Selection Is Not Learning: Choosing the Future for Self-Evolving Language Models
Selection Is Not Learning: Choosing Model Futures
EVOHEIR uses a new probe to decide inheritance; after ten rounds it has the highest average math score across four backbones, with Qwen3-8B at +7.0% relative to Agent0.
- Self-evolutionary models generate their own problems and train on their own outputs, so each round they must decide which update to inherit. R-Zero, Agent0, and Skill-SP admit tasks by learnability and inherit all updates, but learnable tasks do not necessarily produce useful updates, and regressions can enter the subsequent curriculum.
- EVOHEIR explores native branches each round, filters the parameter recipient, repairs failures and retains successes using current and archived experience, then confirms with new probes whether to inherit. HEIRCOMMIT commits only when the frozen successor is better than the current root; eligible experience is archived regardless of submission.
- After ten rounds on four backbones, EVOHEIR has the highest math and general reasoning averages, with Qwen3-8B-Base's Math Avg. at +7.0% relative to Agent0 (Table 1). HEIRCOMMIT reduces the proportion of harmful submissions across three hosts to 13.33%–16.00%.
- The paper does not have a separate section for limitations or future work. Appendix B.3 states: verified defect references can bias repair, absorption may lose successes, and zero-threshold sampling noise may include harmful proposals. The main experiments use four backbones, ten rounds, and three seeds; ablations and equal-budget comparisons are on Qwen3-4B-Base.
- Evaluated models
- Qwen3-4B-BaseQwen3-8B-BaseOLMo-3-7B (allenai/Olmo-3-1025-7B)Gemma-4-E4B-it
- Benchmarks
- MATH-500MinervaOlympiadBenchOmni-MATH-2MATH-BeyondAIME 2025AIME 2026SuperGPQAMMLU-ProBBEHLiveBench ReasoningKOR-Bench
- Metrics
- Math Avg.General Avg.exact-answer accuracypass@1harmful commitscorrect-target NLL gainabsorption gainselected-target optimism
The paper points out that existing self-evolving language model training loops have a gap from selection to update, and proposes two inheritance decision methods, HeirCommit and EvoHeir. Loops such as R-Zero, Agent0, and Skill Self-Play decide which version of the update to inherit based on the task itself, accepting tasks that appear learnable and inheriting all their updates, but most updates on validation tasks do not improve the probability of answering correctly. HeirCommit freezes the updated model and uses new evaluation questions to decide whether to inherit or keep the current model; EvoHeir separates parameter inheritance from experience inheritance, first using current and archived experience to fix failures, retaining successes, and then making a final confirmation. The analysis characterizes selection-induced optimism and decomposes regret into three types of errors: selection, absorption, and confirmation. The authors report that at the same acceptance rate, HeirCommit reduces the proportion of harmful updates by 68-83% relative to random inheritance; on four backbone models, EvoHeir achieves the best average performance in math and general reasoning, with math averaging 5.3% higher than the strongest baseline Agent0.
Paper analysis
What problem does this paper attempt to solve?
Self-evolution inherits updates by task learnability, but learnable tasks do not necessarily bring a useful model future.
Self-evolution must decide each round which updated model to inherit, but filtering by task learnability fails to recognize useful updates.
- Scenario: The model generates its own problems and trains on its own outputs. The authors state that inherited updates simultaneously change the learner and the subsequent curriculum, so progress or regression accumulates; self-training can also first raise a target metric and then crush it within the same training run (Lin, 2026).
- Existing shortcomings: R-Zero, Agent0, and Skill Self-Play (Skill-SP) make decisions per task: they admit tasks that look learnable and inherit every resulting update. The authors state that pass rates, divergence, and frontier proximity describe pre-training behavior, not the transfer produced by optimizer updates.
- Core observation: The authors call this the selection-to-update gap. Under independent answers, 52.2% and 42.0% of Agent0's and R-Zero's native admitted tasks leave the frontier (Section 6.1; the introduction summarizes as 42–52%). Even when tasks are exactly valid, 60 of 96 Skill-SP updates have non-positive correct-target negative log-likelihood (NLL) gains.
- This paper proposes: HEIRCOMMIT freezes the successor and uses new problems to decide whether to inherit it or keep the current model (the root). EVOHEIR separates parameter inheritance from experience inheritance: filtering decides the parameter starting point of the next learner, then uses current and archived references to repair failures and retain successes; confirmation occurs after absorption. Eligible experience enters the archive regardless of whether that round is submitted.
What related research exists?
Related work covers self-evolution acceptance, privileged distillation, and experience replay; this paper separates selecting the recipient from retaining experience.
The paper divides related work into three groups: self-evolution and update acceptance, privileged context distillation, and experience persistence, and accordingly separates recipient selection, experience donors, and post-absorption confirmation.
Self-evolution and update acceptance
- Self-generated supervision: STaR, Self-Instruct, SPIN (Zelikman et al., 2022; Wang et al., 2023; Chen et al., 2024) learn from self-generated supervision; Alami et al. (2024) regularize SPIN with a KL term and cross-iteration fictitious games.
- Tasks and curricula: Absolute Zero, SPELL (Zhao et al., 2025; Yang et al., 2026) generate tasks via self-play, and DataEnvGym (Khan et al., 2025) uses student feedback. R-Zero, Agent0, and Skill-SP use adaptive curricula. Earlier automatic curricula select problems by measured learning progress (Graves et al., 2017; Matiisen et al., 2020); the paper paraphrases Willems et al. (2020) as arguing that signal is flawed and proposing mastery rate.
- Whether to accept updates: SOAR and Actor-Curator use realized improvements (Sundaram et al., 2026; Gu et al., 2026). Update acceptance appears in AlphaGo Zero but not AlphaZero (Silver et al., 2017b;a). APE evaluates candidate fine-tunes before acceptance (Marín, 2025); PACE does non-parametric agent evolution (Shawn, 2026). Shang et al. (2026) and Sengupta (2026) verify distilled skills before submission, or provide certificates for adapter and harness updates, while the base remains frozen. Lin et al. (2026) discuss that harness updates may not bring gains either.
Privileged context distillation
- Student prefixes: GKD trains on student-generated prefixes (Agarwal et al., 2024).
- Conditional self-teachers: SDFT, SDPO, OPSD have the self-teacher condition on demonstrations, feedback, or privileged context (Shenfeld et al., 2026; Hübotter et al., 2026; Zhao et al., 2026); OPCD internalizes contextual knowledge (Ye et al., 2026b).
Experience persistence
- Replay and merging: experience replay retains past experience (Rolnick et al., 2019); genetic distillation merges policies (Gangwani & Peng, 2018).
- Internalizing accumulated experience: online experience learning and SPEE internalize accumulated experience into the model (Ye et al., 2026a; Ren et al., 2026).
Baselines and benchmarks
- Baseline methods: The ten-round comparison includes R-Zero, Agent0, and No Recovery, which removes absorption while keeping exploration, filtering, and confirmation. HEIRCOMMIT audit controls are commit-all and matched random at the same acceptance rate.
- Benchmarks: Seven math benchmarks MATH-500, Minerva, OlympiadBench, Omni-MATH-2, MATH-Beyond, AIME 2025, AIME 2026; five general reasoning benchmarks SuperGPQA, MMLU-Pro, BBEH, LiveBench Reasoning, KOR-Bench.
The authors position EVOHEIR as: separating the selection of parameter recipients and experience donors, then confirming the successor after absorption. The authors state that Recovery combines recipient-specific repair and retention, cross-branch references, fixed recipient trajectories, and post-absorption confirmation; experience is re-evaluated against the current recipient and retained independently of whether parameters are accepted, with rejected states still able to serve as donors.
How does the paper address this problem?
First filter parameter recipients, then absorb current and archived experience, and finally use a new probe to decide whether to inherit.
EVOHEIR gives exploration two paths into the next learner each round: the selected parameters and the absorbed experience. The four steps correspond to Figure 3 and Algorithm 1: explore native branches from the root, filter the parameter recipient, absorb current and archived experience into that recipient, then confirm against the root with new probes.
Environment and Value
- State of a round: The root state is St=(θt, ωt), where θt are parameters and ωt is the recoverable training state (optimizer moments, learning rate scheduler, random number generator). The Curriculum Scout generates problems, the Evolution Controller admits based on the Evolving Learner's frontier, and native training updates two agents in an R-Zero-style loop (Figure 2). Each round trains two isolated native branches from the root.
- Disjoint data roles: Training, Recovery, filtering, confirmation, and final evaluation use disjoint problem pools. Online scores use frozen consensus labels. Appendix A: 1,024 problem generations per round; native admission uses ten answers per problem; each branch has a 64-problem cohort. Consensus support is the proportion of the largest answer cluster among valid answers, with probes requiring at least 0.7; problems with no valid answers are excluded.
- Value: For a declared distribution Pt and utility q, the value of a state relative to the root is Vt(S)=E[q(S;z)−q(St;z)], where z is drawn from Pt, and the root's value is 0. Capability uses exact correctness or sampled pass@1; update audits use the decrease in correct-target NLL, which the authors state is not an accuracy difference. Under binary correctness, value equals the probability of newly gained successes minus the probability of lost successes.
- Three-step decision: Filtering selects the recipient from native endpoints and the root; Recovery turns it into a proposal; confirmation either keeps the root or adopts the proposal. Transfer regret is the maximum value among constructed states minus the value of the actually inherited state. Absorption gain is the value change of the proposal relative to the recipient.
Filtering the Recipient and Recovering Experience
- Who provides parameters: Native endpoints and pre-declared training prefix checkpoints form the candidates. The controller freezes candidates before scoring on a shared filtering pool. Paired root-relative scores select the recipient; ties favor the root first, then a fixed branch order. Filtering can select the root itself, so the next learner's parameter starting point need not come from a branch endpoint. The filtering pool is 16 problems, with four answers per state per problem, temperature 0.7, top-p 0.95 (Appendix A).
- Who provides experience: The Experience Donor contributes only eligible answers; it need not provide the inherited parameters or continue as a teacher. Current exploration records and a bounded archive provide problems and references, annotated with generation checkpoint, round, and verification trajectory. Eligibility checks completeness, answer validity, and problem-level deduplication. The archive holds at most 256 entries, deduplicated by normalized problem (Appendix B.4).
- Grouping by recipient: The recipient answers without references, using exact checkpoint caching or batched rollouts. Failures with eligible references enter the repair group; successes enter the retention group. The two groups are mutually exclusive and recipient-specific: archived successes that the current recipient cannot reproduce become repair material. Each round nominates at most 32 exploration records and 32 learner-success records; the recipient attempts within a budget of 64 generations.
- Frozen recipient as teacher: Both teacher and student initialize from the selected parameters and predict the next token on a fixed prefix already generated by the recipient. The repair teacher additionally sees one donor reference; the retention teacher and student see only the problem. The student matches the repair teacher on failed attempts and the retention teacher on successful attempts. The loss is Lrec(θ)=⟨d(Trepair,Pθ)⟩repair+λkeep⟨d(Tkeep,Pθ)⟩keep. d is forward KL over the full vocabulary, averaged per problem within each group. Reported runs use lambda_keep=1, LoRA rank 8, scaling 16, target projections q_proj and v_proj, learning rate 5×10^−5, up to 16 updates (Appendix B.4). If the experience pool is empty, the recipient is unchanged. Only the resulting learner is deployed.
HEIRCOMMIT
- Confirming what will actually remain: including changes from absorption. The final proposal is frozen before scoring on a held-out confirmation pool. The proposal is adopted only if the paired value estimate on the confirmation pool exceeds a threshold; otherwise the root is kept. The reported runs use a threshold of 0, meaning the gain must be strictly positive. A positive confidence margin corresponds to the certification mode of Theorem 1.
- No retry after rejection: On rejection, the full root state is restored, including scheduler and random number generator, without falling back to the recipient or testing another proposal. A proposal identical to the root is a zero-value no-op. The confirmation pool is 32 new problems, with four answers per state per problem (Appendix A).
- Independent of construction: External hosts can apply HEIRCOMMIT to a single native update; EVOHEIR handles construction and search. Accepted absorption weights are merged into the next native root, and the native optimizer and scheduler are reinitialized. Eligible experience is archived regardless of submission. Appendix H gives message formats for the Scout, problem-solving inputs, and repair teacher; filtering and confirmation have no natural language judge prompts.
Identifiability and Regret
- Task-side non-identifiability: Proposition 1 states that for a loss-decrease utility and any positive value gap, any persistent rule based only on pre-training behavior observations faces two differentiable learning worlds with identical observations, target validity, and native step sizes, but values of opposite sign. Relative to "the better of that update and the root," worst-case expected regret is at least half that gap. The observation interface does not include update gradients.
- Filtering optimism and separate pools: Proposition 2 states that even if each candidate score is unbiased, the expected score of the filtering winner is no less than its true value. Theorem 1: reusing filtering to accept a branch pays a union bound over the number of branches; after absorption, confirming an already fixed proposal with new probes gives an exponential bound that does not contain the number of branches. The authors state that the zero threshold used in reports trades acceptance sensitivity for this positive-margin guarantee; Appendix C states that in this mode the exponential bound is 1.
- Regret decomposition: Corollary 1 states that within the realized candidate set, regret is at most twice the filtering error plus absorption loss, confirmation error, and margin. No Recovery has zero absorption loss. Across rounds, the sum of per-round error budgets does not exceed the total budget, yielding simultaneously valid bounds (Appendix C.3).
What experiments did the paper conduct?
After ten rounds, EVOHEIR has the highest average scores on both metrics across four backbones; HEIRCOMMIT reduces the proportion of harmful commits across three hosts.
Experimental Setup
- Backbones and protocols: Qwen3-4B-Base, Qwen3-8B-Base, OLMo-3-7B (checkpoint in Appendix A as allenai/Olmo-3-1025-7B), Gemma-4-E4B-it (google/gemma-4-E4B-it). The first three are pretrained base models; Gemma is an instruction-tuned starting point. Each method runs ten rounds with three training seeds, evaluating the actually inherited checkpoint after each round, including rounds with a retained root.
- Host audit: HEIRCOMMIT is applied to native updates from R-Zero, Agent0, and Skill-SP, with inheritance decided by an ensemble adapter. Each host has 64 updates, with a scoring pool disjoint from the decision probes; there are also 96 audits of exactly valid Skill-SP tasks. The paper does not state the backbone for this set of audits.
- Baselines: The ten-round comparison includes R-Zero, Agent0, No Recovery (removing absorption, keeping exploration, filtering, and confirmation), and EVOHEIR (Full). The inheritance audit controls are commit-all and matched random at the same coverage. Skill-SP is not in the ten-round comparison of Table 1.
- Benchmark sizes: After deduplication, MATH-500 500, Minerva 272, OlympiadBench 674, Omni-MATH-2 4,159, MATH-Beyond 146, AIME 2025/2026 30 each. SuperGPQA 26,529, MMLU-Pro 12,032, BBEH 4,520, LiveBench Reasoning 200, KOR-Bench 1,250 (Appendix E).
- Metrics and evaluation: Math Avg. and General Avg. are equally weighted over their respective benchmarks. The first five math benchmarks and the general reasoning benchmarks use one greedy answer and exact-answer accuracy. AIME uses 32 answers per problem per year (temperature 0.6, top-p 0.95, top-k 20), estimating pass@1 via average answer correctness. LiveBench averages its three Reasoning tasks, and KOR-Bench averages its five categories. Update audits use correct-target NLL decrease. Filtering and confirmation are numeric paired gains; the paper states that no natural language review prompts and no model-generated acceptances or rejections were used in the runs (Appendix H.4).
- Replication: Scores in Table 1 are means over three seeds; ± is marked only on Math Avg. and is the sample standard deviation over three seeds. The intervals in Figure 7d are pointwise 95% intervals from 2,000 stratified paired bootstraps over three seed groups. These ten-round results are fixed-round comparisons (Section 6.3).
Main Results
Math Avg. after ten rounds (%, from Table 1). ± is the sample standard deviation over three seeds; Base has no standard deviation.
Swipe horizontally to read the table
Backbone Base R-Zero Agent0 No Rec. EVOHEIR Qwen3-4B-Base 25.81 28.79±0.65 30.01±0.57 30.29±0.40 31.82±0.55 Qwen3-8B-Base 28.06 32.65±0.65 34.11±0.63 34.35±0.38 36.51±0.67 OLMo-3-7B 19.92 21.73±0.61 23.35±0.44 23.50±0.56 24.94±0.48 Gemma-4-E4B-it 43.77 45.11±0.50 46.66±0.59 46.53±0.42 47.21±0.44 - Authors' summary and different phrasings: The authors state that EVOHEIR (Full) achieves the highest Math Avg. and General Avg. on every backbone. The vs. Agent0 column in Table 1 is +6.0%, +7.0%, +6.8%, +1.2%. The authors state that on the three base models it exceeds Agent0 by 1.59–2.40 percentage points, with the smallest margin on Gemma, still leading by 0.55 percentage points; the Math Avg. improvement over the base is 7.9–30.1%, the largest among compared methods. The abstract writes a 5.3% improvement in math average over Agent0; the conclusion writes up to 7.0%. Figure 1 additionally marks the ten-round math gain averaged over the four backbones: EVOHEIR +5.7, Agent0 +4.1, with R-Zero peaking at +2.7 in round 8 and then declining. These three numbers are not the same.
- General Avg.: In Table 1, EVOHEIR is 30.52, 36.48, 21.40, 40.93, corresponding to Agent0 at 29.51, 35.24, 20.05, 40.37, and No Recovery at 29.82, 35.75, 20.34, 40.73, in order of the four backbones. R-Zero's General Avg. is lower than Agent0's in all four rows. In the table, only Gemma's No Recovery Math Avg. is lower than Agent0's; on the other three backbones, No Recovery is slightly higher than Agent0.
- Per-benchmark and trajectories: The authors state that EVOHEIR is higher than Agent0 in 45 of 48 benchmark–backbone cells, with the three exceptions all being the 30-problem AIME sets (Figure 5). The figure shows negative values of −0.28, −0.73, −0.77; the text does not specify the year per cell, so years are not assigned here according to possibly misaligned columns. In Figure 6, Full is highest at round 10 on all four backbones, with bands as ±1 standard deviation over three seeds, and each panel has a different y-axis scale.
Task Selection and HEIRCOMMIT
- Selection is not learning: The optimism of selected targets, i.e., support on the answers used to select them minus support on new answers, is 8.60 percentage points for Agent0 and 4.10 for R-Zero; fixed-target controls are 0.20 and −0.10 percentage points (Section 6.1). Under new answers, the proportion of admitted tasks leaving the native frontier is 52.2% for Agent0 and 42.0% for R-Zero. Among 96 exactly valid Skill-SP updates, 60 (also written as 62.5% in the paper) do not decrease the correct-target NLL (Figure 4b).
- Harmful submissions: In Table 4, HEIRCOMMIT's coverage is 23.44% for R-Zero, 39.06% for Skill-SP, and 51.56% for Agent0 (summarized in the text as submitting 23–52%). The proportion of harmful submissions drops from 76.56%, 57.81%, 46.88% of all submissions to 13.33%, 16.00%, 15.15% (Figure 4c rounds to 77/58/47 and 13/16/15). Matched random maintains the harmful proportion of all submissions. The authors state a reduction of 68–83% relative to it. The retention rate of useful updates is 86.67%, 77.78%, 82.35%, summarized in the text as 78–87%.
- Pool gains: In Table 4, the pool-average NLL gain (10^−3, higher is better) changes from −23.17, −3.47, +2.80 for all submissions to +10.19, +2.49, +14.35 for HEIRCOMMIT, in order R-Zero, Skill-SP, Agent0. The authors report a sign flip from negative to positive on two hosts and state that Agent0's gain increases fivefold. The backbone for this set of audits is not stated in the paper.
Does Experience Persist with the Model
- Weak donors can still be useful: 16 interventions, four fixed recipients from rounds 3, 5, 7, 10, each with two current donors and two archived donors; each time the recipient is restarted and evaluated on the same 1,024 problems before confirmation, with gains relative to that recipient. Eight of ten donors weaker than the current root still improve the recipient; their donor gains average −1.09 percentage points, and absorption gains average +0.71 percentage points (Figure 7a).
- Repair and retain: On the same 64 references, Cache SFT gains +2.08 percentage points in new successes and loses 2.61, net −0.54; Recovery gains +2.54 in new successes and loses 1.15, net +1.39 (Figure 7b). The authors state that Recovery is higher than Cache SFT at all four checkpoints, on average by 1.93 percentage points; Recovery itself has a net gain of −0.20 percentage points at round 7.
- Rejected rounds: With a fixed set of 64 references, current material alone gives +0.39 percentage points, adding history from submitted rounds gives +0.73, and including all eligible history (including rounds where the model was rejected) gives +1.39 (Figure 7c). The contribution list states that experience from rejected rounds adds another 0.66 percentage points beyond submitted history. In matched replacement, non-inherited sources still outperform same-age inherited records by 0.34–0.73 percentage points at ages 1 to 4, with pointwise 95% intervals above zero (Figure 7d). Figure 8: Qwen3-4B-Base round 5, an answer corresponding to a round-4 archived reference that was not inherited; Recovery reproduces it after removing the reference; under the same budget, Cache SFT changes a remainder problem the recipient originally solved correctly, while Recovery keeps the original answer. The new-probe gain for that round is −0.31 percentage points, and HEIRCOMMIT retains the root.
Other Ablations and Analyses
- Component ablation only on Qwen3-4B-Base, three seeds (Table 2): Full's Math / General is 31.82±0.55 / 30.52±0.35; Same-Lineage 30.77±0.57 / 29.99±0.35; Random Experience 30.54±0.44 / 29.87±0.43; Current-Round Only 30.85±0.69 / 30.04±0.35; NoConfirm 29.80±1.13 / 29.33±0.65. The authors state that removing any mechanism lowers both averages; the first three have Math lower by 0.97–1.28 percentage points and General lower by 0.48–0.65; NoConfirm has the largest drop, at 2.02 and 1.19 percentage points.
- Equal accounting budget (Qwen3-4B-Base, ten rounds, three seeds): single-stream native baseline 100.0 units, No Recovery default 111.9, Full default 123.7, with the authors stating a construction overhead of 23.7% (Appendix G). Full has the lowest mean at 0.8× the default budget and the highest from 1.0× onward (Figure 9a). Units are not measured GPU hours or FLOPs.
- Candidate count and confirmation pool: At budget 1.0, four candidates give the largest mean, close to the default of two; confirmation pools of 32 and 64 problems are also close (Figures 9b, 9c). The text does not give Math Avg. readings for each setting. Default filtering is 16 problems, confirmation 32.
- Retention weight: The default lambda_keep=1 gives the largest net absorption gain (Figure 9d). On a fixed recipient with 2,400 problems (initially solving 1,224, i.e., 51%), net gains over three absorption-order seeds are Recovery 3.57±0.30, Repair only 2.26±0.45, Cache SFT 1.11±0.86 percentage points (Appendix F).
- Wall-clock time: On two approximately 96 GB GPUs, Full takes about 8–9 hours per round, about 80–90 hours for ten rounds; benchmark evaluation totals about 4–5 hours (Appendix G).
What are some points that can be further explored?
The paper does not list limitations separately; the appendix notes verification misses, absorption losing successes, and the zero threshold potentially admitting harmful proposals.
Limitations and Future Directions Noted by the Authors
The paper has no dedicated Limitations, Discussion, or Future Work sections, and does not write out a future work plan. The failure modes explicitly written in the appendix are as follows.
- Defective references: Appendix B.3 writes that incomplete or conflicting references are excluded before training, but defective references that pass the verifier still guide the repair teacher, so repair depends on verification quality.
- Absorption loses successes: Appendix B.3 writes that absorption can also lose existing successes, pointing to the net gain of −0.20 percentage points at round 7 in Figure 7.
- Zero threshold: Appendix B.3 writes that under the empirical zero threshold, sampling noise may admit harmful proposals; Theorem 1's positive-margin mode provides a bound for this risk. Appendix C writes that the exponential bound in this empirical mode is 1. Section 5 writes that the zero threshold used in reports trades acceptance sensitivity for the positive-margin guarantee.
Experimental Coverage
- Main experiments: Four backbones each run ten rounds with three training seeds; evaluation covers seven math benchmarks and five general reasoning benchmarks, with sizes in Appendix E; Table 1 reports round 10, the actually inherited checkpoint.
- Inheritance audit: HEIRCOMMIT is applied to native updates from R-Zero, Agent0, and Skill-SP, 64 per host, plus 96 audits of exactly valid Skill-SP tasks (Sections 6.1–6.2, Table 4). The paper does not state the backbone for this set of audits.
- Ablations and interventions: The component ablation in Table 2 and the equal-budget comparisons in Figure 9 are both on Qwen3-4B-Base with three training seeds. Experience interventions use four fixed recipients from rounds 3, 5, 7, 10 and 16 donors, evaluated on the same 1,024 problems before confirmation (Section 6.4).
- Evaluation: Filtering uses 16 problems, confirmation 32, with four answers per state per problem; consensus probe support threshold is 0.7 (Appendix A). The paper states there is no natural language judge (Appendix H.4) and does not report agreement with human judgment. Main results give sample standard deviations over three seeds; hypothesis testing appears only in the bootstrap intervals of Figure 7d.
- Cost accounting: Native baseline 100.0 units, No Recovery 111.9, Full 123.7; the paper states these units are not measured GPU hours or FLOPs. Full takes about 8–9 wall-clock hours per round on two approximately 96 GB GPUs (Appendix G).
Summarize the main content of the paper
Judge the update itself rather than just the task, and let experience stay with the model that produced it after rejection.
This work addresses inheritance decisions in self-evolutionary training: which update to carry into the next round, rather than only whether the current task looks learnable.
- Problem: R-Zero, Agent0, and Skill-SP admit tasks based on pre-training behavior and inherit every resulting update. The authors call the gap left by this the selection-to-update gap. Under independent answers, 52.2% and 42.0% of the two types of native admissions leave the frontier; among 96 exactly valid Skill-SP updates, 60 do not decrease the correct-target NLL.
- Approach: HEIRCOMMIT freezes the successor and inherits it only if it is strictly better than the current root on new probes; on rejection, it restores the full root state. EVOHEIR then separates two things: filtering decides the parameter starting point, and Recovery uses current and archived references to repair that recipient's failures and retain its successes; confirmation occurs after absorption. Eligible experience enters the archive regardless of submission, and rejected states can still serve as donors later. Filtering and confirmation use disjoint problem pools; the authors use a union bound to show that reusing filtering evidence to accept a branch and confirming a fixed proposal with new probes are not the same guarantee. The reported runs use a threshold of 0.
- Inheritance audit: On 64 native updates per host across three hosts, HEIRCOMMIT reduces the proportion of harmful submissions from 76.56%, 57.81%, 46.88% to 13.33%, 16.00%, 15.15% (Table 4). The authors state a reduction of 68–83% relative to random inheritance at the same acceptance rate. Pool gains for R-Zero and Skill-SP flip from negative to positive. The paper does not state the backbone for this set of audits.
- Ten-round results: On four backbones, EVOHEIR's Math Avg. and General Avg. are higher than R-Zero, Agent0, and No Recovery (Table 1). The Math Avg. gains over Agent0 are +6.0%, +7.0%, +6.8%, +1.2%. The abstract additionally writes 5.3%, and the conclusion writes up to 7.0%; Figure 1's four-backbone mean math gain is +5.7 percentage points. EVOHEIR is higher than Agent0 in 45 of 48 benchmark–backbone cells, with the authors stating that the three exceptions are all 30-problem AIME sets.
- Experience: Eight of ten donors weaker than the current root still have experience that improves the recipient. On the same 64 references, Recovery has a net gain of +1.39 percentage points, while Cache SFT is −0.54; at round 7, Recovery itself is −0.20. The authors state that the fact that the model producing the experience was not inherited does not prevent that experience from being used later.
- Authors' conclusion: Self-evolutionary systems should judge the updates they produce, not just the selected problems, and retain useful experience even when the model is not inherited.