Imagine training a sales-support AI that gets stronger automatically—adjusting its conversational skills after each dialogue based on user feedback.
Automatic self-evolution already works well for math and coding: when the AI answers a question, it's either right or wrong, and the signal is crisp. Dialogue doesn't work that way.
The problem lies in a subtle piece of logic: changing the AI's reply changes the entire subsequent course of the conversation.
Suppose a user asks: "Can I return this product?" The AI originally answers: "Yes, returns are accepted within 7 days." The user leaves satisfied. You think the reply could be better and revise it to: "Yes, returns are accepted within 7 days, but you'll need to keep the receipt." The catch—if the AI had said the second version, the user might not have simply walked away satisfied. They might have asked, "What if I lost the receipt?" and the conversation would have gone in a completely different direction.
This is the core difficulty of self-evolution in open-ended dialogue: the feedback signal is a moving target. You can't compare a revised reply against the user's pre-revision reaction, because the two don't exist on the same timeline.
Traditional self-evolution methods (like SkillOpt) assume "the user's reaction stays the same after the reply is changed," and then use the pre-revision reaction to evaluate the revised reply. This assumption holds for math problems (whether an answer is right is independent of the user's reaction), but not in dialogue.
A July 2026 paper proposes a clever workaround: stop trying to predict whether a revised reply is good, and instead predict whether the observed reply will elicit positive feedback.
Future-Feedback Prediction: Turning a Moving Target into a Fixed One
The paper's core idea works in three steps:
Step 1: Don't optimize replies directly—optimize a feedback predictor first.
Given a dialogue history (including the AI's reply), predict whether the user's next feedback will be positive or negative. This prediction task is verifiable—you have a corpus of historical dialogues, each followed by actual user feedback (satisfied/dissatisfied, follow-up question/leaving, like/complaint, etc.). The predictor's accuracy can be measured directly.
Step 2: Use the predictor as a proxy for skill evaluation.
Now suppose you want to improve one of the AI's skills (e.g., "how to introduce product advantages"). You propose a skill modification, generate a batch of replies with the modified skill, and use the predictor from step 1 to assess whether these replies will elicit positive feedback. If the predictor says "yes," the skill modification is accepted; if "no," it's rejected.
Step 3: Iterate.
Verified skill modifications are adopted and feed into the next round of evolution.
The key insight: the predictor evaluates the observed answer, not the counterfactual answer. You don't ask "if the AI had said something else, how would the user have reacted?"—an unanswerable question. You only ask "what feedback did the actually-uttered reply trigger?"—a question with data behind it.
That's the meaning of "Future-Feedback" in the title: not predicting "what would happen if we changed it" but "what the actual outcome triggered." The former is a moving target; the latter is a fixed one.
Why This Sidesteps the Moving-Target Problem
Let's look at the logic more carefully.
Traditional approach (SkillOpt-style): 1. Observe dialogue D = (user question, AI answer A, user reaction R) 2. Propose a revised answer A′ 3. Evaluate how good A′ is — problem: no R′ exists, because A′ never actually happened 4. Compromise: use R to approximate the evaluation of A′ — problem: R is a reaction to A, not to A′
Future-Feedback approach: 1. Observe dialogue D = (user question, AI answer A, user reaction R) 2. Train a predictor f: given (user question, AI answer A), predict R 3. The predictor f learns "what kinds of replies elicit what kinds of feedback" 4. Propose a skill modification; generate a new reply A_new with the modified skill 5. Evaluate A_new with f: f(user question, A_new) predicts what feedback A_new would trigger 6. If f predicts positive feedback, adopt the skill modification
Key difference: the traditional approach tries to evaluate "how much better A′ is than A" (a counterfactual comparison); Future-Feedback evaluates "will A_new trigger positive feedback" (an absolute prediction). The former requires counterfactual data; the latter needs only observational data.
It's like exam grading. The traditional approach asks "is your answer better or worse than the reference answer?"—which requires a reference answer. Future-Feedback asks "can your answer earn points?"—which requires only grading criteria, not a reference answer.
Industrial Case: A Sales-Assistant Dataset
The paper validates the approach on a real industrial setting: a privacy-protected sales-assistant dialogue dataset.
Data characteristics:
- Drawn from actual sales conversations (anonymized)
- Each conversation labeled "resolved" or "unresolved"
- Data cleaning: low-quality and ambiguously labeled dialogues removed
- Balanced splits: test sets contain roughly 50% resolved / 50% unresolved (to prevent gaming via label prevalence)
- No longer equating fluent sales scripts with success: early versions were misled by smooth phrasing; later versions learned to look at substance
- No longer equating a completed workflow with success: finishing the process doesn't mean the user's problem was solved
- No longer equating successful tool calls with success: a tool running without errors doesn't mean the user's need was met
- Weighting coverage, specificity, and next steps more heavily: does the reply address all the user's questions, is it concrete, does it give an executable next action
- Future-Feedback Prediction is suited to screening skill-modification proposals—picking promising candidates from a pool
- Whether a final skill actually works still requires online A/B testing or human evaluation to confirm
- The method's value is in reducing the cost of online testing—not every proposal needs to go live; the predictor filters first
- SkillOpt: uses an LLM as optimizer to propose skill modifications, evaluated on a fixed validation set. Problem: the validation signal is unstable in dialogue settings
- DSPy: a similar approach using "demonstrations" as the validation signal; faces the same dialogue verification difficulty
- Future-Feedback Prediction (this paper): doesn't evaluate skill modifications directly; instead trains a verifiable feedback predictor first and evaluates indirectly through it
- 75% accuracy isn't high enough: at 90%+, a predictor could support more aggressive automatic optimization; 75% only permits conservative screening
- The predictor may fail under distribution shift: if skill modifications drastically change reply style, rules learned from old data may no longer apply
- Only validated in sales scenarios: effectiveness for other dialogue types (support, education, companionship, etc.) is unknown
- Feedback-signal definition: what counts as "positive feedback" must be manually defined, and definitions vary across settings
- Verifying AI alignment: directly verifying "is the AI aligned with human values" is impossible, but you can verify "does the AI's output pass a measurable alignment test"
- Verifying AI capability: directly verifying "can the AI do X" is sometimes impossible (e.g., "can it autonomously complete scientific research"), but you can verify "can it complete a research subtask"
- Verifying AI safety: directly verifying "will the AI deceive humans" is impossible, but you can verify "does the output contain known deception patterns"
The predictor achieved 75%+ accuracy on this dataset.
What does 75% mean? Random guessing yields 50% (given balanced labels), so 75% means the predictor genuinely learned "what kinds of replies resolve problems" rather than exploiting label prevalence.
But the paper is honest about this: 75% is observed accuracy, not counterfactual accuracy. The predictor is 75% accurate on replies that actually happened; how it performs on replies that never happened cannot be directly verified. This is the method's fundamental limitation.
The Learned Rules Are Interpretable
An unexpected bonus: the judgment rules the predictor learns can be read out.
Analyzing the predictor's decision patterns, the paper found it progressively learned to distinguish "surface success" from "actual resolution":
These rules align closely with how human sales experts judge conversations. The predictor didn't just learn statistical correlations—it learned semantic-level judgment criteria.
This is a good example of "skill interpretability." Unlike black-box models, the predictor's rules can be extracted for human review and correction. If the predictor learned a biased rule (e.g., "users with northern accents are easier to satisfy"), humans could detect and fix it.
Positioning: Offline Optimization, Not an Online Replacement
The paper is explicit about the method's scope:
This is an offline optimization stage, not a substitute for online evaluation.
Specifically:
This positioning matters. Many self-evolution papers overclaim "fully automatic evolution," but this paper clearly understands what problem it solves and what it doesn't.
The paper also discusses a deeper philosophical issue: observational verification vs. counterfactual validity.
Observing that "A triggered R" lets you verify "the predictor can predict R." But you cannot infer from this "if A′ had been said, what would have happened." This is a basic problem in causal inference—observational data contains no counterfactual information.
The cleverness of Future-Feedback Prediction: it doesn't attempt to answer counterfactual questions. It makes only observational-level predictions and leaves counterfactual questions to online experiments. It's a design that "knows its limits"—knowing what it can and cannot do, and not overstepping.
Relation to Other Self-Evolution Methods
The paper situates itself in the "textual skill optimization" lineage:
Key distinction: the first two try to evaluate skill quality directly; this paper evaluates indirectly—through a verifiable intermediate layer (the feedback predictor).
It's like how you can't directly measure "does this drug cure the disease" (that requires clinical trials), but you can first measure "does it kill bacteria in a petri dish" (a verifiable intermediate metric). The surrogate doesn't replace the final trial, but it can drastically reduce trial costs.
Limitations
The paper candidly lists several limitations:
Why This Matters More Than It Looks
On the surface, this is an engineering paper about "self-evolving dialogue." At a deeper level, it articulates an important methodology:
When direct verification is impossible, look for a verifiable surrogate task.
This methodology applies to many AI safety and evaluation problems:
This "know your limits before you act" mindset is especially precious in AI. Too many papers attempt unsolvable problems (like "completely eliminating hallucinations") and can only offer approximations. This paper admits "counterfactual verification is impossible" and then works within what's verifiable—an honesty more valuable than a fake breakthrough.
---
Paper: Verifiable Self-Evolution for Open-Ended Dialogue Skills via Future-Feedback Prediction Authors: ChaoJin Zhao, Xuan Jiang Date: July 2026