EncouRAGe: Incentives in Adaptive Systems

• EncouRAGe is a research paradigm that alters system processes (e.g., gating gradients or input friction) to make desired behaviors more attractive.
• It spans diverse domains such as disentangled representation learning, strategic classification, AI prompt design, and retrieval-augmented generation with measurable performance improvements.
• The framework emphasizes incentive design, inductive bias, and mechanism-level interventions to encourage genuine improvement over superficial adaptations.

EncouRAGe is a recurrent research label used for systems, methods, and frameworks that alter a learning, interaction, or evaluation process so that a target behavior becomes easier, more attractive, or more strongly rewarded. In current arXiv usage, the name has been attached to a weakly supervised variational autoencoder for disentanglement, a strategic-classification method for constructive adaptation, interaction techniques that elicit longer prompts in AI writing, and a local framework for evaluating retrieval-augmented generation (RAG) systems (Vowels et al., 2019, Chen et al., 2020, Joshi et al., 4 Jul 2025, Strich et al., 31 Oct 2025). Rather than denoting a single canonical architecture, the term spans multiple domains while retaining a common emphasis on incentive design, inductive bias, and mechanism-level intervention.

1. Scope and recurring design logic

Across its explicit uses, EncouRAGe refers to interventions that do not merely score outcomes after the fact but instead restructure the process that produces them. In representation learning, this takes the form of gated gradient flow that pushes specific factors into designated latent partitions. In strategic classification, it appears as an objective that discourages manipulation while rewarding genuine improvement. In human–AI interaction, it is realized as prompt-entry friction that makes users contribute more before submitting. In RAG research, it denotes a modular evaluation infrastructure that standardizes retrieval, generation, and measurement (Vowels et al., 2019, Chen et al., 2020, Joshi et al., 4 Jul 2025, Strich et al., 31 Oct 2025).

EncouRAGe instantiation	Domain	Core mechanism
Gated-VAE / EncouRAGe	Disentangled representation learning	Latent partitioning with gated backpropagation
EncouRAGe	Strategic classification	Joint minimization of manipulation risk and improvement risk
EncouRAGe augmented prompt entry	Human–AI writing	Time-cost or motor-cost submission friction for short prompts
EncouRAGe framework	RAG evaluation	Type Manifest, RAG Factory, Inference, Vector Store, Metrics

This suggests that EncouRAGe functions less as a single technical stack than as a design stance: desired behavior is encouraged by modifying the optimization geometry, the subject’s best response, the user interface, or the evaluation protocol itself.

2. Weakly supervised disentanglement in variational autoencoders

In "Gated Variational AutoEncoders: Incorporating Weak Supervision to Encourage Disentanglement" (Vowels et al., 2019), EncouRAGe denotes a weakly supervised training strategy for VAEs. The method starts from the standard ELBO and partitions the latent representation $\mathbf z \in \mathbb R^M$ into $P$ subspaces,

$$\mathbf z=[\mathbf z_1,\mathbf z_2,\dots,\mathbf z_P].$$

All partitions are used during the forward pass, and the decoder reconstructs from the full latent code. The distinctive intervention is on the backward pass: gradients are gated so that, for a given input/target pairing, only the designated partition receives learning signal. Weak supervision enters through pair construction rather than direct label injection. Pairs are chosen so that the two images share a subset $\mathbf s \subset \mathbf v$ of factors, and the corresponding latent partition is trained from that sharing relation.

The method is explicitly designed as a plug-in for existing VAE variants. The paper applies it to $\beta$-VAE, InfoVAE, and DIP-VAE-II without replacing their underlying objectives. Its contribution is therefore not a new divergence or prior, but a training protocol that imposes factor-specific pressure on latent organization. The stated goal is not necessarily full disentanglement within each partition; rather, the partitions themselves become disentangled from one another.

On dSprites, with latent dimensionality $M=8$ and $P=4$ equal partitions, the reported quantitative trend is substantial. Using the Eastwood and Williams metrics with a Lasso regressor, weighted-average disentanglement rises from $0.237$ to $0.609$ for $\beta$-VAE, from $P$0 to $P$1 for InfoVAE, and from $P$2 to $P$3 for DIP-VAE-II. Completeness also improves, for example from $P$4 to $P$5 for $P$6-VAE and from $P$7 to $P$8 for InfoVAE. Informativeness improves through lower NRMSE, with $P$9-VAE dropping from $$\mathbf z=[\mathbf z_1,\mathbf z_2,\dots,\mathbf z_P].$$0 to $$\mathbf z=[\mathbf z_1,\mathbf z_2,\dots,\mathbf z_P].$$1 and InfoVAE from $$\mathbf z=[\mathbf z_1,\mathbf z_2,\dots,\mathbf z_P].$$2 to $$\mathbf z=[\mathbf z_1,\mathbf z_2,\dots,\mathbf z_P].$$3.

The face experiment uses grayscale CelebA with alignment and weak labels derived from OpenFace 2.0. Head-pose labels are built from pitch and yaw, expression labels from FACS/action units, and noisy pairings are formed via K-means into 2,500 head-pose clusters and 4,000 expression clusters. The latent space is split into a 6-dimensional head-pose partition and an 18-dimensional expression partition. Qualitative traversals show that head-pose is separated from expression fairly well, with little head-pose leakage into the expression partition; the paper also notes that only two dimensions are effectively used for head-pose, consistent with pitch and yaw requiring two degrees of freedom.

3. Constructive adaptation in strategic classification

In "Linear Classifiers that Encourage Constructive Adaptation" (Chen et al., 2020), EncouRAGe is a strategic-classification method for settings in which decision subjects respond to a deployed classifier by changing their features. The paper distinguishes immutable features, improvable features, and manipulable features, writing

$$\mathbf z=[\mathbf z_1,\mathbf z_2,\dots,\mathbf z_P].$$4

and models interaction as a two-stage Stackelberg-style game. The model designer first selects a classifier, typically a linear threshold rule $$\mathbf z=[\mathbf z_1,\mathbf z_2,\dots,\mathbf z_P].$$5. Decision subjects then choose an adapted feature vector $$\mathbf z=[\mathbf z_1,\mathbf z_2,\dots,\mathbf z_P].$$6 to maximize

$$\mathbf z=[\mathbf z_1,\mathbf z_2,\dots,\mathbf z_P].$$7

where the cost is a Mahalanobis norm over actionable features.

The central design objective is to make the subject’s best response resemble genuine improvement rather than superficial manipulation. To that end, the method defines an improving best response $$\mathbf z=[\mathbf z_1,\mathbf z_2,\dots,\mathbf z_P].$$8, a manipulating best response $$\mathbf z=[\mathbf z_1,\mathbf z_2,\dots,\mathbf z_P].$$9, and optimizes

$\mathbf s \subset \mathbf v$0

Here $\mathbf s \subset \mathbf v$1 penalizes prediction error after pure manipulation, while $\mathbf s \subset \mathbf v$2 rewards a classifier when the improving best response yields a positive prediction. This differs from conventional strategic classification, which treats all adaptation as adversarial and typically optimizes only against an unconstrained best response.

The theory develops closed-form best responses for linear classifiers under Mahalanobis costs and derives several qualitative consequences. One result states that if a manipulable feature has nonzero weight in the classifier, then almost every subject will change it under strategic adaptation. Another shows that if a subject can flip the decision by improving only improvable features, then the subject can also flip it under unconstrained adaptation, while the converse need not hold. The paper also proves a group-disparity statement: if two groups share the same cost for improving features but differ in the cost of manipulating features, then the group with higher manipulation cost must pay more to flip its label.

The link to true outcome improvement is handled carefully. The improvement term is justified under a covariate shift assumption in which $\mathbf s \subset \mathbf v$3 remains unchanged while only $\mathbf s \subset \mathbf v$4 changes. Under that assumption, shifting the adapted distribution toward regions with positive predictions can align with shifting it toward regions with positive true labels. Empirically, on a synthetic causal benchmark and on credit, adult, german, and spambase datasets, the method maintains competitive deployment accuracy while inducing higher improvement rates and less manipulation than the comparison baselines.

4. Augmented prompt entry and psychological ownership

In "Interaction Techniques that Encourage Longer Prompts Can Improve Psychological Ownership when Writing with AI" (Joshi et al., 4 Jul 2025), EncouRAGe refers to two augmented prompt-entry techniques for chat-style generative AI systems. The motivating claim is that writing longer prompts for an AI assistant to generate a short story increases psychological ownership, defined as the user’s feeling that the writing belongs to them. The intervention is not to require long prompts outright, but to make short prompts more costly to submit.

The first technique is a time-cost design. A gauge-like progress bar beneath the submission button shows progress toward a target of 150 words. If the prompt is shorter than 150 words, normal submission is disabled and the user must press and hold the button until the bar fills; releasing the button resets hold progress to zero. The bar advances every 100 ms, and typing more words also increases the bar’s width, allowing the user to trade waiting time for additional writing. The second technique is a motor-cost design with the same 150-word target, but submission requires repeatedly moving a vertical slider up and down to “pump” the gauge when the prompt is short. Editing the prompt resets the gauge based on current word count, and the submit button remains disabled until the bar turns green.

Experiment 1 was a within-subjects study with 29 valid participants. The interface resembled ChatGPT, and each participant completed baseline, time-cost, and motor-cost conditions. Both augmented techniques increased prompt length relative to baseline: the reported median word counts were 15 for baseline, 35 for time, and 49 for motor. Both also increased text similarity between prompt and generated story, from a baseline median of .28 to .37 for time and .36 for motor. Psychological ownership likewise increased, with medians of 28.25 for baseline, 41 for time, and 47.5 for motor. The two techniques did not differ significantly from each other on word count or ownership. Workload rose as well: mental demand, physical demand, effort, and frustration were all significantly higher than baseline, and motor produced the highest physical demand.

The process analysis shows that the mechanisms altered submission behavior rather than serving as decorative interface elements. In the time condition, participants attempted to press the button about 10 times on average; 41% increased word count, while 55% showed no change between initial and final word count even though almost none of those no-change cases had already reached 150 words. In the motor condition, participants tried to move the slider about 6 times on average; 41% increased word count, and 17% increased it by 50 words or more.

Experiment 2, with 30 new participants, added AI-generated expansion suggestions. After one second of typing inactivity, a blue tooltip appeared in the upper-right corner of the prompt box with a one-sentence suggestion phrased as a question and generated with GPT-3.5 Turbo. The tooltip remained fully visible for 4 seconds and then faded over 3 seconds; it disappeared when typing resumed and was not auto-inserted into the prompt. This augmentation caused prompt length to increase sharply: the median word counts were 19 for baseline, 155 for time, and 150.5 for motor. Text similarity also rose, from .32 in baseline to .41 and .42 in the two augmented conditions. Yet the between-experiment comparison showed that suggestions did not produce significant gains in psychological ownership relative to the corresponding no-suggestion conditions. The paper interprets this as evidence that ownership may depend less on word count alone than on self-generated effort; the correlational analyses support that reading, with mental demand correlating more strongly with ownership than word count in both experiments.

5. Local, modular evaluation of retrieval-augmented generation

In "EncouRAGe: Evaluating RAG Local, Fast, and Reliable" (Strich et al., 31 Oct 2025), EncouRAGe is a Python framework for developing and evaluating RAG systems with an emphasis on local deployment, reproducibility, and modularity. The framework has five components: Type Manifest, RAG Factory, Inference, Vector Store, and Metrics. The Type Manifest standardizes data into objects such as Document, Context, Prompt, and PromptCollection, with Jinja2 templates used for prompt rendering. The RAG Factory provides 10 methods, including Pretrained-Only, Oracle Context, Base RAG, BM25, Hybrid BM25, Reranker, HyDE, Summarization, and SumContext. Inference is built around vLLM and OpenAI-compatible APIs; Vector Store currently supports Chroma and Qdrant; Metrics includes retrieval, generation, and LLM-based evaluators.

The task is formalized as

$\mathbf s \subset \mathbf v$5

where $\mathbf s \subset \mathbf v$6 is the question, $\mathbf s \subset \mathbf v$7 the ground-truth answer, and $\mathbf s \subset \mathbf v$8 the gold context. A retriever–generator pair $\mathbf s \subset \mathbf v$9 produces retrieved contexts $\beta$0 and an answer $\beta$1, after which metrics are computed as $\beta$2. This explicit decomposition lets the framework evaluate retrieval quality, generation quality, and their interaction separately.

The reported evaluation spans four datasets and totals 24,968 QA pairs and 51,068 documents.

Dataset	QA pairs	Documents
HotPotQA	7,395	7,395
FeTaQA	7,324	6,929
FinQA	6,237	2,110
BioSQA	4,012	34,634

The experiments embed contexts with multilingual e5-large instruct, generate with Gemma 3 27B, retrieve the top 10 documents, and run on two NVIDIA H100 GPUs. The framework evaluates retrieval with metrics such as MRR, MAP, nDCG, Recall@k, and HitRate@k, and generation with Exact Match, BLEU, ROUGE, GLEU, Precision, Recall, F1, and Number Match.

The main empirical conclusion is that RAG still underperforms Oracle Context, meaning retrieval remains a bottleneck even when generation is strong. Oracle Context achieves 100 on retrieval metrics wherever applicable, establishing an upper bound. Among practical methods, Hybrid BM25 is reported as the best overall across all four datasets. Its scores include HotPotQA F1 40.8, MRR 70.4, Recall@10 96.9; FeTaQA F1 50.1, MRR 87.6, Recall@10 92.4; FinQA Number Match 51.8, MRR 46.7, Recall@10 82.1; and BioSQA F1 49.9, MAP 43.6, Recall@10 55.7. By contrast, reranking yields only marginal improvements in the main evaluation and introduces a substantial latency cost. In the dedicated reranker analysis, cross-encoder reranking increases execution time by 2–4×, and although increasing the reranker ratio can improve performance by up to 10%, the gains depend strongly on dataset format and reranker alignment.

6. Related encouragement paradigms

A broader literature uses the same basic logic of encouragement without adopting EncouRAGe as the system name. In human–robot interaction, "Exploring Collaborative Game Play with Robots to Encourage Good Hand Hygiene Practises among Children" studies the collaborative game Land of Hands with the custom social robot HakshE (Pasupuleti et al., 2022). The game combines gamification with the Computers as Social Actors paradigm and includes keys, scrolls, an Ask Robot For Help button, points, and a timer. In a Zoom-based study with 32 children aged 6–10 years, learning improved overall from pre- to post-study, including a repeated-measures result of $\beta$3 for demonstrated handwashing steps. However, pro-social nudges did not significantly improve rubric scores relative to the non-nudging condition, while they did significantly increase interaction level, verbal responses, and facial expressions. The result is an important distinction between engagement gains and learning gains.

In computing education, "Using Assignment Incentives to Reduce Student Procrastination and Encourage Code Review Interactions" ties bonus marks to synchronous staff code review rather than to early file upload alone (Wang et al., 2023). The incentive schedule grants 10% if the assignment is completed and code-reviewed on or before Monday, 5% on or before Wednesday, and 0% otherwise. In an upper-level database systems course with 107 students and 10 weekly pair-programming assignments, 234 of 523 submissions were completed in help sessions, amounting to 45%, and 31% of assignments were completed four days before the Friday deadline. Survey responses indicate that 82% strongly agreed and 10% agreed that the bonus marks motivated earlier completion. The paper also reports no meaningful increase in marking time: traditional LMS grading took about 5 minutes per assignment, while live code-review marking typically took 3–5 minutes, sometimes 10+ minutes.

In novice programming support, "Code Perfumes: Reporting Good Code to Encourage Learners" introduces positive feedback as the counterpart to bug and smell detection (Obermüller et al., 2021). The paper defines a code perfume as a code idiom indicative of the correct application of a programming concept or pattern and presents a catalogue of 25 perfumes for Scratch. These include Backdrop Switch, Boolean Expression, Correct Broadcast, Movement In Loop, Parallelisation, and Valid Termination Condition. On 74,907 publicly shared Scratch projects, all 25 perfumes were observed, yielding 4,712,055 total perfume instances; 73,787 projects, or 98%, contained at least one perfume. On 37 Fruit Catching solutions, the number of perfume instances correlated moderately with the number of passed Whisker tests, with Pearson $\beta$4 and $\beta$5. The paper’s broader claim is that automated tools should not only criticize but also identify what learners already do correctly.

A related strand redesigns evaluation itself to encourage a research community toward a desired direction. "New Metrics to Encourage Innovation and Diversity in Information Retrieval Approaches" introduces rareness-weighted variants of precision and average precision that reward relevant documents missed by most systems (Türkmen et al., 2023). On TREC collections, these metrics reorder system rankings, increase discriminative power, and improve stability under topic resampling. "Our Evaluation Metric Needs an Update to Encourage Generalization" proposes WOOD Score, which weights test examples by their semantic dissimilarity to the training set so that more OOD-like examples matter more in evaluation (Mishra et al., 2020). In both cases, encouragement is embedded in the metric: the leaderboard no longer rewards only conventional success, but also innovation or robustness.

7. Debates, limitations, and unresolved questions

The literature also contains explicit cautions against assuming that encouragement mechanisms automatically translate into the desired substantive outcome. "Does Journaling Encourage Healthier Choices? Analyzing Healthy Eating Behaviors of Food Journalers" provides a clear counterexample (Achananuparp et al., 2018). Using MyFitnessPal diaries from 9,896 users over six months, the paper finds that active journalers do not eat as healthily as expected and that their food choices often resemble those of the general populace. Journaling duration is described as only a marginal determinant of healthy eating outcomes, while sociodemographic factors such as gender and region of residence are much more predictive. The paper therefore separates calorie-awareness and self-monitoring from broader dietary improvement.

A second unresolved issue concerns what exactly should be encouraged inside large models. "Encourage or Inhibit Monosemanticity? Revisit Monosemanticity from a Feature Decorrelation Perspective" argues that, during preference alignment, monosemanticity should be encouraged rather than inhibited (Yan et al., 2024). The paper introduces Decorrelated Policy Optimization (DecPO), which adds a feature decorrelation regularizer

$\beta$6

to DPO with $\beta$7. On Llama2-7b-base, Llama2-7b-chat, and Llama3-8b-instruct, the reported effects include increased representation diversity, greater activation sparsity, higher reward margins, and improved preference-alignment performance. The paper explicitly positions this against an earlier inhibition hypothesis and argues that cross-model comparisons are too confounded to settle the question. In this case, even the target of encouragement remains contested.

Taken together, these debates imply that encouragement is not a unitary guarantee but a mechanism whose effect depends on what is being encouraged, how that encouragement is implemented, and which endpoint is measured. In some studies, the intervention mainly improves engagement or interpretability; in others, it changes strategic behavior, prompt elaboration, or system rankings; in still others, the causal route to real-world improvement remains only partial. That variability is central to the contemporary meaning of EncouRAGe: it names a family of interventions that reshape incentives and inductive biases, not a promise that any encouraged behavior will automatically yield the intended downstream benefit.