Experiments on ESCI-English show that free-generation and likelihood-based forced-choice evaluation can yield qualitatively different robustness conclusions even when valid-label generation is nearly perfect and average task performance is similar, and SFT does not uniformly reduce instruction sensitivity.
Abstract
Large language models can exhibit substantial performance variation across alternative formulations of the same task instruction, yet it remains unclear how conventional task-specific supervised fine-tuning (SFT) changes this instruction sensitivity. We study this question by evaluating fixed model checkpoints under multiple paraphrased instructions and defining instruction sensitivity as the standard deviation of task performance across them. We conduct a controlled scale analysis with Qwen3 models at 1.7B, 4B, and 8B on MS MARCO, together with targeted cross-family checks using Mistral-7B and Gemma-2-9B. Before SFT, instruction sensitivity decreases sharply with Qwen3 model scale. At 1.7B and 4B, SFT consistently reduces sensitivity across training instructions, with reductions of approximately 54--71%. At 8B, individual sensitivity changes are not statistically distinguishable from zero, but paired contrasts between training instructions are statistically reliable under query-level bootstrap analysis and have consistent directions across all three random seeds. Gemma-2-9B shows the same directional training-instruction contrast as Qwen3-8B, whereas Mistral-7B does not, suggesting that the strength of this effect also varies across models. Experiments on ESCI-English further show that free-generation and likelihood-based forced-choice evaluation can yield qualitatively different robustness conclusions even when valid-label generation is nearly perfect and average task performance is similar. Overall, SFT does not uniformly reduce instruction sensitivity: its robustness effect depends on the adaptation setting, while measured sensitivity can additionally depend on the prediction and scoring protocol.
IFHierBench is introduced, a hierarchical instruction-following benchmark of 600 prompts stratified across four constraint-tree depths and 35 distinct constraints, each prompt paired with a deterministic checker that verifies satisfaction at every scope.
This work proposes a measure of training data influence that does not require selecting a downstream task or validation set as the attribution target, and defines an example's influence by how much its gradient update reduces the squared distance to the final parameters of a given pretraining run.
Yuto Nishida, Hirokazu Kiyomaru, Yusuke Oda et al.· 0 citations
Natural-language requirements for program synthesis are often incomplete or ambiguous, yet large language models are commonly expected to generate code in a single pass. Prior clarification-based methods address this issue by asking follow-up questions when sampled candidate programs disagree, but fixed clarify-on-disagreement policies can overuse clarification and can also overtrust weak behavioral agreement. We present an adaptive routing framework for LLM-based program synthesis that treats clarification as an inference-time control decision. The framework augments a ClarifyGPT-style pipeline with execution-driven confidence estimation, semanticdifference analysis, and bounded candidate expansion, allowing the system to choose among direct generation, additional evidence gathering, and clarification. We evaluate the framework on MBPP, HumanEval, and extended-test variants using GPT-4.1 mini, Claude Haiku 4.5, and GPT-5.4 mini. Adaptive routing improves pass@1 accuracy by up to 7.60 percentage points over single-pass baselines. Compared with fixed-policy clarification, it preserves accuracy while reducing token usage by up to 57.2% for GPT-4.1 mini, and reallocates computation toward harder cases for Claude Haiku 4.5. These results suggest that clarification is most useful when triggered selectively based on execution evidence and semantic disagreement, even when ambiguity is observed indirectly through candidate behavior rather than through explicitly annotated ambiguous requirements.
Muhammad Ahmed, Edwar Tiu, Niyati Nikunj Kapadia et al.· Annual International Compute...· 0 citations
This work introduces a search-based approach that identifies and evolves a set of natural language transformation rules with strong downstream effects on coding performance, and proposes DUALFIX, a staged repair pipeline that combines the evolved transformation rules with execution-feedback repair, addressing both specification-level and implementation-level failures.
Amal Akli, Melissa Akli, Cedric Richter et al.· 0 citations
We study how instruction-tuned LLMs arbitrate direct conflicts between system and user instructions. We introduce a benchmark of 41 paired constraints with deterministic verifiers and evaluate eight models under matched baseline, conflict, and same-channel control conditions. Behaviourally, the models split into three regimes by System Authority Delta: hierarchy-respecting models use the system channel as an authority signal, anti-hierarchy models follow the system less often than their same-channel baseline predicts, and no-effect models show little channel sensitivity. Llama-3.1-8B is the strongest anti-hierarchy case in our suite, following the system in only 0.10 of conflict trials. We use this behavioural failure case to ask whether user-preferring arbitration reflects the absence of an internal conflictresolution signal. It does not: on Llama-3.1-8B, the conflict outcome is linearly decodable from residual-stream activations at 0.97 balanced accuracy, 17 percentage points above a metadata-only baseline, with analogous signals on Qwen2.5-7B and gpt-oss-20b. Steering with a layer-12 mean of four per-conflict logistic-regression directions raises genuine system compliance from 0.132 to 0.530, while directions selected mainly for pooled separability steer poorly. User-preferring conflict resolution can therefore coexist with a readable internal arbitration signal, and successful intervention depends on the geometry of the readout rather than probe accuracy alone
Enrique Balp-Straffon, Chih-Hao Hsu, Rushiraj Gadhvi et al.· 0 citations
Accuracy changes after language-model self-revision are usually interpreted as changes in reasoning. We show this can fail at the answer-extraction boundary, and test the failure causally rather than only observationally. Across Qwen3.5 (0.8B-9B), Gemma-4-12B, and two frontier models via API (Tencent Hy3, Nvidia Nemotron-3-Ultra-550B) in 29 primary cells plus a frontier arm, we decompose the always-revise accuracy shift into a content margin (both answers parseable) and format-recovery/loss margins (parseability changes). On 12 cells with meaningful unparseable-answer rates, format effects exceed content effects (Wilcoxon p=1.7e-3). To test this causally, we force already-generated reasoning through grammar-constrained decoding so every answer is parseable by construction: across 14 cells this closes a median 71% of the gap between the naive total effect and the content-margin estimate, with two cells converging exactly and a residual on the two largest-effect cells reported rather than dismissed. A clustered model confirms floor-scale (0.8B/2B) models have far higher odds of content-level change and harm than capable-scale models (p<1e-7). Replicating a cited confidence-gating protocol verbatim on Qwen3.5 does not reproduce its reported gain and shows the same near-zero content margin. A frontier check on much larger models shows format-dominance intensifying with scale: content margin is exactly zero in all 5 cells despite total effects up to +0.275, though this arm is lower-powered. The calibration-floor criterion on the content margin reveals a squeeze: floor-scale cells have headroom but insufficient signal, capable-scale cells have signal but little headroom; only one cell is marginally viable, with negligible sealed-holdout gain. Content is a minority share of what the field has measured as self-correction. We release the instrument, code, and derived results.