Keywords: Automated Essay Scoring, Large Language Models, Rubric-Based Assessment, Construct Validity, Mechanistic Interpretability

Introduction

Automated essay scoring (AES) systems built on LLMs are typically validated by score agreement with human raters, which says little about how a model arrives at a score: whether it tracks individual rubric criteria, or leans on one undifferentiated quality signal that happens to correlate with each. This construct-validity question cannot be settled from output scores alone.

We address it mechanistically: probing and steering 25 open-weight LLMs (0.13B-10.31B parameters, 14 families) as they score BESTEP (Taiwan’s Ministry-of-Education English exam) responses across three task types and six analytic rubric criteria. RQ1: is the holistic score decodable from a model’s activations, and where does that signal live? RQ2: do the six criteria occupy distinguishable directions in representation space? RQ3: does steering a criterion’s direction move the score specifically for that criterion, or does any rubric-adjacent push move it about equally? Across all three, decodability, separation, and causal control diverge: rubric information is present and partly separable, but not independently controllable under steering. A larger companion study (80 to 100+ models) extends this design elsewhere.

Related Work

AES and LLM-based scoring. Automated essay scoring first moved from feature-engineered methods to BERT-based deep learning: domain-adaptive training, multi-scale essay representations, and hierarchical transfer learning each closed part of the gap to human-rater agreement (Ramesh and Sanampudi, 2022; Wang et al., 2022; Cao et al., 2020; Xue et al., 2021; Settles et al., 2020). A second wave prompts general-purpose LLMs directly as scorers or raters, with a genuinely mixed record. Some studies report LLM ratings holding up under real psychometric scrutiny – generalizability-theory and many-facet Rasch analyses across thousands of essays and multiple genres (Song et al., 2025; Wang et al., 2025a) – and case studies where an LLM produces both a score and trait-level comments across several analytic criteria at once (Wang et al., 2025b; Cho et al., 2024). Others find the opposite: robustness failures under adversarial stress-testing and demographic bias in ratings (Kabra et al., 2022; Yamashita, 2025), systematic disagreement between LLM and human grading behavior (Mathew et al., 2026), and validity that shifts by language, genre, and proficiency level rather than holding uniformly (Mizumoto and Eguchi, 2023; Pack et al., 2024; Mansour et al., 2024; Gaggioli et al., 2025; Barbosa et al., 2025; Yavuz et al., 2025; Yancey et al., 2023). A related strand looks past scoring accuracy to classroom use, from ChatGPT’s promise and pitfalls as a writing tutor to its measurable effect on students’ own writing when used for feedback (Barrot, 2023; Ding et al., 2025). Closest to our work, Wang et al. (2026) open an LLM-based AES model’s decision process through feature-weighting analysis; we instead intervene causally on internal representations rather than analyzing weights alone.

Rubric-aware and rationale-based scoring. A separate line augments scores with rubric-level explanations or rationales (Bueno et al., 2026; Chu et al., 2025; Jiao et al., 2025; Kucia et al., 2026), and directly probes what implicit rubric structure a transformer-based scorer has learned (Fiacco et al., 2023). These works evaluate rubric outputs; we evaluate rubric representations.

Mechanistic interpretability and faithfulness. Our probing and steering methodology builds on activation-patching practice (Zhang and Nanda, 2023), causal concept-based explanation (Wu et al., 2023b), and work questioning whether model-generated explanations are faithful to internal computation (Atanasova et al., 2023; Jacovi and Goldberg, 2021; Chuang et al., 2026; Ferrando et al., 2023; Demircan et al., 2024), a concern directly relevant to AES, where a plausible-sounding rationale is not evidence of criterion-specific control.

Criterion-adjacent NLP. For the grammar and vocabulary criteria specifically, we draw on grammatical error correction (Bryant et al., 2023; Wu et al., 2023a; Fang et al., 2023) and lexical and proficiency assessment (Aleksandrova and Pouliot, 2023; Bannò et al., 2024; Bannò et al., 2025) literatures, and for inter-rater reliability, the standard kappa statistic (Cohen, 1960).

Setup

Models. We probe and steer 25 open-weight, base (non-instruction-tuned) LLMs spanning 0.13B to 10.31B parameters and 16 to 42 layers across 14 model families (Llama, Gemma, Falcon3, SmolLM2/3, Granite, Mistral, OLMo, TAIDE, Apertus, and others; see Appendix B, Table 5 for the full catalog), selected for complete output coverage across every phase of the pipeline (probing, geometry, and steering). All 25 are cataloged as base models; the current set does not support comparing base against instruction-tuned or chat variants, a gap noted in Limitations.

Dataset. All models are probed and scored on the same fixed set of 75 human-verified BESTEP writing responses (the gold-75), spanning two test forms and three task parts (Table 1): Part 1 (Answering Questions, a short factual-response task), Part 2 (Expressing Opinions), and Part 3 (Integrated Writing, the longest and most demanding task). BESTEP’s rubric assesses six analytic criteria, but not every criterion applies to every part (Table 3): grammar and vocabulary apply throughout, spelling and punctuation is scored only in the short-answer part (Part 1), and argument relevance only in the integrated-essay part (Part 3).

Activation extraction. For every model, we extract final-token residual-stream activations at every layer for each of the 75 responses; these activations are the shared input to the probing (RQ1), contrastive-direction (RQ2), and steering (RQ3) procedures described in the next three sections.

Experiments

RQ1: Score-Information Representation

Method. For each (model, layer, part) cell we fit a Ridge regression probe on final-token residual-stream activations to predict the response’s holistic band score, using leave-one-out cross-validation (LOOCV; Figure 1): each of the 21 (Part 1), 24 (Part 2), or 30 (Part 3) responses is held out once, the probe is trained on the rest, and every held-out prediction is pooled to compute one R2 per cell. We report the peak-R2 layer, the peak R2 itself, and the peak-to-final-layer drop, using the pooled-form scope (both test forms combined) as the headline result.

Results. Score information is reliably decodable across all 25 models, but not in a simple “easy parts early, hard parts late” order. The one robust timing pattern: Part 1 (the shortest, short-answer task) peaks latest or second-latest among the three parts in 23 of 25 models (Figure 2). A response-length explanation for this timing does not hold up: relative peak-to-final signal drop ranges from 0.0% to 2611.3% across models, and the part with the largest drop varies by model, so timing alone cannot be read as a complexity gradient.

Peak timing is not the same as decodability strength, and the two diverge. Part 1 has the lowest peak R2 in 24 of 25 models and is never the highest; across models, mean peak R2 is 0.432 for Part 1, 0.633 for Part 2, and 0.773 for Part 3 (Table 2, Figure 3). Part 1 is also the least stable across models (peak-R2 variance 0.0338, versus 0.0095 for Part 2 and 0.0038 for Part 3). Part 1 has the fewest gold responses (21), the shortest responses (about 47 words on average, versus about 117 for Part 2 and about 214 for Part 3), and the coarsest scale, which plausibly explains why its signal is both weaker and more model-dependent: less textual evidence and fewer observations for the probe to exploit.

The practical reading for RQ1: models clearly carry information about essay quality, and BESTEP’s shortest task type gives that information a later-peaking but weaker and less stable signal than the two longer, more integrated tasks, a pattern this paper cannot yet explain mechanistically, but that RQ2 and RQ3 return to.

RQ2: Construct Representation

Method. RQ2 uses the same 75 gold-standard responses plus criterion-targeted clean-versus-degraded mutations, since holistic scores blend criteria and cannot support construct-geometry claims on their own. For each (model, layer, part, criterion) cell we estimate a contrastive direction as the normalized difference between clean and mutated-essay activation centroids, and its split-half reliability by recomputing the direction from random halves of the essay pairs. We compare same-criterion to different-criterion cosine similarity within the same essay part (a same-part, cross-criterion comparison avoids confounding “different construct” with “different essay”), retaining only reliability-gated rows (split-half reliability at least 0.6).

Results.

Is separation uniform across the assessment? No. Averaged across 1,470 reliability-gated layer-part cells, the mean same-versus-different geometry gap is 0.334 (74.5% of cells positive), but this is not stable across task parts: Part 1’s four criteria are, on average, negative (mean gap -0.220, 6.2% of rows positive) — no more separated from each other than a single criterion is from itself across parts — while Parts 2 and 3 separate cleanly (89.7% and 99.2% of cells positive): reliable, criterion-specific directions are broadly available in Parts 2 and 3 but sparse in Part 1.

Which specific pairs are, or are not, separated? Pooling all reliability-gated rows by literal criterion pair, using disattenuated cosine similarity (a correction for each direction’s own measurement noise, so a value of 1 means statistically indistinguishable from the same construct and 0 means unrelated) shows separation is highly uneven across the 12 comparable pairs (three of the 15 possible pairs never co-occur in the same part and are not comparable at all; Table 4, alongside RQ3’s steering results below). argument_relevance paired with grammar is the most separated pair (-0.102, anti-correlated beyond noise), and argument_relevance, a Part-3-only criterion, carries most of the separation signal in the whole rubric. At the opposite end, grammar paired with spelling_punctuation sits at the 1.0 ceiling: the two are statistically indistinguishable from the same construct, and spelling_punctuation paired with vocabulary is nearly as entangled (0.984). Both of the most-entangled pairs are Part-1-only comparisons, consistent with Part 1’s negative geometry gap above.

The RQ2 picture is therefore not “criteria are separated” or “criteria are collapsed,” but a specific, checkable pattern: separation is broad in the two longer, more integrated task types, concentrated in one criterion (argument relevance) beyond that, and absent in the shortest task type.

RQ3: Causal Specificity

Method. RQ3 reuses RQ2’s criterion directions and tests them causally rather than only representationally: for degraded essays with a known target criterion, we inject the matching criterion direction (self), a mismatched rubric direction (other), an orthogonal control, or a generic holistic-score direction into a target layer at four injection strengths (alpha = 0.05, 0.1, 0.25, 0.5), and measure the resulting change in expected band score. Self and other effects are paired at the finest valid unit, (model, layer, part, form, degradation, alpha), giving 74,304 paired cells across all 25 models, and we test whether the matching direction beats a mismatched one by a practically meaningful margin, not just a statistically detectable one.

Results.

Does steering work at all? Yes, for five of six criteria: grammar, vocabulary, cohesion, spelling/punctuation, and task relevance each show a confidently positive self effect (0.168 to 0.359 score points). Only argument_relevance’s own direction produces no detectable effect, not a backwards lever, simply indistinguishable from doing nothing.

Is steering selective? No. Pooled across all paired cells, the matching direction beats a mismatched one by only 0.0074 score points on average (95% CI [0.0036, 0.0117]; 1.74% of the absolute self effect, 95% CI [0.85%, 2.74%]), and this holds under every robustness check we ran: the gap holds at the cell level (0.0085, 95% CI [0.0051, 0.0119]) and never exceeds 0.0083 under leave-one-model/family-out.

The pooled number, however, hides two different mechanisms. Broken out pair by pair (Figure 4), among the four criteria that both work and are reliably measured (grammar, vocabulary, cohesion, spelling/punctuation), the matching direction shows no detectable advantage over a mismatched one in 11 of their 12 head-to-head comparisons; the one exception is vocabulary losing to spelling/punctuation, not because vocabulary is weak, but because spelling/punctuation happens to be the single strongest general-purpose lever of the six. argument_relevance loses every comparison it has data for, for the different reason that its own direction never moves the score. Re-run separately within each task part (36 within-part comparisons among comparably-strong criteria), the same pattern holds in every part: not one instance of a criterion beating a comparably strong rival specifically because it matched the essay’s degradation.

Tying back to RQ2, a pair’s representational separation (Table 4, above) does predict how large a causal difference that pair shows under steering (r=0.717 across all comparable pairs), but it does not predict which direction wins: geometric distance sets the scale of the effect, not its correctness.

Discussion

Taken together, the three research questions tell a coherent but sobering story about what a rubric-based AES model actually does inside. The model carries real, decodable information about essay quality (RQ1), and it represents at least some rubric criteria as distinguishable directions rather than one undifferentiated quality signal (RQ2), but distinguishable representation does not translate into independently controllable behavior (RQ3). Steering a rubric direction reliably moves the score, but pushing the matching direction is, for the criteria that are both effective and reliably measured, essentially interchangeable with pushing a different one. The model appears to exercise one shared control with several equally effective handles, not six independent, criterion-specific ones.

This has a direct implication for how AES validity is usually argued. High agreement between a model’s scores and human raters, or the mere fact that rubric criteria are statistically decodable from a model’s internals, is not evidence that the model is using those criteria as separately controllable scoring levers. A model can appear to track grammar, vocabulary, and cohesion separately in its representations while still collapsing them into one causal pathway when the score is actually produced. Construct validity claims that stop at score agreement or decodability risk overstating how analytically faithful an LLM-based scorer really is.

The pattern is also not uniform across the assessment design, which matters for how broadly this conclusion should be read. Part 1, BESTEP’s shortest, short-answer task, shows the weakest and least stable score signal (RQ1) and no meaningful criterion separation at all (RQ2), while the two longer, more integrated task parts support both a stronger signal and real, if not independently controllable, criterion structure. Whatever a model is doing with an assessment’s shortest task type may look different, representationally, from what it does with a longer one, and validity claims should be checked at that grain rather than assumed to generalize across task format.

Limitations

This study is a 25-model proof of concept, not the full-scale analysis. All 25 models are base (non-instruction-tuned); we cannot yet compare base against instruction-tuned or chat variants, a factor plausibly relevant to how a model handles rubric-following instructions. The gold-standard set is 75 essays from one assessment (BESTEP); a larger 700+ essay background corpus exists but is not yet human-verified at this scale, and we are currently completing that verification so the same experiments can be re-run against the larger corpus. Model coverage tops out at 10.31B parameters, so results may not hold for substantially larger (around 50B or more) models. argument_relevance’s own steering direction never reaches significance in this study; whether that reflects a genuine representational gap or a criterion that needs more than surface language features to steer is an open question. More models, additional writing assessments, and non-English rubric-based assessments could all be tested to check whether these findings hold beyond BESTEP.

Conclusion

We asked whether an LLM-based AES model’s rubric criteria are genuinely represented and causally controllable, not just correlated with agreement. Across 25 open-weight models, holistic scores are decodable and rubric criteria are partly separable in representation space, but neither property implies independent causal control: steering a rubric direction moves the score as a general lever, not as a criterion-specific one, and this holds for the great majority of the criteria this design can measure reliably. Distinguishable but not independently controllable is, therefore, not a criticism of one model or one assessment, but a validity check that any rubric-based LLM scorer should be expected to pass before its rubric labels are treated as separately meaningful, one that, on this evidence, current models do not yet pass.

Per-Model RQ1 Depth Profiles

Figure 5 shows the RQ1 depth-profile curve (Figure 2’s pooled version, disaggregated) for each of the 25 models individually, for readers who want to check whether the pooled pattern in Figure 2 holds for a specific model rather than only in aggregate.

Model Catalog

Table 5 lists the 25 open-weight base models probed and steered in this study, sorted by parameter count.

References