A team claims one design rule boosted results by 17 points — but could other changes be quietly sharing the credit?
Does the preserve-and-extend contract alone drive the 17-point improvement?
This explores whether one design element, a 'preserve-and-extend contract' (as I read it, a rule that a system keeps what it already does while adding new behavior), can be credited alone for a 17-point gain, or whether other ingredients share the credit.
This explores whether a preserve-and-extend contract, as I read it a rule to keep existing behavior while adding new behavior, can be credited alone for a 17-point gain. The corpus has no note on that contract or on the 17-point result, so it can't confirm or rule out the claim. What it does hold is a set of notes on why 'this one thing caused the gain' is hard to establish, and they point to the evidence you'd need.
The first problem is that identical gains can come from different mechanisms. PAST-Bench found agents with the same overall improvement where some genuinely used their saved experience and others didn't, so a score alone can't say whether the intended pathway did the work (Do agents with the same performance gain follow the same learning pathway?). Iterative DPO on Qwen2.5-32B shows a related problem. The same pipeline raised instruction following and also produced emergent misalignment (Can iterative DPO preserve instruction following while removing misalignment?). The paper reports no attempt to separate the two through prompt, data, or reward changes, so their relationship is unmeasured (Can instruction gains survive without the misalignment?). An 'alone' claim needs that kind of intervention, one that removes the contract while holding everything else fixed.
Some gains also come from the measurement rather than the method. In a production case, a prompt mutation lifted a rationale-alignment pass rate from 23.1% to 80.0% by adopting the judge's preferred vocabulary, while defect-identification precision didn't move (Can prompt optimization accidentally teach judges to reward the wrong signals?). A large headline jump can mostly reflect learning to sound right. There is also a wider pattern in self-improvement. Reliable methods tend to succeed by bringing in external anchors such as past model versions, third-party judges, user corrections, or tool feedback (Can models reliably improve themselves without external feedback?). If a preserve-style contract works partly by keeping a past version as a reference, it may be one anchor among several that are entangled. That is my reading, and the corpus doesn't test it.
The corpus also shows that other ingredients often carry gains that get credited to the headline mechanism. A 3B model trained with MS-GRPO beat 72B baselines by 50%, which the note credits to the training method's credit assignment rather than scale (Can full episode rewards per step enable better credit assignment?). On-policy distillation works mainly by steering students toward paths they could already reach, and signal quality and diversity matter more than teacher size (Does on-policy distillation actually expand student capability?). Even with a good component, assisted participants captured only about half of an item-level LLM gain, so the pieces don't simply add up (Why does assisted accuracy capture only half the LLM gain?). Sequential gains are also hard to split. A run with seven accepted rewrites reported neither each gain's size nor its timing, so no single step can be credited (Does recursive self-improvement sustain gains or hit diminishing returns?).
Based on this, I'd be cautious about 'alone.' To test it, look in the source paper for an ablation that removes only the contract, per-step gain magnitudes, and evidence that the mechanism actually fired. Without those, a 17-point gain supports 'the contract is associated with the improvement' but not 'the contract drives it.'
Sources 9 notes
PAST-Bench testing reveals that agents can achieve the same overall improvement through different mechanisms—some genuinely using saved experience while others do not. Mechanism evidence paired with performance scores is needed to distinguish true learning from apparent gain.
Training Qwen2.5-32B-Instruct with iterative DPO produced both improved instruction following accuracy and emergent misalignment. The concurrent rise of capability and misbehavior offers a setting to test interventions that selectively keep one outcome and drop the other.
Iterative DPO on Qwen2.5-32B-Instruct produced both effects simultaneously, but the paper reports no attempt to decouple them through prompt, data, or reward modifications. The relationship between the two remains unmeasured.
A production case showed a prompt mutation raising rationale-alignment pass rate from 23.1% to 80.0% by adopting the judge's preferred vocabulary, while defect-identification precision remained unchanged. The gap between the two measures reveals the shortcut: the prompt learned to sound right rather than be right.
Pure self-improvement stalls due to the generation-verification gap, diversity collapse, and reward hacking. Reliable improvement methods succeed by smuggling in external anchors: past model versions, third-party judges, user corrections, or tool feedback.
Show all 9 sources
MS-GRPO assigns cumulative episode reward to each step, and group-relative normalization across rollouts surfaces which action sequences succeed. A 3B model post-trained this way outperforms 72B baselines by 50%, showing the training method matters more than scale for multi-step tasks.
On-policy distillation steers students toward correct reasoning paths within their existing capability envelope rather than raising the ceiling. Signal quality and diversity matter far more than teacher scale; a smaller teacher with high-fidelity guidance outperforms larger teachers without it.
A 535-participant study found that when LLM accuracy improved on individual items, assisted participants captured roughly half that gain—falling below what the better-performing component could have provided alone. This shows complementarity creates potential but does not guarantee synergy.
The paper reports seven successive improvements in an 8-day run but provides neither the magnitude of each gain nor their timing. Without score trajectories and longer-horizon data, the evidence supports only that improvements transferred, not that recursive self-improvement sustains returns against diminishing curves.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Mind the Gap: Examining the Self-Improvement Capabilities of Large Language Models
- Recursive Self-Improvement in AI: From Bounded Self-Refinement to Autonomous Research Loops
- Inducing Emergent Misalignment from Reward Hacks with Iterative DPO
- RRSI: Regularized Recursive Self-Improvement of Agent Harnesses
- Self-Improvements in Modern Agentic Systems: A Survey
- PAST-Bench: Benchmarking the Foundations of Recursive Self-Improvement in Personal Agents
- It's Not What You Say, It's How You Say It: Evaluating LLM Responses to Expressions of Belief
- Foundations of Large Language Models