What concrete numbers show that chains of experience deliver real gains?
The most direct evidence comes from a 2026 study that introduced Chain-of-Experience (CoE), where LLMs accumulate experiential traces through iterative interactions at test time. Across 8 LLMs (including GPT-5, Gemini-2.5 Pro, Claude-4.5 Sonnet) and tasks in math, coding, and knowledge, CoE achieved a 5.6% overall improvement over feedback-free baselines and a 19% reduction in API cost [1]. That means the model not only got better answers but did so more efficiently, which is hard to dismiss as a benchmark trick because it also saves money.
The same study found that combining complementary feedback channels (e.g., model self-feedback plus correctness signals) yielded additional gains, and that CoE delivered higher accuracy per token than existing test-time strategies [1]. This suggests the improvement is not just about memorizing benchmark answers but about using feedback to refine reasoning, which is a more general capability.
Another study, Evo-Memory, benchmarked LLM agents with self-evolving memory across 10 datasets and found that experience reuse (via their ExpRAG baseline) improved performance on sequential task streams [5]. This reinforces that the benefit appears across different setups, not just one specific benchmark.
Does the improvement hold up under weak feedback or when tasks change?
A key concern with benchmark tricks is that they only work under ideal conditions. The CoE study directly tested this: models remained robust under weak or spurious feedback, and different feedback types contributed to distinct improvement aspects, with most gains emerging early in iterations [1]. This means even imperfect feedback still helps, and the model doesn't collapse when feedback is noisy.
Continual learning research also shows that models can learn new tasks without forgetting old ones, which is a sign of genuine improvement rather than overfitting to a single benchmark. For example, attribution-guided continual fine-tuning (2026) achieved better retention of old tasks while maintaining competitive performance on new ones [4]. Similarly, a continual learning model for aspect-based sentiment analysis (2024) obtained state-of-the-art performance across 19 datasets while preserving history-domain abilities [3].
These results converge: improvement from experience is not just about getting better on one benchmark but about maintaining and building on knowledge across tasks, which is what real continual learning should look like.
How do we know the gains aren't just from benchmark leakage?
Benchmark leakage—where training data includes test set answers—can inflate scores artificially. A 2023 study showed that leakage can dramatically boost evaluation results, leading to unreliable assessments [6]. This is a real risk, but the CoE study and others use evaluation protocols that avoid this by testing on unseen tasks or using feedback that is not derived from the benchmark answers [1][5].
For instance, CoE uses environmental signals like public coding test pass rates, which are not part of the benchmark's answer key [1]. Evo-Memory uses streaming task streams where memory evolves after each interaction, so the model is not simply recalling pre-seen answers [5]. These designs reduce the chance that gains are due to memorization.
Additionally, the PandaLM benchmark (2023) provides an automatic evaluation method that focuses on subjective factors like clarity and adherence to instructions, not just objective correctness, and it achieves 93.75% of GPT-3.5's evaluation ability [2]. This suggests that evaluation can be more nuanced, and improvements in such subjective metrics are less likely to be a simple benchmark trick.
About These Sources
This answer is built on 6 studies (1 peer-reviewed, 5 preprints) — published from 2023 to 2026, 4 from 2024 or later, collectively cited 53 times — selected as the most relevant from 10 studies that passed quality screening, drawn from 46 papers retrieved from a database of over 500 million.
Sources used in this answer
Chain-of-Experience for Continual LLM Improvement
Chain-of-Experience (CoE) improved accuracy by 5.6% and cut API cost by 19% across 8 LLMs and multiple tasks, with robustness to weak feedback and gains emerging early in iterations.
PandaLM: An Automatic Evaluation Benchmark for LLM Instruction Tuning Optimization
PandaLM, a judge LLM, achieved 93.75% of GPT-3.5's and 88.28% of GPT-4's evaluation ability in F1-score, enabling fairer and cheaper evaluation without API dependence.
Boosting Large Language Models with Continual Learning for Aspect-based Sentiment Analysis
An LLM-based continual learning model for aspect-based sentiment analysis achieved state-of-the-art performance across 19 datasets while maintaining history-domain abilities.
Attribution-Guided Continual Learning for Large Language Models
Attribution-guided continual fine-tuning outperformed baselines, achieving better retention of old tasks while maintaining competitive performance on new tasks.
Evo-Memory: Benchmarking LLM Agent Test-time Learning with Self-Evolving Memory
Evo-Memory benchmarked self-evolving memory in LLM agents across 10 datasets, showing that experience reuse (ExpRAG) improves performance on sequential task streams.
Don't Make Your LLM an Evaluation Benchmark Cheater
Benchmark leakage can dramatically boost evaluation results, leading to unreliable assessments; guidelines are proposed to avoid it.