Skip to content

Prior Art and Defensible Novelty Boundary

This document maps EQ-Layer against pre-LLM dialogue-system research. Its goal is not to manufacture a "first" claim. It records which ideas clearly existed before modern large language models, which combinations are close to EQ-Layer, and where a narrower contribution may still be defensible.

Bottom line

EQ-Layer is not the first system to:

  • infer a user's goal or dialogue act;
  • maintain dialogue state across turns;
  • detect affect, frustration, uncertainty, or satisfaction;
  • adapt dialogue strategy to a user model;
  • select a dialogue action before surface realization;
  • use confidence/uncertainty in action selection;
  • repair misunderstandings or react to corrections;
  • adapt content, politeness, or social behavior to the user;
  • optimize a dialogue policy;
  • evaluate alternative dialogue strategies with human users.

Those ideas are established pre-LLM research topics.

The potentially defensible contribution is narrower:

A model-agnostic control layer for frozen/general-purpose generative language models that explicitly factor intent, affect, conversational subtext, stance evidence, and uncertainty into inspectable state; select a discrete interaction policy before token generation; steer the unchanged base model; preserve provenance for direct/derived/structural/verified signals; and test the layer using same-base-model blinded response-level comparisons.

This wording is intentionally a novelty hypothesis, not a first-in-history claim. It must still be checked against contemporary LLM middleware and steering literature.

High-value pre-LLM prior art

1. TRAINS / grounding / repair (1994-1996)

Key work - Traum & Allen (1994), formal theory of repair. - Allen et al. (1995), TRAINS conversational planning agent. - Traum & Dillenbourg (1996), miscommunication and repair. - Traum & Heeman (1996), grounding in spoken dialogue.

What already existed - explicit discourse context; - shared plans and discourse obligations; - mixed-initiative action selection; - models of grounding, misunderstanding, and repair; - repair as a consequence of misaligned agent mental states.

EQ-Layer overlap - correction/repair policies; - conversational history; - stance/grounding-related state; - separating a dialogue move from its surface wording.

What to borrow - represent repair as a state transition, not merely a keyword class; - add explicit misalignment_probability, repair_target, and repair_distance fields; - distinguish self-correction from user-corrects-system repair.

Reference - https://trips.ihmc.us/trips/TRAINS.html - David Traum publication archive: https://people.ict.usc.edu/~traum/Papers/

2. How May I Help You? / open-ended intent routing (1997)

Key work - Gorin, Riccardi, Wright et al., How May I Help You?, Speech Communication, 1997.

What already existed - open-ended user utterances; - automatic semantic action / call-type classification; - learning from large sets of real user interactions rather than requiring users to speak in a fixed grammar.

EQ-Layer overlap - learned intent classification; - routing a free-form user request into a compact control state.

What to borrow - evaluate intent on naturally occurring user language; - replace synthetic intent data with a genuinely held-out human corpus where possible.

DOI - 10.1016/S0167-6393(97)00040-X

3. PARADISE dialogue evaluation (1997-1998)

Key work - Walker, Litman, Kamm & Abella (1997), PARADISE.

What already existed - systematic comparison of dialogue strategies; - separation of task requirements from dialogue behavior; - user satisfaction as an empirical outcome; - combining task success and interaction costs.

EQ-Layer overlap - blinded response-level evaluation; - explicit separation of component scores from final user-facing performance.

What to borrow - add task success / constraint satisfaction as an objective endpoint alongside human preference; - report interaction cost: turns, tokens, clarification count, and latency; - do not reduce evaluation to "empathetic preference" alone.

DOI - 10.3115/976909.979652

4. COLLAGEN / collaborative user-goal structure (1998)

Key work - Rich & Sidner (1998), COLLAGEN.

What already existed - application-independent collaboration management; - hierarchical interaction history organized around user and agent goals; - mixed-initiative assistance.

EQ-Layer overlap - model-agnostic control layer; - intent and action organization outside the surface generator.

What to borrow - represent user goals hierarchically instead of a single flat canonical_request; - allow subgoals and abandoned/corrected goals to persist explicitly.

Reference - MERL TR97-21a, COLLAGEN: A Collaboration Manager for Software Interface Agents.

5. GoDiS / information-state dialogue management (1999-2000)

Key work - Larsson et al. (2000), GoDiS - An Accommodating Dialogue System.

What already existed - explicit information state; - update phase separated from move selection; - private/shared state; - questions under discussion; - user information could arrive in flexible order.

EQ-Layer overlap - state -> policy -> realization pipeline; - explicit intermediate state instead of direct generation.

What to borrow - split state into private control state and shared commitments; - add a Questions Under Discussion (QUD)-like stack for unresolved user goals; - make clarification target a named unresolved slot/reference.

Reference - ACL Anthology W00-0302.

6. Corrections in spoken dialogue systems / TOOT (2000-2002)

Key work - Swerts, Litman & Hirschberg (2000), Corrections in Spoken Dialogue Systems. - Litman & Pan (2002), Designing and Evaluating an Adaptive Spoken Dialogue System.

What already existed - empirical study of user corrections of system errors; - evidence that correction behavior depends on dialogue strategy; - dynamically changing user models; - automatic dialogue-strategy adaptation when recognition problems occur.

EQ-Layer overlap - structural correction handling; - repair policy; - dynamic, turn-by-turn state.

What to borrow - classify correction type and its target; - track repeated system failure separately from user affect; - use error/recovery history as an independent policy feature.

DOIs - Corrections: 10.21437/ICSLP.2000-344 - Adaptive TOOT: 10.1023/A:1015036910358

7. Emotional cues inside dialogue management (2002)

Key work - Holzapfel, Fuegen, Denecke & Waibel (2002), Integrating Emotional Cues into a Framework for Dialogue Management.

What already existed - emotion represented explicitly inside dialogue management; - emotional cues used as information for dialogue behavior; - design intended to transfer across domains/tasks.

EQ-Layer overlap - affect as an explicit intermediate control signal; - affect influencing policy rather than merely changing wording.

Novelty implication - EQ-Layer must not claim to be the first system that inserts affect into a dialogue-management layer.

DOI - 10.1109/ICMI.2002.1166983

8. RavenClaw (2003; extended 2009)

Key work - Bohus & Rudnicky (2003/2009), RavenClaw.

What already existed - reusable, task-independent dialogue engine; - explicit separation of task logic from domain-independent conversational behavior; - error handling, turn taking, and conversational skills outside domain logic.

EQ-Layer overlap - reusable middleware; - discrete policy before realization; - task semantics separated from conversational handling.

What to borrow - keep domain/task action separate from social/conversational action; - treat the control layer as a reusable engine, not a bag of prompts.

DOIs - 10.21437/Eurospeech.2003-255 - 10.1016/j.csl.2008.10.001

9. User-tailored response generation / MATCH-SPUR (2004)

Key work - Walker et al. (2004), Generation and Evaluation of User Tailored Responses in Multimodal Dialogue.

What already existed - explicit user models; - response content adapted to the user; - content selection and conciseness manipulated algorithmically; - experimental evidence that tailored responses can improve efficacy.

EQ-Layer overlap - user constraints/preferences; - verbosity/directness control; - response realization conditioned on a user model.

What to borrow - make content selection and surface style two separate decisions; - evaluate concise/direct versus verbose/supportive adaptations independently.

DOI - 10.1207/s15516709cog2805_8

10. Relational agents and embodied social dialogue (2000-2006)

Key work - Cassell et al., Embodied Conversational Agents (2000). - Bickmore & Cassell (2000), social dialogue with embodied agents. - Bickmore dissertation (2003), relational agents. - Bickmore & Picard (2005), long-term human-computer relationships.

What already existed - social dialogue separated from task dialogue; - relationship-building behavior; - adaptation over repeated interactions; - verbal and non-verbal social signals.

EQ-Layer overlap - social calibration; - non-patronizing / relationship-aware response behavior; - conversational policy beyond pure task completion.

What to borrow - optional session-level relational state; - distinguish immediate affect from longer-term interaction history; - do not collapse rapport, politeness, empathy, and affect into one scalar "EQ".

11. POMDP dialogue management / uncertainty (2007)

Key work - Williams & Young (2007), Partially Observable Markov Decision Processes for Spoken Dialog Systems.

What already existed - multiple uncertain dialogue-state hypotheses; - confidence scores; - policy/action selection under uncertainty; - globally optimized behavior rather than hard deterministic state.

EQ-Layer overlap - confidence-gated intent; - ambiguity handling; - policy selection.

What to borrow - replace single-label confidence with a belief distribution over plausible intent/subtext states; - propagate uncertainty into policy choice; - clarification should be an action selected because expected loss is lower, not only because a fixed threshold fired.

DOI - 10.1016/j.csl.2006.06.008

12. Affect-sensitive AutoTutor (2006-2011)

Key work - D'Mello et al. (2006-2008), affect detection in AutoTutor. - D'Mello, Picard & Graesser (2007), affect-sensitive AutoTutor. - later real-time uncertainty detection and adaptive responses.

What already existed - simultaneous cognitive + affective user state; - detection of boredom, confusion, frustration, engagement/flow; - confidence in affect classification; - previous affect state; - selection of pedagogical/motivational dialogue moves based on those states.

This is one of the closest pre-LLM architectural relatives to EQ-Layer.

EQ-Layer overlap - intent/cognitive state + affective state; - history; - confidence; - state-conditioned discrete dialogue move.

What to borrow - explicit previous_affect and persistence duration; - separate cognitive/task state from affect state all the way through policy selection; - evaluate whether adaptation helps only when affect detection is correct.

References - MIT Media Lab, AutoTutor Detects and Responds to Learners Affective and Cognitive States (2008). - Speech Communication (2011), real-time uncertainty detection/adaptation.

13. Affective dialogue management with factored POMDPs (2009)

Key work - Bui, Zwiers, Poel & Nijholt (2009), Affective Dialogue Management Using Factored POMDPs.

What already existed - user affect explicitly factored into dialogue state; - affect changes system action selection; - proof-of-concept policy optimization.

EQ-Layer overlap - affect + task state -> policy.

Novelty implication - any broad claim such as "first architecture to choose dialogue policy from intent and emotion" is indefensible.

What to borrow - factor state/action spaces instead of using a single monolithic policy label.

14. Quality-adaptive dialogue (2014)

Key work - Ultes, Dikme & Minker, quality-adaptive dialogue.

What already existed - online estimate of interaction quality/user satisfaction; - adaptation of confirmation strategy based on current interaction quality; - focus on rescuing problematic dialogue trajectories.

EQ-Layer overlap - escalation/quality trend; - dynamic policy adaptation.

What to borrow - add a rolling interaction_quality state independent of emotion; - distinguish "user is angry" from "interaction is failing"; - use recovery latency and quality improvement as evaluation endpoints.

Reference - ACL Anthology L14-1092.

15. NEMO / externally coupled affect adaptation (pre-modern-LLM)

Key line of work - need-inspired task-independent emotion model (NEMO) integrated externally with a spoken conversational agent.

Why it matters - affective adaptation did not always require rewriting the dialogue manager; external modules could alter the host system.

EQ-Layer overlap - middleware/wrapper architecture.

Novelty implication - "external affect module attached to an existing conversational system" is not by itself new.

16. Sentiment-adaptive end-to-end dialogue (2018)

Key work - Shi & Yu (2018), Sentiment Adaptive End-to-End Dialog Systems.

What already existed - multimodal user sentiment included directly in dialogue learning; - both supervised and reinforcement-learning variants; - sentiment adaptation improved task success and reduced dialogue length in the reported bus-information setting.

EQ-Layer overlap - learned user affect modifying dialogue behavior.

What to borrow - this is a strong pre-modern-LLM baseline family for any claim that affect adaptation improves task-oriented dialogue.

Reference - ACL Anthology P18-1140.

Architectural comparison

Capability Clear pre-LLM precedent? Representative prior art EQ-Layer status
Free-form intent routing Yes How May I Help You? Learned intent
Dialogue-act classification Yes classic SDS / GoDiS Learned
Explicit dialogue state Yes TRAINS, GoDiS Yes
Uncertainty/confidence Yes POMDP dialogue systems Learned posterior + one-step Bayes-risk routing implemented
User model Yes TOOT, MATCH, relational agents Partial
Affect detection Yes affective computing, AutoTutor Yes
Affect changes policy Yes Holzapfel, AutoTutor, Bui Yes
Correction/repair Yes TRAINS/grounding, TOOT Structural repair lifecycle implemented
Task vs discourse separation Yes RavenClaw Partial
Social/relationship behavior Yes REA, relational agents Partial
User-tailored content/style Yes MATCH/SPUR Partial
Satisfaction/quality adaptation Yes PARADISE, Ultes et al. Structural interaction-quality state implemented
Learned policy optimization Yes NJFun, POMDP work Not current core
External affect middleware Yes NEMO line Yes-ish
Frozen general-purpose LLM steering No pre-LLM analogue by definition — Yes
Same base model A/B isolation not an architectural novelty evaluation methodology Implemented
Evidence provenance: direct/derived/structural/verified no clear direct match found in this review — Implemented
Explicit no-go preservation in routing evidence no clear direct match found in this review — Implemented

What EQ-Layer should change next

These are the highest-value lessons from the historical literature.

A. Factored action architecture — IMPLEMENTED

EQ-Layer now compiles the selected policy into:

state + intent -> task_move + social_move + repair_move + realization_controls

The legacy policy selector remains for backward-compatible routing, but it no longer carries the full response decision by itself. task_move is derived primarily from explicit user intent, while social and repair behavior are compiled separately. This prevents affective adaptation from silently replacing the user's task.

Implemented in: - eq_layer/actions.py - eval/factored_action_eval.py

Example:

task_move      = answer_status
social_move    = low_validation
repair_move    = none
verbosity      = low
directness     = high
clarification  = false

This carries forward the factorization lessons from RavenClaw and affective POMDP work and explicitly prevents "emotion policy" from replacing the user's task. Core CI enforces this orthogonality.

B. Interaction Quality as a separate variable — IMPLEMENTED

Conversation failure is no longer inferred from emotion alone. EQ-Layer now tracks a separate structural state:

interaction_quality:
  current
  delta
  repeated_failure_count
  unresolved_repair_count
  clarification_count
  evidence_level

The initial implementation is intentionally conservative. It uses observable repair pressure and clarification overhead rather than pretending to have a learned satisfaction model. A low interaction-quality score can shorten and make realization more direct, but it is explicitly not interpreted as user emotion.

Implemented in: - eq_layer/interaction.py - eval/interaction_state_eval.py

This carries forward the distinction emphasized by quality-adaptive dialogue work: "the user is emotionally activated" and "the interaction is failing" are different variables.

C. Structural repair lifecycle — IMPLEMENTED

Correction is now relational rather than only a flat subtext label:

repair:
  active
  kind
  target_turn_index
  target_excerpt
  correction_excerpt
  repeated
  recent_repair_count
  confidence
  evidence_level

A correction becomes active only when an observable user correction follows a prior assistant turn. Once the assistant responds, the repair is labelled responded_repair rather than "resolved": the system does not claim that its reply actually fixed the misunderstanding without later evidence. Repeated repairs compile to stop_restatement_and_repair.

Crucially, repair state does not imply user_right or user_wrong; factual stance remains externally verified or unknown.

Implemented in: - eq_layer/interaction.py - eq_layer/tracker.py - eq_layer/actions.py - eval/interaction_state_eval.py - eval/factored_action_eval.py

D. Intent belief state + Bayes-risk routing — IMPLEMENTED

The learned intent adapter now preserves a posterior over all supported intent classes instead of collapsing uncertainty to one confidence score. A transparent one-step Bayes-risk router compares the expected interaction cost of each response move with clarification and selects the lowest-risk option.

This is POMDP-inspired uncertainty handling, not a full POMDP. EQ-Layer does not currently learn a transition model or long-horizon value function. The current decision-cost table is hand-specified and auditable; calibration and sensitivity analysis remain open scientific work.

Implemented in: - eq_layer/intent_belief.py - eq_layer/trained_intent.py - eq_layer/intent.py - eval/intent_belief_eval.py

Legacy confidence/margin thresholds remain stored for compatibility and diagnostics, but they no longer determine learned intent routing.

E. Separate short-term affect from long-term user model

Do not store "the user is angry" as a stable user property.

Use: - turn/session affect; - session interaction quality; - optional persistent user preferences only when appropriate.

This avoids conflating affective computing with relational/user modeling.

F. Expand evaluation beyond preference

Add: - intent/constraint satisfaction; - task success; - repair success; - recovery latency; - number of unnecessary clarifications; - response length/token overhead; - latency; - user preference; - interaction quality before/after the response.

This combines lessons from PARADISE, adaptive TOOT, and affect-sensitive systems.

Claims to avoid

Do not write:

  • "the first emotionally intelligent dialogue manager";
  • "the first system to use emotion to choose a response";
  • "the first architecture combining intent and affect";
  • "the first adaptive dialogue policy based on user emotion";
  • "the first middleware that adapts dialogue to user state" without a much narrower qualifier;
  • "EQ-Layer gives LLMs emotional intelligence" as an empirical conclusion.

All have strong prior-art problems or exceed current evidence.

Claims that may be defensible after contemporary prior-art checking

Safer formulation:

EQ-Layer investigates whether an explicit, model-agnostic control layer can improve frozen general-purpose LLM responses by separating user intent, affective state, conversational evidence, discrete interaction-policy selection, and generation steering.

Stronger but still testable formulation:

Unlike prompt-only empathy instructions, EQ-Layer exposes the intermediate control state and policy decision, preserves the provenance and uncertainty of those signals, and can be evaluated by holding the underlying language model fixed across baseline and controlled conditions.

Do not add "first" unless a dedicated contemporary systematic search supports it.

Core references

  1. Picard, R. W. (1997). Affective Computing. MIT Press.
  2. Gorin et al. (1997). How May I Help You? Speech Communication.
  3. Walker et al. (1997). PARADISE: A Framework for Evaluating Spoken Dialogue Agents. ACL/EACL.
  4. Rich & Sidner (1998). COLLAGEN: A Collaboration Manager for Software Interface Agents. User Modeling and User-Adapted Interaction.
  5. Larsson et al. (2000). GoDiS - An Accommodating Dialogue System.
  6. Swerts, Litman & Hirschberg (2000). Corrections in Spoken Dialogue Systems.
  7. Litman & Pan (2002). Designing and Evaluating an Adaptive Spoken Dialogue System.
  8. Holzapfel et al. (2002). Integrating Emotional Cues into a Framework for Dialogue Management.
  9. Bohus & Rudnicky (2003/2009). RavenClaw.
  10. Walker et al. (2004). Generation and Evaluation of User Tailored Responses in Multimodal Dialogue.
  11. Bickmore & Picard (2005). Long-term relational agents.
  12. Williams & Young (2007). POMDPs for spoken dialogue systems.
  13. D'Mello et al. (2007/2008). Affect-sensitive AutoTutor.
  14. Gratch et al. (2007). Creating Rapport with Virtual Agents.
  15. Bui et al. (2009). Affective Dialogue Management Using Factored POMDPs.
  16. Ultes et al. (2014). Quality-adaptive dialogue.
  17. Shi & Yu (2018). Sentiment Adaptive End-to-End Dialog Systems.

Current research conclusion

The historical record strongly supports the following interpretation:

EQ-Layer is an integration/reframing contribution, not a de-novo invention of affective dialogue management.

Its strongest research path is therefore:

  1. make the integration technically cleaner than historical systems;
  2. exploit properties unique to modern general-purpose LLMs;
  3. keep the base model frozen so the layer's causal contribution can be isolated;
  4. expose state/policy decisions for inspection;
  5. preserve uncertainty and evidence provenance;
  6. demonstrate response-level gains under blinded same-model evaluation.

That is a substantially more defensible research story than claiming that intent-aware or emotion-aware dialogue policy itself is new.