Self-Evaluation Elicitation (SEE)

• Self-Evaluation Elicitation (SEE) is a set of methods that recover latent evaluative signals inherent in systems, enabling self-assessment without learning from scratch.
• In LLM applications, SEE allows models to predict judge scores with notable metrics such as 77% top-5 token accuracy and calibration improvements up to 0.746.
• Related approaches in fairness assessment and reinforcement learning use SEE to disentangle self-consistency from reward signals, promoting equitable evaluation and robust performance.

Searching arXiv for papers on "Self-Evaluation Elicitation (SEE)" and closely related usages to ground the article. Self-Evaluation Elicitation (SEE) denotes a set of elicitation-oriented methodologies that recover evaluative information already latent in a system rather than treating that information as a capability that must be learned from scratch. In recent arXiv literature, the label is used in at least two distinct but related senses. In LLM research, SEE most directly refers to judge-aligned self-evaluation: a model predicts how an external judge will score its own open-ended output, and post-training is framed as surfacing a latent capability rather than acquiring a new one (Zhang et al., 3 Jun 2026). In fairness-oriented human evaluation, SEE denotes “Equitable Evaluation via Elicitation,” an interactive assessment framework that infers latent skill while reducing dependence on stylistic differences in self-report (Du et al., 24 Feb 2026). Closely related reinforcement-learning work does not use the acronym, but formalizes “self-consistency elicitation” as a distinct causal component of performance gains under majority pseudo-reward, thereby sharpening the conceptual boundary between elicitation and genuine reward-design signal (Gao, 4 Jun 2026).

1. Conceptual scope and term usage

Within the LLM self-evaluation literature, SEE is centered on a precise question: can a model predict how an external judge will score its own answer? The 2026 formulation treats this as a problem of judge-aligned self-evaluation in the open-ended setting, where there is no ground-truth answer to compare against and the judge scores responses on multiple quality dimensions (Zhang et al., 3 Jun 2026). The method’s defining claim is that the relevant ability is “largely present before any targeted training,” so the primary role of post-training is elicitation.

A different 2026 paper uses SEE to mean “Equitable Evaluation via Elicitation.” There, the target is not model self-scoring but latent human skill assessment under heterogeneous self-presentation styles. The core problem is that equally qualified individuals may differ in demeanor, self-promotion, or modesty, making direct comparison of self-descriptions systematically unreliable. The proposed remedy is an interactive AI for skill elicitation that provides accurate determination of skills while simultaneously allowing individuals to speak in their own voice (Du et al., 24 Feb 2026).

A third strand is conceptually adjacent rather than terminologically identical. In reinforcement learning from verifiable rewards, “self-consistency elicitation” is defined as the effect of a majority pseudo-reward or spurious reward that sharpens a policy toward its modal answer. That mechanism can improve accuracy when the modal answer is already correct, or reinforce error when it is not. The central claim is that prior “true minus random” comparisons conflate this elicitation effect with genuine reward-design signal (Gao, 4 Jun 2026).

Taken together, these usages assign “elicitation” a common role: extracting latent structure from an existing system, whether the system is a base LLM, a human respondent, or a policy trained under RLVR. The objects being elicited, however, differ: judge-aligned quality estimates, latent human skill, and modal-answer consistency.

2. Judge-aligned self-evaluation in base LLMs

In the judge-aligned setting, SEE asks whether a LLM can estimate how an external evaluator would score its own output without requiring a separate judge at inference time. The motivating applications are reranking, deferral, and escalation: if a model can estimate its own judged quality, it can rerank its own samples, defer low-confidence or low-quality queries, or escalate difficult prompts to a stronger model (Zhang et al., 3 Jun 2026).

The empirical centerpiece is that the ability is already present in the base model. Qwen3-4B-Base, prompted few-shot, predicts GPT-5.4’s five-attribute HelpSteer2-style scores “well above chance across three benchmarks.” The reported HelpSteer2 validation calibration score is 0.63, calibration across the three open-ended benchmarks is about 0.50–0.70, and on HelpSteer2 validation the judge’s score is already in the model’s top-5 predicted score tokens 77.07% of the time before training (Zhang et al., 3 Jun 2026). This supports the paper’s conclusion that judge-aligned self-evaluation is “mostly an elicitation problem, not an acquisition problem.”

The output format is explicitly structured. The model first generates a normal response and then emits a [SELF_EVAL] ... [/SELF_EVAL] block containing integer scores in [0,9] for five HelpSteer2 attributes: helpfulness, correctness, coherence, complexity, and verbosity. An external judge scores the same answer on the same five attributes. This couples answer generation and self-evaluation without collapsing them into a single scalar.

The earlier selective-generation literature provided an important precursor. Rather than scoring free-form generations by sequence likelihood or perplexity, it reformulated open-ended generation into token-level self-evaluation tasks. The two primary modes were “Sample and Select,” which turns candidate answers into a multiple-choice selection problem, and “Sample and Eval,” which asks whether a given answer is correct. A “None of the above” option makes uncertainty explicit, and the hybrid “Sample and Select and Eval” combines selection with pointwise rescoring (Ren et al., 2023). That line of work established the general proposition that token-level self-evaluation can be a better proxy for quality than sequence-level likelihood.

3. SEE training cycle, supervision design, and empirical results

The 2026 SEE pipeline alternates two phases: Calibration-Coupled Reinforcement Learning and Masked Judge Distillation. The first phase improves the answer and nudges the self-evaluation toward the judge at the same time. The second phase replaces the self-evaluation block with the judge’s scores and fine-tunes only on the self-evaluation tokens, leaving the answer untouched (Zhang et al., 3 Jun 2026).

This separation is methodologically central. In the reinforcement-learning phase, malformed outputs receive a fixed reward of -1, so the system “fails closed.” The quality term averages only helpfulness, correctness, and coherence, while the calibration term is computed over all five attributes, including complexity and verbosity. The calibration exponent $\gamma > 1$ penalizes large disagreements more strongly than small ones. RL is optimized with GRPO, using rollouts per prompt $N = 8$, cycles $C = 15$, RL steps per cycle $K = 10$, learning rate $1 \times 10^{-6}$, KL loss coefficient 0.01, entropy coefficient 0.001, reward weights $(w_q, w_c) = (0.7, 0.3)$, and calibration exponent $\gamma = 2$; the distillation stage uses SFT_MAX_SAMPLES = 400, one epoch, learning rate $2 \times 10^{-6}$, and batch size 32 (Zhang et al., 3 Jun 2026).

Data efficiency is one of the paper’s strongest empirical claims. SEE uses 160 unique examples, roughly 31x fewer than the reinforcement-learning baseline, which uses about 5,000 unique examples. The total exposure is also smaller: SEE uses 2,400 total sample-passes, whereas Adapted RLCR uses about 10,000 sample-passes (Zhang et al., 3 Jun 2026).

Evaluation spans HelpSteer2 validation plus three open-ended instruction-following benchmarks: LC AlpacaEval 2.0, Arena-Hard-Auto v2.0, and WildBench v2. On LC AlpacaEval 2.0, response win-rate rises from 0.500 for the base model and 0.534 for Adapted RLCR to 0.592 for SEE; quality increases from 0.788 and 0.789 to 0.792; calibration increases from 0.702 and 0.716 to 0.746. On Arena-Hard-Auto v2.0, the corresponding values are 0.500, 0.502, and 0.518 for response win-rate; 0.568, 0.569, and 0.575 for quality; and 0.517, 0.547, and 0.609 for calibration. On WildBench v2, response win-rate moves from 0.500 and 0.537 to 0.581, quality from 0.560 and 0.561 to 0.567, and calibration from 0.504 and 0.541 to 0.609 (Zhang et al., 3 Jun 2026).

The localization claim is equally important. The paper argues that the elicited self-evaluation is sharply localized in the model’s own token distribution rather than buried in a low-probability tail. The reported top-5 token accuracy is 0.8776 on HelpSteer2 validation and 0.9078 on LC AlpacaEval 2.0. Cross-judge generalization is also tested without retraining: the ranking SEE > Adapted RLCR > base is preserved under Claude Sonnet 4.6 and Gemini 3.1 Flash-Lite, even though absolute values shift. The paper interprets this as evidence for a transferable notion of quality rather than a single judge’s preference, while also noting that this does not establish alignment with human preference because all judges are themselves LLMs (Zhang et al., 3 Jun 2026).

4. Causal partition of self-consistency elicitation in RLVR

The RLVR literature introduces a more formal causal decomposition of elicitation. The paper defines four reward conditions: Frozen, Random, Spurious, and True. Frozen is the base model $\pi_0$ with no training update; Random uses an iid Bernoulli(0.5) reward independent of response; Spurious rewards whether the response matches the group majority answer; True rewards ground-truth correctness. The associated accuracies are $a_F$, $N = 8$0, $N = 8$1, and $N = 8$2 (Gao, 4 Jun 2026).

Its key identity is a telescoping decomposition:

$N = 8$3

Here $N = 8$4 is the null or random-control effect, $N = 8$5 is self-consistency elicitation, and $N = 8$6 is reward design. The naive estimator is

$N = 8$7

The paper’s central claim is that $N = 8$8 is systematically biased as an estimate of reward-design value because it conflates self-consistency elicitation with genuine reward-design signal (Gao, 4 Jun 2026).

The simulator is a controlled tabular-GRPO environment parameterized by a prior-strength variable $N = 8$9, interpreted as the probability that the base model’s modal answer is correct. The prior sweep uses $C = 15$0. When $C = 15$1, majority pseudo-reward strengthens correct behavior; when $C = 15$2, it strengthens wrong behavior; near $C = 15$3, the elicitation term is small and unstable (Gao, 4 Jun 2026).

The reported pattern is strongly prior-dependent. The abstract states that the reward-design fraction of the naive estimator ranges from 0.139 at weak prior $C = 15$4 to 0.05 at strong prior $C = 15$5, and that the elicitation term flips sign at the self-consistency crossover (Gao, 4 Jun 2026). A pre-registered $C = 15$6 factorial, with factors reward type, filtering, and prior strength, confirms non-additivity: the interaction ratio is 0.385 and the $C = 15$7 effect is $C = 15$8. Since the pre-registered invalidation threshold is $C = 15$9, the additivity assumption is rejected. A points-vs-bounds pilot gate further distinguishes strong-prior regimes, which are point-identified, from near-crossover regimes, which are only bounded. Re-audits of two named published families yield “ELICITATION DOMINATED” with elicitation share 0.98 and “REWARD DESIGN DOMINATED” with reward-design share 1.18, and the paper releases a reusable one-command harness for running the same audit on alignment papers (Gao, 4 Jun 2026).

This decomposition matters for SEE because it isolates an elicitation term that is neither mere noise nor equivalent to verifier value. In this setting, elicitation is a causal component with its own sign, magnitude, and dependence on prior strength.

5. Equitable Evaluation via Elicitation

In “Equitable Evaluation via Elicitation,” SEE is redirected from model self-scoring to the fair assessment of people. The paper begins from the observation that individuals with similar qualifications and skills may vary in their demeanor, or outward manner, and that some tend toward self-promotion while others are modest to the point of omitting crucial information. Comparing the self-descriptions of equally qualified job-seekers with different self-presentation styles is therefore problematic (Du et al., 24 Feb 2026).

The proposed solution is an interactive AI for skill elicitation. Rather than treating a person’s self-description as ground truth, the system asks a sequence of questions, updates its belief about the latent trait, and chooses the next question adaptively to reduce uncertainty. The abstract presents two goals simultaneously: accurate determination of skills and preservation of the person’s own voice (Du et al., 24 Feb 2026).

Training relies on synthetic humans. To obtain sufficient training data, the paper trains an LLM to act as synthetic humans. This makes it possible to simulate individuals who are equally skilled but differ in how modestly, candidly, or strategically they answer. In the paper’s framing, elicitation mitigates endogenous bias arising from individuals’ own self-reports, and synthetic humans serve as a training and calibration tool for learning robust questioning strategies across presentation styles (Du et al., 24 Feb 2026).

The fairness criterion is “equitability.” The abstract describes it as a mathematically rigorous notion ensuring that the covariance between self-presentation manner and skill evaluation error is small. This distinguishes the framework from standard one-shot prediction from self-reports: the target is latent skill, while the nuisance factor is self-presentation manner. The paper situates possible deployments in settings such as onboarding on a professional networking platform and matching employees to needs during a company reorganization (Du et al., 24 Feb 2026).

Relative to judge-aligned LLM SEE, the main structural similarity is the treatment of observed answers as signals rather than as direct truth. The difference is that the latent quantity is human skill rather than judged response quality, and the normative constraint is equitability rather than calibration to a judge.

6. Broader elicitation paradigm, boundaries, and terminology

SEE sits within a wider research trend that interprets many post-training gains as amplification of latent capabilities rather than insertion of entirely new ones. A theory-grounded elicitation instrument, NeuReasoner, makes that claim explicit for reasoning. It is training-free, uses one model only, and integrates a Neuro Lens with a Cognitive Lens through internal modularization. On CogBench and standard mathematical and coding benchmarks, it matches or exceeds thinking-mode baselines on arithmetic reasoning, code generation, Bayesian reasoning, and reward learning; the gains persist against self-consistency and iterative-refinement baselines matched to its per-decision call budget. At the same time, the paper identifies clear boundaries: risk-taking and decision making under uncertainty remain hard to recover through elicitation alone, and model scale can either widen or erase elicitation advantages depending on the cognitive signature (Javadov et al., 29 Jun 2026).

That broader result clarifies the status of SEE-like methods. It suggests that elicitation is not a universal solvent. In judge-aligned self-evaluation, the latent signal appears recoverable and sharply localized in the token distribution (Zhang et al., 3 Jun 2026). In RLVR, the elicited component can even become harmful below the self-consistency crossover (Gao, 4 Jun 2026). In fairness-oriented human evaluation, elicitation is valuable precisely because the observable reports are endogenously biased by presentation style (Du et al., 24 Feb 2026). Across these settings, what is being elicited, and whether elicitation helps, depends on the causal structure of the task.

The acronym itself is not unique to these literatures. “SEE” is also used for “Self-Evaluated Expertise” in experimental physics, where it denotes a measure of the self-recognition component of student physics identity extracted from E-CLASS responses (Fox et al., 10 Apr 2025). That unrelated usage underscores a terminological point: in current research, SEE is best interpreted through local context rather than acronym expansion alone.

The main unresolved questions are correspondingly domain-specific. For judge-aligned self-evaluation, the 2026 LLM study is explicit that evidence comes from one base model, one model family, one data scale, and LLM judges rather than humans (Zhang et al., 3 Jun 2026). For causal RLVR attribution, the key issue is identification: some regimes are point-identified, whereas near-crossover regimes are only bounded (Gao, 4 Jun 2026). For equitable evaluation, the open question is how well synthetic-human training distributions capture real-world self-presentation heterogeneity (Du et al., 24 Feb 2026). These limitations do not collapse the concept of SEE; they define its current research frontier.