Evidence-Decision-Feedback (EDF)

• Evidence-Decision-Feedback (EDF) is a modular, theory-driven framework integrating evidence collection, policy-based decision-making, and adaptive feedback.
• It is applied in intelligent pedagogical agents and variable-length coding to optimize system efficiency and learning outcomes.
• Its closed-loop design enhances interpretability and performance by dynamically aligning actions with real-time evidence and structured decision metrics.

Evidence-Decision-Feedback (EDF) refers to a modular, theory-driven framework in which system actions are dynamically governed by an explicit loop of evidence gathering, decision-making based on structured policies, and adaptive feedback delivery. The EDF paradigm appears in distinct but related contexts across intelligent pedagogical agents and high-performance error-control coding with variable-length feedback. In both domains, it formalizes a closed-cycle process where the “evidence” is continuously extracted from system-user or system-channel interaction, informing real-time decisions and optimized feedback.

1. Modular Structure and Loop Dynamics

EDF organizes system behavior into three interdependent modules:

1. Evidence Module: Ingests raw input data (e.g., learner actions, code traces, received channel symbols), processes relevant features, and infers structured representations. For pedagogical LLM agents, this involves updating mastery indicators, strategy inferences (“TINKERING”, “DEPTH-FIRST ENACTING”), and conceptual gap analysis relative to expert solutions (ZPD alignment) (Cohn et al., 1 Feb 2026). In feedback coding, it involves computing probabilistic reliability metrics (e.g., ROVA posterior probability) based on received sequence statistics and trellis path-metrics (Williamson et al., 2014).

1. Decision Module: Consumes evidence and policy criteria to select the next system action. In educational agents, the module enacts a pedagogical policy (e.g., PROBE_UNDERSTANDING, SUGGEST_ACTION, PUSH_LIMIT) guided by learning-theoretic principles (self-efficacy, constructivism, ZPD). In coding, the decision simplifies to threshold-based stopping: if posterior reliability surpasses $1-\epsilon$, an ACK is sent; otherwise, additional channel symbols are requested.
2. Feedback Module: Renders the abstract decision as concrete system output—dialogic scaffolds in LLM agents (questions, hints, encouragement per social constructivist doctrine), or binary feedback to the transmitter in coding (ACK/NACK). This feedback both closes the system loop and alters subsequent interaction.

This cyclical architecture facilitates adaptive, context-aware guidance, whether for supporting human learning or for optimizing code transmission efficiency.

2. Theoretical Foundations

In AI-driven learning environments, EDF synthesizes evidence-centered design (ECD), stealth assessment, social cognitive theory, and social constructivism. ECD specifies which behaviors serve as valid indicators of underlying skills; ZPD analysis locates the learner’s “immediate next step”; social constructivist feedback is employed to foster dialogic engagement rather than rote answer-giving (Cohn et al., 1 Feb 2026).

In feedback coding, EDF leverages finite-blocklength information theory. The evidence is rooted in posterior probabilities derived from the Reliability-Output Viterbi Algorithm (ROVA) or its tail-biting variant (TB-ROVA), providing rigorous reliability metrics on the decoded message. The decision module implements a probabilistic thresholding rule, aligning channel resource utilization with prespecified error constraints (Williamson et al., 2014).

3. Mathematical and Algorithmic Formulation

The EDF loop is governed by an explicit sequence of update operations:

$$\begin{align*} &\text{(1) Evidence:} \qquad e_t = \text{Evidence}(m_{t-1}, a_t) \ &\text{(2) Decision:} \qquad d_t = \text{Decision}(m_{t-1}, e_t) \ &\text{(3) Feedback:} \qquad f_t = \text{Feedback}(d_t, e_t) \ &\text{(4) Model Update:} \qquad m_t = \text{UpdateModel}(m_{t-1}, e_t) \end{align*}$$

In coding, the evidence metric—posterior probability $P(W = \hat{W}_n | Y^n = y^n)$—is computed via

$$P(W = \hat{W}_n | Y^n = y^n) = \frac{\exp\{\Lambda_{\max}(y^n)\}}{Z(y^n)}$$

with $\Lambda_{\max}(y^n)$ denoting the MAP trellis path-metric and $Z(y^n)$ the sum over code word likelihoods.

The stopping time is

$$\tau = \inf \{ n \geq 1 : P(W = \hat{W}_n | Y^n) \geq 1 - \epsilon \}$$

guaranteeing the undetected error probability constraint (Williamson et al., 2014).

In educational agents, the policy function is

$$d_t = \arg\max_{p \in \text{Policies}} U(p | m_{t-1}, e_t)$$

where $U$ is utility (e.g., maximal conceptual probing early, autonomy-promoting later) (Cohn et al., 1 Feb 2026).

4. System Instantiations

4.1 LLM-Based Adaptive Scaffolding (Copa)

EDF is exemplified in Copa, a multi-agent system embedded in C2STEM. Its sub-agents instantiate the modular EDF workflow:

Sub-Agent/Module	EDF Role	Functionality
StrategyAgent	Evidence	Analyzes code/action logs for strategy
AssessmentAgent	Evidence	Infers learner state (e.g., STRUGGLING)
KnowledgeAgent	Evidence	Performs expert reference/ZPD alignment
DialogueAgent	Decision/Feedback	Picks policy, generates talk moves/feedback

LLM calls are distributed: asynchronous, high-reasoning agents (e.g., GPT-5) for deep inference; synchronous chat agents (e.g., GPT-5-Chat) for genuinely interactive feedback. Scaffold fading logic is tied tightly to mastery quintiles: PROBE_UNDERSTANDING policies dominate for <20% mastery, SUGGEST_ACTION in 20–60%, and PUSH_LIMIT for >60% (Cohn et al., 1 Feb 2026).

4.2 Variable-Length Coding with Feedback

EDF is operationalized in variable-length feedback (VLF) schemes using rate-compatible punctured convolutional codes and TB-ROVA. The system dynamically adjusts codeword length based on posterior reliability, with single-bit feedback (ACK/NACK) eliminating the need for CRC bits (Williamson et al., 2014).

Simulation results in the binary symmetric channel (BSC, $p=0.05$) and AWGN channel (SNR=2 dB) show TBCC codes achieving 76–82% of theoretical capacity with average latencies of tens to hundreds of symbols—beating the random-coding lower bound for blocklengths <100 symbols. Direct use of ROVA reliabilities as evidence reduces decoding latency relative to CRC-based schemes, with the undetected error bounded by $P(\text{error}) \leq \epsilon$.

5. Empirical Evaluation

5.1 LLM Agent Classroom Study

A six-week C2STEM study involving 33 high-school dyads and open-ended modeling tasks measured policy adaptivity, understanding-mastery alignment, student reliance, and interpretability (Cohn et al., 1 Feb 2026):

• Dialogue policy frequency exhibited strong alignment with mastery quintiles (e.g., PROBE_UNDERSTANDING, ρ=−0.34, p=0.034; SUGGEST_ACTION, ρ=+0.33, p=0.039; PUSH_LIMIT, ρ=+0.42, p=0.007).
• Higher mastery deciles predicted better verbal explanation success (ρ=+0.40, p=0.014).
• Proportion of agent interactions declined as mastery increased (ρ=−0.26, p<0.001).
• Quantitative Trace Analysis demonstrated significant gains in grounding, alignment, and faithfulness compared to baseline dialogue.

Students rated appropriate questioning positively (means ≈3.8/5) but expressed preference for more direct answers, revealing tensions between pedagogical scaffolding and learner expectations.

5.2 Variable-Length Feedback Coding

Simulations compared EDF-based ROVA reliability stopping with CRC-based error detection:

Code/Channel	Latency λ	Throughput $R_t$	% Capacity
TBCC, 64-state (BSC)	44.1	0.543	76.2%
TBCC, 1024-state (AWGN, m=5)	121.0	0.529	82.4%

At short blocklengths (λ<75 bits), CRC overhead imposes severe rate loss; CRCs do not robustly achieve $P_{UE} \leq \epsilon$ at all blocklengths, whereas EDF with ROVA does.

6. Interpretability and Design Implications

EDF’s chain-of-thought (CoT) traces in every module provide high transparency, supporting interpretability for learners and educators. In the LLM context, QTA metrics—keyword recall, SBERT-based similarity—demonstrate statistically significant improvements over baseline groundedness and alignment (Cohn et al., 1 Feb 2026). In coding, the explicit evidence-based stopping rule simplifies reliability auditing and error analysis relative to CRCs.

EDF supports continuous, theory-driven scaffold fading, dynamically shifting the cognitive support locus as learners approach independence (or as a code message approaches decodability). The architecture robustly avoids learner overreliance by decreasing intervention as mastery or reliability increases.

7. Domain-Specific Considerations and Trade-Offs

Key design trade-offs emerge in both application families:

• Pedagogical Agents: The granularity of evidence extraction and clarity of policy alignment are critical; LLM resource allocation (high-reasoning vs. chat-optimized agents) impacts latency and fidelity; scaffold-fading calibration affects autonomy.
• Feedback Coding: Constraint length (state size) versus decoder complexity, symbol-by-symbol versus packetized decoding, and optimal selection of decoding opportunities are pivotal (Williamson et al., 2014). ROVA/TB-ROVA-based EDF eliminates CRC overhead and minimizes average blocklength for a fixed undetected error.

A plausible implication is that the EDF framework, by systematizing the closed loop from data-driven evidence to context-aware feedback, supports optimal adaptivity and interpretability across domains where responsive interaction is central.

---

References

• "Evidence-Decision-Feedback: Theory-Driven Adaptive Scaffolding for LLM Agents" (Cohn et al., 1 Feb 2026)
• "Variable-length Convolutional Coding for Short Blocklengths with Decision Feedback" (Williamson et al., 2014)