Separating Constraint Compliance from Semantic Accuracy: A Novel Benchmark for Evaluating Instruction-Following Under Compression

AI summary

Overview

Research area: Natural Language Processing — evaluation of large language model instruction-following and prompt compression.

Technical level: Intermediate. The paper is readable without deep mathematical background, but it assumes familiarity with LLM prompting, RLHF, and basic statistics (correlation, p-values, Cohen's d, Fleiss' kappa).

Scope: The paper introduces the Compression-Decay Comprehension Test (CDCT), a benchmark that separately scores constraint compliance and semantic accuracy for 9 frontier LLMs across 8 concepts and 5 compression levels.

What This Paper Is About

When a prompt is shortened (compressed), language models get worse — but existing evaluations blend two different kinds of failure together: breaking a formatting rule versus losing knowledge. That mixing makes it impossible to tell whether a model failed because it could not follow an instruction or because it no longer had enough context to know the answer. This paper builds a benchmark that scores those two things independently across prompt lengths, from roughly 2 words to roughly 135 words, and then runs an experiment to test why the instruction-following failures happen.

Key Contributions

1. A benchmark that separates two failure modes. The Compression-Decay Comprehension Test (CDCT) measures Constraint Compliance (CC) — whether the response obeys an exact 35-word limit — and Semantic Accuracy (SA) — whether the explanation of a scientific concept is correct — as independent scores, using a three-judge LLM jury (Claude Opus 4.1-2, GPT-5.1, DeepSeek-v3.1) that rates each dimension from 0 to 10.

2. Discovery of a near-universal U-shaped compliance curve. Constraint compliance is high at extreme compression (c=0.0, roughly 2 words) and at no compression (c=1.0, roughly 135 words), but dips sharply at medium compression (c=0.5, roughly 27 words). The paper reports this pattern in 97.2% of experiments (70 of 72), with a mean magnitude of 0.381 ± 0.111.

3. Evidence that the two dimensions are statistically orthogonal. Across 81 experiments reported, the Pearson correlation between CC and SA is r = 0.193 (95% CI: [-0.025, 0.396]), p = 0.084 — not significant at α = 0.05. The average absolute CC change magnitude (mean = 0.381) is 2.9× larger than the average absolute SA change magnitude (mean = 0.090).

4. Experimental validation of a mechanism via RLHF ablation. Removing "helpfulness" language from system prompts at c=0.5 improved constraint compliance by 598% on average (median 525%), across 71 of 72 trials (98.6%), with 57 of 72 trials (79.2%) reaching perfect compliance. The author reads this as showing that RLHF-trained helpfulness behavior is the dominant cause of constraint failures at medium compression, not a mere contributing factor.

Main Findings

• A universal U-curve in constraint compliance. Of 72 total experiments (9 models × 8 concepts), 70 (97.2%) show the U-shape, with compliance highest at both extremes and lowest at c=0.5. Mean CC at c=0.0 is 8.12 ± 0.92 (95% CI: [7.91, 8.33]); at c=0.5 it is 6.54 ± 1.18 (95% CI: [6.28, 6.80]); at c=1.0 it is 8.03 ± 1.01 (95% CI: [7.81, 8.25]). The trough is significant against both extremes by paired t-test (p < 0.001 for both).

• Constraint compliance is a reliable measurement; semantic accuracy is much less so. Fleiss' kappa across all 72 experiments on discretized ratings (threshold = 0.7) is κ = 0.90 for CC (almost perfect agreement), κ = 0.25 for SA (fair agreement), and κ = 0.19 for Functional Completeness (slight agreement). The author uses this asymmetry to argue that constraint failures are objective behavioral-mode errors rather than subjective judgments.

• The dimensions do not move together. The CC drop from c=1.0 to c=0.5 (mean delta = 1.49, 95% CI: [1.21, 1.77]) is 33.1× larger than the SA change over the same range (mean delta = 0.045, 95% CI: [0.025, 0.065]). By contrast, CC change from c=1.0 to c=0.0 is only 0.09 (95% CI: [-0.03, 0.21]).

• Semantic accuracy rises monotonically with context. Mean SA deltas step upward at every level: +0.018 from c=0.0 to c=0.25 (95% CI: [-0.002, 0.038]), +0.021 from c=0.25 to c=0.5 (95% CI: [0.005, 0.037]), +0.026 from c=0.5 to c=0.75 (95% CI: [0.011, 0.041]), and +0.025 from c=0.75 to c=1.0 (95% CI: [0.010, 0.040]). Unlike CC, SA shows no U-curve.

• Reasoning-optimized models beat efficiency-optimized models. Reasoning models (O3, GPT-5, O4-Mini) outperform efficient models by 27.5% on constraint compliance (mean CC 8.20 versus 7.30; paired t-test p < 0.001, Cohen's d = 0.96). The gap is largest at c=0.5, where reasoning models stay above CC 7.5 while efficient models fall to roughly CC 6.0.

• Per-model differences span a wide band. Averaged across all compression levels, O3 leads at CC 8.32 ± 0.88 and SA 8.91 ± 0.52; Gemini 2.5 Flash is lowest at CC 6.98 ± 1.31 and SA 8.19 ± 0.76. GPT-5 scores CC 8.18 ± 0.95 / SA 8.76 ± 0.61, O4-Mini 8.09 ± 1.01 / 8.54 ± 0.68, Claude Opus 4.1 7.67 ± 1.09 / 8.61 ± 0.58, Claude Sonnet 4 7.45 ± 1.18 / 8.48 ± 0.65, DeepSeek-v3 7.34 ± 1.22 / 8.41 ± 0.69, GPT-4.5 7.21 ± 1.24 / 8.32 ± 0.71, and Llama 4.1 405B 7.12 ± 1.28 / 8.25 ± 0.73.

• Domain matters for semantics but not for the U-curve. SA deltas from c=1.0 to c=0.0 are -0.012 ± 0.082 for formal sciences, +0.067 ± 0.103 for natural sciences, and +0.145 ± 0.178 for applied sciences. The author interprets this as formal concepts (modus ponens, recursion, derivative) being well encoded in model weights so that minimal context suffices, while applied concepts (harm principle, impressionism) benefit more from added context.

• The RLHF ablation was dramatic and uniform. At c=0.5, the reported baseline-to-ablated CC shifts per model are: gpt-oss-120b 0.08 → 1.00 (+1150%), o4-mini 0.16 → 0.99, mistral-medium-2505 0.16 → 0.99, claude-haiku-4-5 0.17 → 0.99, gpt-5 0.17 → 1.00, phi-4 0.18 → 1.00, Llama-4-Maverick 0.21 → 0.98, grok-4-fast-non-reasoning 0.22 → 0.95, and o3 0.29 → 1.00 (+245%). All 9 models and all 8 concepts showed improvement. The author notes that grok-4-fast had one trial already at perfect baseline, pulling its post-ablation average to 0.95.

• A concrete qualitative example. On "impressionism" at c=0.5, the baseline GPT-5 response ran 149 words (violating the 35-word limit by 414%), while the ablated response ran 24 words and retained semantic accuracy.

• The author flags two predictions the ablation overshot and one inconsistency between tables. The ablation improved CC by 598% average, far beyond the 40–50% the Constraint Salience Hypothesis predicted. Also note that the paper describes model coverage inconsistently: Table 1 lists the 9 evaluated models as O3, GPT-5, O4-Mini, GPT-4.5, Claude Sonnet 4, Claude Opus 4.1, Gemini 2.5 Flash, Llama 4.1 405B, and DeepSeek-v3, while Table 3 (the ablation) lists gpt-oss-120b, o4-mini, mistral-medium-2505, claude-haiku-4-5, gpt-5, phi-4, Llama-4-Maverick, grok-4-fast-non-reasoning, and o3 — a different set of nine names. The paper also alternates between describing 72 experimental conditions (9 models × 8 concepts) and 81 experiments, and between 97.2% and 97.5% U-curve prevalence.

Methodology in Plain English

The researchers wrote a detailed prompt asking a model to explain a scientific concept in exactly 35 words, then used GPT-5.1 to rewrite that prompt at five shorter lengths, from about 135 words down to about 2 words, while keeping the 35-word requirement in the text. They chose 8 concepts spanning formal sciences (modus ponens, recursion, derivative), natural sciences (photosynthesis, natural selection), and applied sciences (harm principle, impressionism, theory of mind). Each of 9 models then answered at each compression level with temperature set to 0. To score the answers without relying on human annotators, they ran a jury of three architecturally different judge models — Claude Opus 4.1-2, GPT-5.1, and DeepSeek-v3.1 — each rating Constraint Compliance, Semantic Accuracy, and Functional Completeness on 0–10 rubrics, with scores averaged. They then checked whether the three judges agreed (Fleiss' kappa), correlated CC against SA to test independence, and compared compression levels pairwise with paired t-tests, Cohen's d, and 95% confidence intervals at α = 0.05. Finally, they ran an ablation: they re-ran the worst compression level (c=0.5) for all 72 conditions using system prompts stripped of RLHF "helpfulness" language such as words encouraging comprehensive or detailed answers, leaving only the 35-word constraint, and measured the change.

Why This Matters

Impact on research. The paper argues that prior compression and instruction-following work (LLMLingua, FollowBench, IFEval) confounds two separable failure modes. By showing that CC and SA are statistically orthogonal and that CC is measurable at κ = 0.90 while SA sits at κ = 0.25, it offers a way to diagnose whether a model failure is a knowledge problem or an instruction-salience problem. If the RLHF ablation result holds, it also reframes a large class of instruction-following failures as an alignment side-effect rather than a capability limit.

Real-world applications:

• Prompt design for constrained outputs — the author recommends avoiding the 20–35 word prompt range (c=0.4–0.6) and instead writing prompts either under 10 words or over 60 words.
• Safety guardrails and format enforcement — the framework generalizes to toxicity constraints, JSON or structured output requirements, and tone/style constraints, where constraint adherence can be scored separately from task quality.
• API and UI character or token limits — the 35-word constraint is presented as representative of character limits in interfaces and API token budgets.
• Multi-turn dialogue consistency — the same separate-scoring approach is proposed for tracking behavioral constraint drift across turns.

Industry relevance. The 27.5% CC advantage for reasoning-optimized models, and the finding that 79.2% of ablated trials reached perfect compliance, both suggest that constraint-following debuggability is a lever teams can pull without retraining on new knowledge. The practical claim is that a prompt that is too short can outperform a prompt that is medium-length for constraint-critical tasks — a counterintuitive but directly actionable result.

Future Directions

• Human annotation studies. The current jury is entirely LLM-based; the authors acknowledge this may introduce systematic biases despite the three-architecture design, and call for human validation.
• Broader model and concept coverage. The study covers 9 models and 8 concepts, all factual; creative domains are untested and the authors flag this as a generalizability gap.
• Comparison with algorithmic compression. Compression here is model-generated rewriting, not algorithmic schemes like LLMLingua, so it is unclear whether the U-curve would survive fragmented, algorithmically compressed text.
• Mechanistic tests of constraint salience. Three predictions remain unvalidated: that formatting the constraint as "EXACTLY 35 WORDS" would raise CC by 30–40% at c=0.5, that attention weights on constraint tokens would be minimal at c=0.5 with attention entropy peaking there, and that U-curve magnitude should be concept-independent because the word-count constraint is constant across concepts.

Target Audience

Researchers and engineers working on LLM evaluation, prompt compression, and instruction-following robustness. It is most useful to practitioners who design constrained-generation systems (structured output, length limits, guardrails), to alignment researchers interested in how RLHF helpfulness signals interact with explicit instructions, and to benchmark designers looking for a template that scores orthogonal behavioral dimensions separately. Readers without a statistics background can follow the core narrative but will need to interpret the correlation, effect-size, and agreement metrics to get full value.

Authors’ abstract

Large language models (LLMs) exhibit degraded performance under prompt compression, but the mechanisms remain poorly understood. We introduce the Compression-Decay Comprehension Test (CDCT), a benchmark that independently measures constraint compliance (CC) and semantic accuracy (SA) across compression levels. We evaluate 9 frontier LLMs across 8 concepts using 5 compression levels from extreme (c=0.0, ~2 words) to none (c=1.0, ~135 words). A three-judge LLM jury achieves almost perfect inter-rater agreement on CC (Fleiss' \k{appa}=0.90). We observe a universal U-curve pattern in constraint compliance (97.2% prevalence), with violations peaking at medium compression (c=0.5, ~27 words). Counterintuitively, models perform better at extreme compression than medium lengths. The dimensions are statistically orthogonal (r=0.193, p=0.084), with constraint effects 2.9x larger than semantic effects. Experimental validation via RLHF ablation confirms our constraint salience hypothesis: removing "helpfulness" signals improves CC by 598% on average (71/72 trials, p&lt;0.001), with 79% achieving perfect compliance. This demonstrates that RLHF-trained helpfulness behaviors are the dominant cause of constraint violations at medium compression. Reasoning models outperform efficient models by 27.5% (Cohen's d=0.96). Our findings reveal a fundamental tension between RLHF alignment and instruction-following, providing actionable guidelines for improving deployed systems.