Experiential Reflective Learning (ERL)

• Experiential Reflective Learning (ERL) is an iterative cycle that couples concrete experience with deliberate reflection, abstraction, and action.
• Key methodologies include guided prompts, retrieval-based frameworks, and adaptations of Kolb’s cycle to transform insights into actionable strategies.
• ERL is applied across educational settings, AI-agent self-improvement, and robotics, demonstrating measurable gains such as improved task success and enhanced learning outcomes.

Experiential Reflective Learning (ERL) denotes an iterative learning regime in which concrete experience is coupled to deliberate reflection, abstraction, and subsequent action. In educational research, ERL combines hands-on, authentic work with structured reflective practice; in recent AI-agent research, it denotes self-improvement mechanisms that transform trajectories into heuristics, insights, or episodic memories and then reuse them through retrieval or test-time adaptation (Saha et al., 21 Apr 2026, Allard et al., 25 Mar 2026). Across these literatures, ERL is consistently associated with cyclical learning, explicit reflective prompts, and a transfer objective: learners or agents are not only expected to complete a task, but to articulate or distill what should guide future practice.

1. Conceptual foundations

ERL is grounded in several overlapping theories of learning and reflection. In the remote-engineering formulation, ERL builds directly on Kolb’s Experiential Learning Theory, with the cycle stated as

$$\mathrm{CE} \rightarrow \mathrm{RO} \rightarrow \mathrm{AC} \rightarrow \mathrm{AE} \rightarrow \mathrm{CE},$$

or equivalently

$$\mathrm{RO}=\rho(\mathrm{CE}), \quad \mathrm{AC}=\alpha(\mathrm{RO}), \quad \mathrm{AE}=\beta(\mathrm{AC}), \quad \mathrm{CE}'=\gamma(\mathrm{AE}),$$

where Concrete Experience, Reflective Observation, Abstract Conceptualization, and Active Experimentation recursively feed one another (Kularatne et al., 2021). Guided-reflection work in physics places this cycle alongside Dewey’s “active, persistent, and careful consideration,” Schön’s distinction between reflection-in-action and reflection-on-action, Boud, Keogh and Walker’s three-stage model, and Zimmerman’s self-regulated learning (Dounas-Frazer et al., 2015).

In work-based computer science learning, ERL is defined as a combination of authentic work and structured reflection, instantiated as an open-source micro-internship plus a guided blogging exercise. That study explicitly draws on Turns et al.’s model of reflection as an “intentional and dialectical thinking process” in which learners revisit concrete features of their experience through technical, social, and identity lenses to create meaning and guide future practice (Saha et al., 21 Apr 2026). In the Holistic Cognitive Development framework, ERL is formulated as the inner cycle of an Autonomy, Ownership, and Scaffolding surround, and is written as

$$\mathrm{ERL}=\sum_{i=1}^{n} F\bigl(\mathrm{Think}_i,\mathrm{Create}_i,\mathrm{Critique}_i,\mathrm{Reflect}_i\bigr),$$

with optional modulation by supervision dimensions $A_i$, $O_i$, and $S_i$ (Anand, 10 Nov 2025).

The AI-agent literature preserves the same experiential-reflective logic while changing its substrate. Rather than essays or blogs, reflection is encoded as heuristics, insights, strategy entries, or cases. In one formulation, after each trajectory $\tau$ with task $x$ and binary outcome $r$, ERL applies

$h=\mathrm{Reflect}(x,\tau,r)$

to produce a heuristic for a persistent pool $$\mathrm{RO}=\rho(\mathrm{CE}), \quad \mathrm{AC}=\alpha(\mathrm{RO}), \quad \mathrm{AE}=\beta(\mathrm{AC}), \quad \mathrm{CE}'=\gamma(\mathrm{AE}),$$0, later retrieving the top-$$\mathrm{RO}=\rho(\mathrm{CE}), \quad \mathrm{AC}=\alpha(\mathrm{RO}), \quad \mathrm{AE}=\beta(\mathrm{AC}), \quad \mathrm{CE}'=\gamma(\mathrm{AE}),$$1 heuristics for a new task (Allard et al., 25 Mar 2026). In another theoretical account, “writing” to episodic memory corresponds to policy evaluation and “reading” from memory corresponds to policy improvement in a Stateful Reflective Decision Process (Wang, 27 Dec 2025). Taken together, these formulations indicate that ERL is less a single method than a family of cyclic architectures linking experience, reflection, and future action.

2. Educational implementations

Educational ERL has been operationalized in work-based CS programs, laboratory science, remote engineering, interaction design, and creative computing. The common pattern is phase-aligned reflection: prompts are attached to identifiable points in the work process rather than deferred to an entirely post hoc narrative.

Setting	Reflective structure	Reported finding
Work-based CS micro-internship	Five-section LinkedIn blog post over a 4-week CodeDay Labs program	Four themes emerged, and students demonstrated deep reflection across all four Knowledge Gain constructs (Saha et al., 21 Apr 2026)
Advanced laboratory physics course	300–500-word guided reflection essay after a six-week team project	The majority reported satisfaction and achievement, functional team dynamics, learning outcomes unique to the experience, modeling, and future improvements (Cai et al., 2018)
Modeling-based physics course	Weekly Guided Reflection Form with instructor feedback for nine weeks	94% of reflections contained at least one of the coded statement-types (Dounas-Frazer et al., 2015)
Remote engineering experiment	Pre-lab, simulation, PowerPoint, remote laboratory activity, final report	The majority liked the approach, and 80% judged it an ideal COVID-era alternative (Kularatne et al., 2021)
Interaction design course	Nine AI exercises in a half-day autonomous workshop	Metaphors and enactments made training and learning, privacy and consent, autonomy and agency more tangible (Murray-Rust et al., 2023)
Creative-computing courses	Iterative Think/Create/Critique/Reflect cycles with AI-augmented feedback	Mean reflection score rose from $$\mathrm{RO}=\rho(\mathrm{CE}), \quad \mathrm{AC}=\alpha(\mathrm{RO}), \quad \mathrm{AE}=\beta(\mathrm{AC}), \quad \mathrm{CE}'=\gamma(\mathrm{AE}),$$2 to $$\mathrm{RO}=\rho(\mathrm{CE}), \quad \mathrm{AC}=\alpha(\mathrm{RO}), \quad \mathrm{AE}=\beta(\mathrm{AC}), \quad \mathrm{CE}'=\gamma(\mathrm{AE}),$$3 on a 1–5 rubric (Anand, 10 Nov 2025)

The work-based CS study is unusually explicit about how reflection is integrated with professional identity formation. Twenty-five juniors and seniors in BS-CS programs at state universities and community colleges in California and Washington participated in a single Fall 2024 cohort within a 4-week CodeDay Labs program. Each student was matched to one “simple” open-source issue in a project with an OSI-approved license and real user base, received one week of onboarding, attended weekly team meetings with an industry mentor, posted twice-weekly asynchronous stand-ups, and drafted one section of a blog post each week until publishing a LinkedIn post in Week 4 (Saha et al., 21 Apr 2026).

Other educational implementations vary the reflective object but preserve the cycle. The advanced laboratory course used a final guided essay after a six-week open-ended experiment project, with prompts about achievement, team process, learning, modeling, and future changes (Cai et al., 2018). The Guided Reflection Form in physics structured weekly reflection around a specific experience, a goal for improvement, and a concrete plan, with timely instructor feedback closing the loop between reflection and revised action (Dounas-Frazer et al., 2015). In the interaction-design course, ERL was embodied through nine exercises such as Uncertain Interactions, Be the ML, Poor Datasets, Thing Ethnography of AI Systems, and Meaningful Human Control, each mapped to Kolb’s phases and to reflection themes such as agency, privacy, consent, and responsibility (Murray-Rust et al., 2023).

3. Scaffolding, prompts, and assessment

A defining feature of ERL is scaffolded reflection. In the CS micro-internship study, the five-section blog-post outline mapped directly onto phases of reflection: project mission, issue, codebase overview, challenges, and solution. The prompts asked students to situate work in a user-focused narrative, analyze the issue’s scope and code locations, draw a system diagram, enumerate debugging and help-seeking attempts, and explain the final solution with testing and a link to the pull request (Saha et al., 21 Apr 2026). The authors recommended embedding a five-section blog or portfolio template aligned to project phases, requiring weekly or biweekly draft submissions, pairing reflection with industry mentorship and stand-up reflections, and adapting the template to “day-to-day” engineering realities such as code reviews, tooling, and time allocation.

The Guided Reflection Form is a more minimal scaffold but makes the reflective kernel explicit. Students chose a skill focus, described a specific experience to improve, selected strategies used, articulated an aspect to improve with at least one concrete future strategy, and recorded resources used. Analytical coding reduced the reflections to three statement-types aligned with three ERL questions: narrative statements $$\mathrm{RO}=\rho(\mathrm{CE}), \quad \mathrm{AC}=\alpha(\mathrm{RO}), \quad \mathrm{AE}=\beta(\mathrm{AC}), \quad \mathrm{CE}'=\gamma(\mathrm{AE}),$$4, goal statements $$\mathrm{RO}=\rho(\mathrm{CE}), \quad \mathrm{AC}=\alpha(\mathrm{RO}), \quad \mathrm{AE}=\beta(\mathrm{AC}), \quad \mathrm{CE}'=\gamma(\mathrm{AE}),$$5, and action statements $$\mathrm{RO}=\rho(\mathrm{CE}), \quad \mathrm{AC}=\alpha(\mathrm{RO}), \quad \mathrm{AE}=\beta(\mathrm{AC}), \quad \mathrm{CE}'=\gamma(\mathrm{AE}),$$6, with

$$\mathrm{RO}=\rho(\mathrm{CE}), \quad \mathrm{AC}=\alpha(\mathrm{RO}), \quad \mathrm{AE}=\beta(\mathrm{AC}), \quad \mathrm{CE}'=\gamma(\mathrm{AE}),$$7

for $$\mathrm{RO}=\rho(\mathrm{CE}), \quad \mathrm{AC}=\alpha(\mathrm{RO}), \quad \mathrm{AE}=\beta(\mathrm{AC}), \quad \mathrm{CE}'=\gamma(\mathrm{AE}),$$8 reflections. Inter-rater reliability on 71 double-coded reflections was assessed via Cohen’s $$\mathrm{RO}=\rho(\mathrm{CE}), \quad \mathrm{AC}=\alpha(\mathrm{RO}), \quad \mathrm{AE}=\beta(\mathrm{AC}), \quad \mathrm{CE}'=\gamma(\mathrm{AE}),$$9: $$\mathrm{ERL}=\sum_{i=1}^{n} F\bigl(\mathrm{Think}_i,\mathrm{Create}_i,\mathrm{Critique}_i,\mathrm{Reflect}_i\bigr),$$0, $$\mathrm{ERL}=\sum_{i=1}^{n} F\bigl(\mathrm{Think}_i,\mathrm{Create}_i,\mathrm{Critique}_i,\mathrm{Reflect}_i\bigr),$$1, and $$\mathrm{ERL}=\sum_{i=1}^{n} F\bigl(\mathrm{Think}_i,\mathrm{Create}_i,\mathrm{Critique}_i,\mathrm{Reflect}_i\bigr),$$2 (Dounas-Frazer et al., 2015).

Assessment instruments in ERL range from qualitative coding to validated psychometric constructs. The CS micro-internship study used Mejia and Turns’s Knowledge-Gain instrument with four constructs—Engineering Self, Course Understandings, Areas for Growth, and Social Impact—measured by 16 Likert items normalized to 1–5. Internal consistency was reported as $$\mathrm{ERL}=\sum_{i=1}^{n} F\bigl(\mathrm{Think}_i,\mathrm{Create}_i,\mathrm{Critique}_i,\mathrm{Reflect}_i\bigr),$$3, $$\mathrm{ERL}=\sum_{i=1}^{n} F\bigl(\mathrm{Think}_i,\mathrm{Create}_i,\mathrm{Critique}_i,\mathrm{Reflect}_i\bigr),$$4, $$\mathrm{ERL}=\sum_{i=1}^{n} F\bigl(\mathrm{Think}_i,\mathrm{Create}_i,\mathrm{Critique}_i,\mathrm{Reflect}_i\bigr),$$5, and $$\mathrm{ERL}=\sum_{i=1}^{n} F\bigl(\mathrm{Think}_i,\mathrm{Create}_i,\mathrm{Critique}_i,\mathrm{Reflect}_i\bigr),$$6 respectively, with Cronbach’s alpha given by

$$\mathrm{ERL}=\sum_{i=1}^{n} F\bigl(\mathrm{Think}_i,\mathrm{Create}_i,\mathrm{Critique}_i,\mathrm{Reflect}_i\bigr),$$7

where $$\mathrm{ERL}=\sum_{i=1}^{n} F\bigl(\mathrm{Think}_i,\mathrm{Create}_i,\mathrm{Critique}_i,\mathrm{Reflect}_i\bigr),$$8 is the number of items, $$\mathrm{ERL}=\sum_{i=1}^{n} F\bigl(\mathrm{Think}_i,\mathrm{Create}_i,\mathrm{Critique}_i,\mathrm{Reflect}_i\bigr),$$9 is the variance of item $A_i$0, and $A_i$1 is the total-score variance (Saha et al., 21 Apr 2026). Braun and Clarke’s coding-reliability thematic analysis yielded $A_i$2 and four themes: identification of problem-solving techniques, reward for perseverance / growth mindset, collaborative development challenges and benefits, and benefits of contribution to users and society.

The Holistic Cognitive Development framework extends ERL assessment into rubric-driven AI support. Reflection is scored as

$A_i$3

with $A_i$4. AI systems such as iReflect, ReflexAI, and a Knowledge Graph–Enhanced LLM then operationalize feedback at scale: ReflexAI averages repeated scoring at $A_i$5 using $A_i$6, while the knowledge-graph component augments prompts with relevant design patterns and pitfalls (Anand, 10 Nov 2025). This suggests that ERL scaffolding can range from light prompt structures to multi-component feedback pipelines, provided that the reflective output still links past action to next action.

4. ERL in LLM and memory-augmented agents

In LLM-agent research, ERL is a non-parametric or lightly adaptive answer to a recurring problem: agents otherwise treat each task “from scratch.” ExpeL organizes this process as four components—Experience Generation, Memory Storage, Reflection / Knowledge Extraction, and Retrieval. Training collects both successful and failed trajectories through ReAct+Reflexion retries, after which an instruction-tuned model distills a compact set of natural-language insights $A_i$7 from same-task success–failure pairs and cross-task success chunks. Each insight can be ADDed, EDITed, UPVOTEd, or DOWNVOTEd; at test time, all surviving insights are prepended to the prompt and the top-$A_i$8 similar successful trajectories are appended as few-shot exemplars, with no further reflection or retries at inference (Zhao et al., 2023).

The 2026 ERL framework for self-improving LLM agents makes the reflective object more structured. A task $A_i$9 produces a trajectory

$O_i$0

with binary outcome $O_i$1; reflection yields a heuristic $O_i$2 whose fields include a concise analysis and a guideline decomposed into a trigger (“When ...”) and action (“I must ...”). A persistent pool $O_i$3 stores all heuristics, and retrieval selects

$O_i$4

using either cosine similarity in an embedding space or an LLM-based ranker. On Gaia2, this ERL formulation improved overall success from 48.3% for ReAct to 56.1% for ERL with LLM-based retrieval and $O_i$5, and pass$O_i$6 gains were especially large, indicating higher reliability (Allard et al., 25 Mar 2026).

Several adjacent frameworks elaborate the same experiential-reflective principle with different granularity. R$O_i$7-Mem uses a Rubric-guided Evaluator to score individual search steps across rubric dimensions, labels steps as good or bad via thresholds, and then has a self-Reflection Learner distill abstract Planning or Reflection experiences for offline banks reused during online deep search. It reported up to +22.6% relative F1, while reducing token consumption by 12.9% and search iterations by 20.2% (Wang et al., 13 May 2026). Memento-II gives a theoretical account in which retrieval from episodic memory and writing to episodic memory induce an equivalent reflected MDP over augmented states $O_i$8, enabling entropy-regularized policy iteration without gradient updates to the LLM parameters (Wang, 27 Dec 2025).

A recurring result across these systems is that reflective abstraction outperforms raw replay. ExpeL found that retrieval-only and insights-only both help, but the full combination is best; the Gaia2 ERL study showed that raw-trajectory few-shot prompting underperformed heuristic retrieval; R$O_i$9-Mem argued that step-granular reflective abstraction captures process-level nuance missing in coarse trajectory memories (Zhao et al., 2023, Allard et al., 25 Mar 2026, Wang et al., 13 May 2026). This suggests that, in agentic ERL, the central technical choice is not merely whether to remember, but how to compress experience into transferable guidance.

5. Embodied and robotic variants

Embodied ERL extends reflection from textual trajectories to physically consequential action. Reflective Test-Time Planning distinguishes reflection-in-action, reflection-on-action, and retrospective reflection. The agent samples candidate actions $S_i$0 from a policy $S_i$1 at high temperature, an internal reflection model $S_i$2 scores them via $S_i$3, and the action with maximal score is executed. After observing the outcome, an external reflection model $S_i$4 produces $S_i$5; later, retrospective reflection revisits earlier steps for long-horizon credit assignment, and test-time training updates both the internal reflection model and the action policy (Hong et al., 24 Feb 2026). On the Long-Horizon Household benchmark, full Reflective Test-Time Planning reached 33.7% average success versus a best baseline of about 11.2%; on MuJoCo Cupboard Fitting it achieved 60.2% fit and 25.3% correct, with LoRA adapters attaining comparable performance to full weight updates at 95% fewer updated parameters.

ExpTeach instantiates ERL for a real robot around a closed loop of planning, execution, verification, reflection, and memory. A pretrained VLM planner $S_i$6 proposes actions conditioned on user instruction, current RGB-D observation, short-term memory, and retrieved long-term memories; a VLM success detector $S_i$7 returns a feedback tuple

$S_i$8

short-term memory stores action–feedback pairs, and task completion triggers summarization into a long-term memory entry retrievable by cosine similarity (Lan et al., 22 Jul 2025). Reflection improved average success on four challenging robotic tasks from approximately 36% to approximately 84%, and grounding with long-term memory increased single-trial success across 12 real-world scenarios from approximately 22% to approximately 80%. A RAG ablation further showed 89% success for top-$S_i$9 cosine retrieval, compared with 67% when the entire long-term memory was placed in the prompt and 27% for random $\tau$0 memories.

ELITE frames embodied ERL as the tandem of Self-Reflective Knowledge Construction and Intent-Aware Transfer in a POMDP

$\tau$1

After each episode, a Reflective Experience Distiller produces success patterns or failure causes, a Context Consolidator updates an evolving strategy pool through Add, Revise, and Remove operations, and later a coarse planner embeds the current intent for top-$\tau$2 retrieval of strategic knowledge (Wei et al., 25 Mar 2026). On EB-ALFRED and EB-Habitat, ELITE improved online performance over base VLMs by 9 percentage points and 5 percentage points respectively; in the supervised setting on EB-ALFRED, average success increased from 49.8% for the base Qwen2.5-VL to 70.8% for ELITE.

These embodied variants clarify an important distinction within ERL. Some systems remain parameter-free and learn through memory and prompt conditioning; others, such as Reflective Test-Time Planning, combine reflective memory with test-time training. The shared feature is not a single optimization regime, but the reuse of experience-specific critique to alter future behavior in the same deployment context.

6. Limits, misconceptions, and research directions

ERL studies repeatedly note that reflection alone is difficult to isolate causally. The CS micro-internship study was descriptive, used a single 4-week cohort with modest $\tau$3, and included mentorship and stand-ups in addition to blogging; the authors explicitly state that future work should compare outcomes against a control group, use randomized designs, compare to other reflective methods such as journals or peer discussion, and measure downstream outcomes such as job-search success and interview performance (Saha et al., 21 Apr 2026). The redesigned laboratory course similarly relied on consensus coding without formal inter-rater statistics, and the authors used reflections to justify curricular modifications such as more scaffolding for modeling and more project-management support (Cai et al., 2018).

Educational ERL also presents timing and framing tensions. In the interaction-design course, many students wanted the exercises earlier for ideation, whereas coaches valued them as “zoom-out” reflection in the final stage; autonomous execution fostered ownership, but facilitation improved focus; and some students drifted toward anthropomorphic “AI as friend” metaphors, motivating calls for more technically grounded exercises (Murray-Rust et al., 2023). In remote engineering, technical hiccups such as oscilloscope downtime and login delays were reported, and the authors recommended modularization, controlled access, multiple data-capture modes, and contingency planning (Kularatne et al., 2021).

Agentic ERL has a different failure profile. The Gaia2 ERL framework adds an LLM retrieval call and larger prompts, incurring about 40% higher API cost; under noisy self-assessed outcomes, success drops by about 4.8%; and growing heuristic pools raise retrieval-quality and conflict-resolution problems (Allard et al., 25 Mar 2026). R$\tau$4-Mem notes dependence on evaluator quality and on distributional similarity between offline trajectories and online tasks (Wang et al., 13 May 2026). ELITE notes that incorrect coarse planning can mislead retrieval and that a flat global strategy pool may become unwieldy in large domains (Wei et al., 25 Mar 2026).

A common misconception is that ERL is equivalent to unrestricted memory accumulation. Multiple agent papers instead show that selective retrieval is decisive: ExpeL uses top-$\tau$5 similar successful trajectories plus distilled insights, the Gaia2 ERL framework finds selective retrieval essential, ExpTeach outperforms both full-memory prompting and random memory selection with top-$\tau$6 cosine RAG, and R$\tau$7-Mem conditions retrieval on abstract situations in separate planning and reflection banks (Zhao et al., 2023, Allard et al., 25 Mar 2026, Lan et al., 22 Jul 2025, Wang et al., 13 May 2026). Another misconception is that ERL is reducible to reflective writing. The literature includes essays, forms, blogs, critique logs, heuristics, JSON-style records, strategy pools, and episodic-memory read–write processes. This suggests that the invariant of ERL is not medium but function: experience must be revisited in a form that can guide later action, whether the learner is a student articulating project meaning or an agent retrieving a failure-avoidance guideline.