set after leakage and duplication audits, then applied unchanged at test time. With Dream-7B as proposer and Qwen2.5-7B as target over ten benchmarks, plus a sweep over seven target models, the human-free variant beats every human-free baseline and exceeds the human instruction on six tasks; the human-seeded variant averages 76.75, above GEPA with a 72B reflection model.
Critical strengths
-
The claim that actually matters here is demonstrated rather than asserted: a 7B open-weight diffusion model, updated only through LoRA and only through the task metric, can replace both the strong external proposer and the natural-language critic that current prompt-optimization pipelines depend on. Table 1 covers seven target models from Phi-3.5-mini to Qwen2.5-72B, and DiPPO-HF beats the no-instruction baseline on every target/task cell reported. Most papers in this line give one target model and call it general.
-
The failure analysis is honest and useful. Section 4.1 admits that two of the five seeds on HaluEval never move beyond the human instruction, and Section 4.3 goes further: in the human-free setting every candidate generated by LLaDA scored zero, so training never began. The backbone-variance discussion explains why the infilling baselines do poorly on two tasks. The archive plus development-set selection plus audits in Section 3.3 is also the right design — it prevents the deployed prompt from being the one that happened to win on its own training batch, which is the standard failure mode of reward-driven prompt search. Reporting failures at this level of specificity is why I trust the rest of the paper more than the tables alone would warrant; it is also exactly why the best-seed policy below matters.
Critical weaknesses
-
Best-seed reporting. Section 4.1 says "Each setting is run with five seeds fixed in advance, and we report the result of the best seed in all figures and tables", while the Figure 1 caption says "Error bars show the seed mean and standard deviation." These cannot both describe the same numbers. The paper's load-bearing claims are narrow-margin and count-based: "improves on the human-written instruction on eight tasks", "highest score on five", "within 1pp of the human instruction on GSM8K and MBPP", "averaged over all tasks, DiPPO-HS scores 76.75". Taking the maximum over five seeds inflates exactly these statements, and the HaluEval discussion tells us the spread is large enough to flip conclusions. The seeds already exist, so re-reporting mean±SD with paired comparisons is a fix, not new research. This is the main reason my score is where it is.
-
No matched target-model evaluation budget, so the headline wins may be compute rather than method. The baselines are given explicit budgets — GEPA keeps its official 6,871 rollouts on HotPotQA and 2,000 on the other single-module tasks — but nowhere in the main text can I find how many target-model evaluations, denoising steps, or updates a DiPPO run consumes. DiPPO's central selling point is that it needs no critic; the honest comparison at matched target-model calls could look quite different, especially against TextGrad and GEPA with Qwen2.5-7B as critic, where DiPPO's margins are small.
-
The "human-free" framing is softer than the abstract, and the largest reported gains depend on it. SQuAD2.0 and HaluEval require a proposer-side meta-instruction describing the task before training can start, because otherwise the proposer never finds the required output convention — with LLaDA on HaluEval, every candidate scored zero. Table 3 is explicitly labelled "results with the proposer-side meta-instruction", and the numbers there are the ones quoted in Section 4.1 as gains of 78.0pp and 80.5pp over the no-instruction baseline; on HaluEval, DiPPO-HF reaches 82.85 against 54.05 for the human instruction. So the two most spectacular improvements in the paper's headline rest on a human-written task description, which is arguably the very knowledge a prompt optimizer is supposed to rediscover. I accept the paper's defense that the meta-instruction states no answer format and no labeling rule. Still, the abstract's claim that the model can "discover prompts without a human-written seed instruction" should carry that qualification where the claim is made, not in Section 4.3, and Table 2/3 should report both conditions side by side.
-
Methodologically the contribution is a composition, which raises the bar on the evidence. Section 2.2 states directly that DiPPO-HF "adopt[s] the trajectory-likelihood-based GRPO of d2 without introducing a new diffusion reinforcement-learning algorithm"; DiPPO-HS is standard DPO with a reconstruction-score log-probability surrogate, and Section 3.2's surrogate needs several masks per candidate to estimate a prompt's probability, which is not the same object the DPO derivation assumes. That is a defensible engineering contribution, but then the empirical case has to be airtight, and with best-seed reporting and no budget control it is not yet.
Questions for authors
-
How many frozen-target-model evaluations does one DiPPO-HF and one DiPPO-HS run consume per task? Can you add score-versus-number-of-target-evaluations curves for DiPPO, GEPA and TextGrad on at least HotPotQA and FEVER? (Rebuttal-feasible analysis.)
-
In Section 3.2, the DPO score for a candidate prompt is a mean log-probability of reconstructing remasked tokens, averaged over several masks. How many masks, and how stable is the resulting preference ordering? A candidate pair whose scores swap under a different mask set would make the preference label noise rather than signal.
-
Please report DiPPO-HF on all ten tasks both with and without the proposer-side meta-instruction, so the human-free claim can be read off directly.
-
When the deployed instruction differs between the seven target models in Table 1, do the winners share structure? If the instruction discovered for Qwen2.5-7B transfers to Qwen3-30B-A3B, that would be a strong practical argument for the method; if not, the per-target retraining cost needs to be stated.
5 comments
W1, continued.
Table R3-W1.1. Human free prompt optimization. Figures are task scores, given as means with standard deviations. The DiPPO-HF row is the revised proposer input without the proposer side meta instruction. GEPA and TextGrad start from the DSPy default string with a Qwen2.5-72B-Instruct reflection model, and their HotPotQA cell does not yet have enough seeds, which we will update.
| Method | GSM8K | MMLU | MBPP | BFCL | SQuAD2 | CLINC | FEVER | HaluEval | Spider | HotPotQA |
|---|---|---|---|---|---|---|---|---|---|---|
| None | 84.76 | 60.19 | 60.40 | 36.14 | 0.00 | 0.13 | 61.94 | 0.00 | 65.70 | 24.67 |
| RLPrompt 82M | 82.21±0.50 | 57.07±0.58 | 61.92±1.03 | 36.60±5.76 | 0.00±0.00 | 0.01±0.02 | 76.58±1.97 | 0.00±0.00 | 64.16±0.73 | 28.47±0.99 |
| RLPrompt 7B | 83.64±0.59 | 55.66±1.84 | 60.88±0.94 | 40.26±2.72 | 0.00±0.00 | 20.01±2.35 | 76.95±2.44 | 0.64±1.18 | 62.36±1.19 | 28.47±0.99 |
| infill-Dream | 81.60±0.52 | 62.19±0.42 | 61.85±0.25 | 28.12±0.79 | 27.60±6.48 | 55.28±2.92 | 79.74±1.13 | 16.92±2.30 | 68.53±2.77 | 36.40±8.52 |
| infill-LLaDA | 83.17±0.72 | 56.67±7.60 | 62.50±1.59 | 33.06±4.16 | 14.71±17.40 | 13.68±23.34 | 19.69±34.11 | 0.00±0.00 | 65.51±0.57 | 32.80±11.26 |
| GEPA | 86.35±1.68 | 68.07±1.03 | 63.36±1.29 | 30.37±0.65 | 74.02±1.52 | 66.75±1.17 | 78.34±1.36 | 31.55±7.76 | 70.79±2.39 | — |
| TextGrad | 88.38±1.86 | 62.44±2.27 | 62.16±2.17 | 31.34±1.06 | 45.39±26.22 | 39.83±29.36 | 66.89±21.64 | 2.64±3.74 | 69.73±1.68 | — |
| DiPPO-HF | 91.28±0.28 | 70.31±0.18 | 64.40±2.59 | 57.46±2.51 | 76.24±1.40 | 73.86±0.65 | 81.29±0.20 | 74.78±4.56 | 73.73±1.02 | 60.53±1.98 |
Table R3-W1.2. Human seeded prompt optimization. Same convention as Table R3-W1.1. The DiPPO-HS row is the revised proposer input without the proposer side meta instruction.
| Method | GSM8K | MMLU | MBPP | BFCL | SQuAD2 | CLINC | FEVER | HaluEval | Spider | HotPotQA |
|---|---|---|---|---|---|---|---|---|---|---|
| Human | 91.81 | 69.82 | 67.20 | 72.98 | 67.63 | 72.38 | 73.68 | 54.05 | 69.05 | 34.00 |
| TextGrad 7B | 91.75±0.35 | 69.00±0.83 | 67.76±0.30 | 71.62±1.25 | 67.54±0.18 | 72.55±0.01 | 76.32±3.52 | 72.59±10.63 | 69.07±0.62 | 47.40±4.22 |
| TextGrad 72B | 89.77±1.37 | 69.13±0.67 | 67.56±0.62 | 73.09±2.17 | 71.83±2.68 | 72.28±0.07 | 80.02±0.29 | 60.87±10.59 | 72.86±1.67 | 41.73±5.62 |
| TextGrad Opus | 92.10±0.55 | 69.79±0.02 | 67.08±1.08 | 71.62±1.93 | 69.17±4.16 | 68.03±2.84 | 80.00±1.59 | 84.32±0.95 | 69.46±2.27 | 58.50±2.83 |
| DLM-PO | 91.89±0.28 | 69.80±0.03 | 67.35±0.44 | 74.02±1.23 | 74.76±5.01 | 72.17±0.25 | 80.70±0.27 | 57.27±4.75 | 68.20±1.15 | 34.00±2.66 |
| GEPA 7B | 90.96±0.86 | 69.75±0.24 | 67.28±0.50 | 70.67±4.50 | 68.99±2.67 | 74.03±0.70 | 80.58±0.69 | 82.76±0.69 | 68.07±1.29 | 57.27±2.32 |
| GEPA 72B | 90.64±1.10 | 68.81±0.90 | 66.28±1.74 | 74.04±3.25 | 76.35±2.17 | 73.91±0.82 | 81.18±0.60 | 81.00±1.22 | 70.64±2.70 | 58.00±4.23 |
| GEPA Opus | 91.11±1.28 | 69.80±0.02 | 67.60±1.37 | 72.82±1.69 | 74.14±1.29 | 73.42±2.50 | 79.78±1.42 | 82.81±1.78 | 68.92±0.37 | 61.07±1.75 |
| DiPPO-HS | 91.55±0.28 | 69.80±0.36 | 67.80±1.46 | 76.79±1.55 | 78.79±0.68 | 73.74±0.85 | 81.04±0.54 | 63.88±15.23 | 73.08±1.22 | 61.44±2.01 |
Q1. Target-model evaluations per run, and score against budget for DiPPO, GEPA and TextGrad.
A. Thank you for raising the budget question. We report the target-model calls per run and performance at matched budget checkpoints.
Calls per run. On a single-module task, DiPPO-HF uses 360,000 target calls in training plus 21,600 for development selection (381,600 total); DiPPO-HS uses 405,000 plus 2,400 (407,400 total). Test evaluation is excluded for every method.
Performance against evaluation budget on FEVER.
GEPA and TextGrad both start from the human instruction here, which makes DiPPO-HS the direct comparison. We also report DiPPO-HF, which starts without one. At each checkpoint we select an instruction on development performance, then evaluate on test.
| FEVER evaluation budget | TextGrad 7B | GEPA 7B | DiPPO-HS | DiPPO-HF |
|---|---|---|---|---|
| 1,000 calls | 80.68 ± 0.56 | 75.80 ± 2.06 | initial instruction 73.71 | initial instruction 60.37 |
| 10,000 calls | 80.33 ± 0.87 | 80.44 ± 1.37 | 73.71 ± 0.00 | 60.37 ± 0.00 |
| 100,000 calls | 80.27 ± 0.95 | 80.04 ± 1.08 | 80.49 ± 0.33 | 80.66 ± 0.31 |
| final evaluated point | 80.27 ± 0.95 | 80.24 ± 0.88 | 80.95 ± 0.57 | 81.09 ± 0.35 |
Test 16,039 examples. The baselines were extended to 405,000 calls, while the DiPPO entries are each run's final selected instruction, so the last row reports endpoints rather than an exactly matched count. DiPPO's first measured checkpoints are at 2,700 (HS) and 2,400 (HF) calls; its first update does not complete until then, which is why the 10,000-call entries still deploy the initial instruction.
GEPA reaches strong performance quickly, above DiPPO-HS at 10,000 calls, and TextGrad is also strong early. DiPPO uses more target evaluations to reach a comparable final mean. The final points lie within one standard deviation of each other, so the three are at a comparable level on FEVER.
| Change from 10,000 calls to the final point | GEPA 7B | TextGrad 7B | DiPPO-HS |
|---|---|---|---|
| FEVER | −0.20 | −0.06 | +7.23 |
| SQuAD 2.0 | +2.49 | not extended | +11.87 |
Extending GEPA and TextGrad to 405,000 calls produces almost no change on FEVER, while DiPPO-HS improves by 7.23 points after 10,000 calls; on SQuAD 2.0 the figures are +2.49 and +11.87. GEPA makes rapid early gains, while DiPPO improves through additional training. Time constraints prevented extending TextGrad and the HotPotQA baselines to the same budget.
SQuAD 2.0.
| SQuAD 2.0 evaluation budget | GEPA 7B | TextGrad 7B | DiPPO-HS | DiPPO-HF |
|---|---|---|---|---|
| 1,000 calls | 67.49 ± 0.22 | 66.21 ± 3.77 | initial instruction 67.52 | initial instruction 0.00 |
| 10,000 calls | 71.08 ± 1.27 | 68.16 ± 6.54 | 67.52 ± 0.00 | 0.00 ± 0.00 |
| 100,000 calls | 73.47 ± 2.61 | — | 73.99 ± 5.91 | 76.90 ± 1.13 |
| final evaluated point | 73.57 ± 1.32 | 68.16 ± 6.54 | 79.39 ± 0.99 | 76.87 ± 1.08 |
Test 11,873 examples. DiPPO-HF starts from an empty instruction, hence the initial 0.00. TextGrad was not extended to 405,000 calls; its final entry is the endpoint of its own configuration, and its call counts are upper bounds computed before candidate deduplication. Additional training raises DiPPO-HS from 67.52 at 10,000 calls to 79.39, and GEPA from 71.08 to 73.57. DiPPO-HS finishes 5.82 points higher.
HotPotQA.
| HotPotQA evaluation budget | GEPA 7B | TextGrad 7B | DiPPO-HS | DiPPO-HF |
|---|---|---|---|---|
| 1,000 calls | 54.56 ± 1.68 | 47.13 ± 10.50 | initial instruction 33.67 | initial instruction 25.67 |
| 10,000 calls | 55.33 ± 1.45 | 52.67 ± 2.95 | 39.67 ± 10.39 | 32.56 ± 11.93 |
| final evaluated point | 55.33 ± 1.45 | 53.27 ± 3.69 | 53.44 ± 3.36 | 59.00 ± 0.67 |
| final call count | 7,816–9,256 | 18,273–20,433 | 204,900 | 84,000 |
Test 300 examples; DiPPO's first measured checkpoints are at 3,750 (HS) and 7,712 (HF) calls. The DiPPO entries use the submission configuration, which does not give the reference answer to the proposer.
DiPPO improves here too, but GEPA finishes above DiPPO-HS far earlier. We acknowledge that GEPA gains more quickly at small budgets, and improving DiPPO's efficiency there is future work.
On cost. All comparisons here are open-weight 7B, and DiPPO uses more target rollouts. Its final mean is higher on FEVER and SQuAD 2.0, and for DiPPO-HF on HotPotQA, where DiPPO-HS falls below GEPA. Separately, GEPA and TextGrad can do better with a commercial model as critic or optimizer, which adds API cost for textual feedback on top of target evaluation; DiPPO trains an open-weight proposer from a scalar reward and makes no commercial critic calls. Proposer training and generation also cost, so this does not establish a lower total cost. The right choice depends on rollout budget, API cost and time; we will work on reducing DiPPO's rollout requirement. We thank the reviewer for prompting this analysis.
W1. Best seed reporting, its relation to Figure 1, and the results as mean and standard deviation.
A-W1. We thank the reviewer for putting this so precisely, and we apologize for not having described this clearly. We reported the maximum because deployment uses a single instruction rather than an average over runs, namely the one with the highest development set score. The same procedure was applied to the methods we compare.
We agree, however, that the best seed value is only a summary of the performance a method can reach, and that it is not an adequate statistic for claiming a small difference between methods, a count of tasks that beat the human instruction, or an advantage on the overall average. On a task such as HaluEval, where the spread across seeds is wide, no conclusion can rest on a single best seed. The revision therefore reports every main comparison as the mean and the standard deviation of the five test scores, and we rewrite claims in the text such as highest and leads on so many tasks to match that statistic.
A separate update: the revised proposer input. The DiPPO of the original submission was an answer blind condition, in which the proposer learned a prompt from the query x and a scalar reward alone. Under that condition the proposer could infer the output convention and the correctness criteria only indirectly, from the score, and on some tasks we needed a separate experiment with a proposer side meta instruction to compensate. GEPA and TextGrad differ in implementation, but both give the optimizer textual feedback produced by comparing a prediction with the answer.
In the revision we therefore give the proposer the reference answer y that is already in the training data as a conditioning input, that is qϕ(p∣x,y). This does not reproduce the feedback mechanism of GEPA or TextGrad. It is a supplement to the input, so that the proposer can reach the reference answer signal during training.
y is given to the proposer only, and only during training. The frozen target still receives [p;x] only, and candidate rewards are computed by the same scorer as before. Development set selection is also unchanged: it compares the target's predictions with the development answers, and the proposer does not read an answer as input at development or at test time. At test time the target has no access to the answer. Apart from the proposer conditioning input, the reward, the GRPO and DPO objectives, the archive and candidate selection, the target side inference and the evaluation procedure are all the same.
In our experiments this supplement reduced the run to run spread, most clearly on the tasks that require discovering an output convention. We will update Figures 1 and 2, the definition of the method, the formulation and the DiPPO result tables accordingly. We would emphasize that this update changes only the proposer’s conditioning input; the core GRPO/DPO optimization objectives and the target-side pipeline remain unchanged.