Instruction Following

Speech Large Language Models (LLMs) that understand and follow instructions in many languages are useful for real-world interaction, but are difficult to train with supervised fine-tuning, requiring large, task-specific speech corpora. While recent distillation-based approaches train performant English-only Speech LLMs using only annotated ASR data by aligning text and speech using only a lightweight projector, these models under-perform when scaled to multilingual settings due to language interference in the shared projector. We address this by introducing language-aware distillation using a query bank and a gating network that selects or mixes query tokens using a Q-Former projector. Our approach shows gains of 14% over matched multilingual distillation baselines on instruction following. We further synthesize Audio-MLQA, a multilingual spoken QA benchmark built on MLQA with high-quality TTS questions. Our best model improves over existing Speech LLM baselines by 32% on Audio-MLQA.

Large Language Models (LLMs), despite advances in instruction tuning, often fail to follow complex user instructions. Activation steering techniques aim to mitigate this by manipulating model internals, but have a potential risk of oversteering, where excessive emphasis on the instruction degrades task accuracy and overall text quality. To address this, we introduce DIRECTER (Dynamic rejection steering), a novel steering method that dynamically modulates steering strength by scaling the KV cache without extra dataset. DIRECTER couples steering with a plausibility-guided decoding loop, which adaptively adjusts steering strength at each step by comparing the steered output distribution to the original. If the steered output is deemed implausible, steering strength is progressively weakened. This strength modulation is guided by a lightweight, one-time attention sensitivity analysis that ranks layers by their influence on model representations. Extensive evaluations show that DIRECTER significantly enhances instruction-following capabilities across diverse benchmarks, improving accuracy by up to 6.5% over baselines without the common trade-offs in generation quality or task fidelity. The proposed dynamic, plausibility-guided control during activation steering further demonstrates its potential as a general mechanism for mitigating oversteering that is compatible with existing baselines.

Instruction following is critical for LLMs deployed in enterprise and API-driven settings, where strict adherence to output formats, content constraints, and procedural requirements is essential for enabling reliable LLM-assisted workflows. However, existing instruction following benchmarks predominantly evaluate natural language generation constraints that reflect the needs of chat assistants rather than enterprise users. To bridge this gap, we introduce FireBench, an LLM instruction following benchmark grounded in real-world enterprise and API usage patterns. FireBench evaluates six core capability dimensions across diverse applications including information extraction, customer support, and coding agents, comprising over 2,400 samples. We evaluate 11 LLMs and present key findings on their instruction following behavior in enterprise scenarios. We open-source FireBench at fire-bench.com to help users assess model suitability, support model developers in diagnosing performance, and invite community contributions.

Instruction-following is a foundational capability of large language models (LLMs), with its improvement hinging on scalable and accurate feedback from judge models. However, the reliability of current judge models in instruction-following remains underexplored due to several deficiencies of existing meta-evaluation benchmarks, such as their insufficient data coverage and oversimplified pairwise evaluation paradigms that misalign with model optimization scenarios. To this end, we propose IF-RewardBench, a comprehensive meta-evaluation benchmark for instruction-following that covers diverse instruction and constraint types. For each instruction, we construct a preference graph containing all pairwise preferences among multiple responses based on instruction-following quality. This design enables a listwise evaluation paradigm that assesses the capabilities of judge models to rank multiple responses, which is essential in guiding model alignment. Extensive experiments on IF-RewardBench reveal significant deficiencies in current judge models and demonstrate that our benchmark achieves a stronger positive correlation with downstream task performance compared to existing benchmarks. Our codes and data are available at https://github.com/thu-coai/IF-RewardBench.

Small open-source language models are gaining attention for healthcare applications in low-resource settings where cloud infrastructure and GPU hardware may be unavailable. However, the reliability of these models under different phrasings of the same clinical question remains poorly understood. We evaluate five open-source models (Gemma 2 2B, Phi-3 Mini 3.8B, Llama 3.2 3B, Mistral 7B, and Meditron-7B, a domain-pretrained model without instruction tuning) across three clinical question answering datasets (MedQA, MedMCQA, and PubMedQA) using five prompt styles: original, formal, simplified, roleplay, and direct. Model behavior is evaluated using consistency scores, accuracy, and instruction-following failure rates. All experiments were conducted locally on consumer CPU hardware without fine-tuning. Consistency and accuracy were largely independent across models. Gemma 2 achieved the highest consistency (0.845-0.888) but the lowest accuracy (33.0-43.5%), while Llama 3.2 showed moderate consistency (0.774-0.807) alongside the highest accuracy (49.0-65.0%). Roleplay prompts consistently reduced accuracy across all models, with Phi-3 Mini dropping 21.5 percentage points on MedQA. Meditron-7B exhibited near-complete instruction-following failure on PubMedQA (99.0% UNKNOWN rate), indicating that domain pretraining alone is insufficient for structured clinical question answering. These findings show that high consistency does not imply correctness: models can be reliably wrong, a dangerous failure mode in clinical AI. Llama 3.2 demonstrated the strongest balance of accuracy and reliability for low-resource deployment. Safe clinical AI requires joint evaluation of consistency, accuracy, and instruction adherence.

The fast-growing demands in using Large Language Models (LLMs) to tackle complex multi-step data science tasks create an emergent need for accurate benchmarking. There are two major gaps in existing benchmarks: (i) the lack of standardized, process-aware evaluation that captures instruction adherence and process fidelity, and (ii) the scarcity of accurately labeled training data. To bridge these gaps, we introduce DARE-bench, a benchmark designed for machine learning modeling and data science instruction following. Unlike many existing benchmarks that rely on human- or model-based judges, all tasks in DARE-bench have verifiable ground truth, ensuring objective and reproducible evaluation. To cover a broad range of tasks and support agentic tools, DARE-bench consists of 6,300 Kaggle-derived tasks and provides both large-scale training data and evaluation sets. Extensive evaluations show that even highly capable models such as gpt-o4-mini struggle to achieve good performance, especially in machine learning modeling tasks. Using DARE-bench training tasks for fine-tuning can substantially improve model performance. For example, supervised fine-tuning boosts Qwen3-32B's accuracy by 1.83x and reinforcement learning boosts Qwen3-4B's accuracy by more than 8x. These significant improvements verify the importance of DARE-bench both as an accurate evaluation benchmark and critical training data.

Large reasoning models (LRMs) produce reasoning traces (RTs) that often contain sensitive information. These leaky thoughts are difficult to control and frequently violate explicit privacy directives. Because RTs can be exposed through prompt injection attacks, this becomes a direct privacy risk to the user. We approach this as a controllability problem: since privacy directives are themselves instructions, improving instruction-following (IF) within the RT provides a direct path to reducing privacy leaks. To this end, we introduce an SFT dataset that teaches models to follow general instructions throughout their reasoning process, and propose Staged Decoding, a simple decoding strategy that decouples RT and answer generation using separate LoRA adapters to maximize IF of each component. We evaluate our approach on six models from two families (1.7B-14B parameters), across two IF benchmarks and two privacy benchmarks. Our method yields substantial improvements, with gains of up to 20.9 points in IF and 51.9 percentage points on privacy benchmarks, though these can come at the cost of task utility due to the trade-off between reasoning performance and IF. Our results show that improving IF in LRMs can significantly enhance privacy, suggesting a promising direction for future privacy-aware LRMs. Our code is available at https://github.com/UKPLab/arxiv2026-controllable-reasoning-models.

Large reasoning models (LRMs) excel at a long chain of reasoning but often fail to faithfully follow instructions regarding output format, constraints, or specific requirements. We investigate whether this gap can be closed by integrating an instruction-tuned model (ITM) into an LRM. Analyzing their differences in parameter space, namely task vectors, we find that their principal subspaces are nearly orthogonal across key modules, suggesting a lightweight merging with minimal interference. However, we also demonstrate that naive merges are fragile because they overlook the output format mismatch between LRMs (with explicit thinking and response segments) and ITMs (answers-only). We introduce RAIN-Merging (Reasoning-Aware Instruction-attention guided Null-space projection Merging), a gradient-free method that integrates instruction following while preserving thinking format and reasoning performance. First, with a small reasoning calibration set, we project the ITM task vector onto the null space of forward features at thinking special tokens, which preserves the LRM's structured reasoning mechanisms. Second, using a small instruction calibration set, we estimate instruction attention to derive module-specific scaling that amplifies instruction-relevant components and suppresses leakage. Across four instruction-following benchmarks and nine reasoning & general capability benchmarks, RAIN-Merging substantially improves instruction adherence while maintaining reasoning quality. The gains are consistent across model scales and architectures, translating to improved performance in agent settings.

Instruction-following benchmarks remain predominantly English-centric, leaving a critical evaluation gap for the hundreds of millions of Indic language speakers. We introduce IndicIFEval, a benchmark evaluating constrained generation of LLMs across 14 Indic languages using automatically verifiable, rule-based instructions. It comprises around 800 human-verified examples per language spread across two complementary subsets: IndicIFEval-Ground, translated prompts from IFEval (Zhou et al., 2023) carefully localized for Indic contexts, and IndicIFEval-Ground, synthetically generated instructions grounded in native Indic content. We conduct a comprehensive evaluation of major open-weight and proprietary models spanning both reasoning and non-reasoning models. While models maintain strong adherence to formatting constraints, they struggle significantly with lexical and cross-lingual tasks -- and despite progress in high-resource languages, instruction-following across the broader Indic family lags significantly behind English. We release IndicIFEval and its evaluation scripts to support progress on multilingual constrained generation (http://github.com/ai4bharat/IndicIFEval).

Despite the strong performance of Large Language Models (LLMs) on complex instruction-following tasks, precise control of output length remains a persistent challenge. Existing methods primarily attempt to enforce length constraints by externally imposing length signals or optimization objectives, while largely overlooking the underlying limitation: the model's intrinsic deficit in length cognition. To address this, we propose LARFT (Length-Aware Reinforcement Fine-Tuning), a training framework that aligns the model's length cognition with its action. Specifically, LARFT integrates length-oriented reinforcement learning with a hindsight length awareness. By transforming on-policy data into hindsight self-awareness tasks where the model learns to identify the actual length of its own generation, LARFT jointly optimizes the model's internal representation of length information and refines its policy to satisfy length constraints, thereby achieving precise and reliable length instruction following. Extensive experiments across four base models demonstrate that LARFT outperforms existing baselines, achieving an average improvement of +20.92 points across three length instruction following benchmarks with only a marginal decline of -1.45 points on four general capability benchmarks.

It is not only what we ask large language models (LLMs) to do that matters, but also how we ask them. Phrases like This is urgent'' or As your supervisor'' can shift model behavior without altering task content. We study this effect as pragmatic framing, contextual cues that shape directive interpretation rather than task specification. While prior work exploits such cues for prompt optimization or probes them as security vulnerabilities, pragmatic framing itself has received comparatively little attention as a target of controlled measurement in instruction following. To support its systematic study as a measurable property, we introduce a framework that combines three components: directive-framing decomposition separating framing context from task specification; a taxonomy organizing 400 instantiations of framing into 13 strategies across 4 mechanism clusters; and priority-based measurement that quantifies influence through observable shifts in directive prioritization. Evaluating five open-weight LLMs across different families and scales, we find that pragmatic framing produces systematic shifts in directive prioritization, and the effectiveness ranking of different strategies proves highly consistent across models. This reveals that susceptibility to pragmatic framing is a structured behavioral property of instruction-tuned systems. Measuring this susceptibility is a prerequisite for any deliberate response to it, and this work provides the framework to do so.

Large language models often lose track of persistent instructions and relevant information as multi-turn conversations grow. We study this cumulative contextual decay through three related failure modes: attention pollution, dilution, and drift. We propose REA(Role-aware Heuristic Episodic Attention), a context-management framework that assigns different persistence and representation policies to instructions and episodic interactions. Instructional Memory retains identified global constraints in a dedicated prefix. Episodic Memory preserves user inputs and compresses model replies, while heuristic retrieval selects raw text, compressed representations, or omission for each historical turn. On Long-MT-Bench+, REA improves the judge score from 6.32 to 7.36 on a 10-point scale, a 16.5% relative gain over the Vanilla baseline, and reduces average latency by 2.91x. Additional evaluations show aggregate gains on three backbones spanning 1.7B-7B parameters and on Chinese and English role-playing tasks. These results support role-aware context management as a practical approach to maintaining conversational continuity and instruction adherence.

As large language model (LLM) based systems take on high-stakes roles in real-world decision-making, they must reconcile competing instructions from multiple sources within a single prompt context. Enforcing an instruction hierarchy, where higher-level directives override lower-priority requests, is critical to the reliability and control of LLMs. In this work, we reframe instruction hierarchy resolution as a reasoning task. The model must first "think" about the relationship between a given user prompt and higher-priority instructions before generating a response. To enable this capability, we construct VerIH, a training dataset of constraint-following tasks with verifiable answers, comprising aligned and conflicting system-user instructions. We show that lightweight reinforcement learning with VerIH effectively transfers general reasoning capabilities of models to instruction prioritization. Our method leads to consistent improvements across multiple model families on both instruction following and instruction hierarchy benchmarks, achieving ~20% absolute improvement in conflict setups. Our method also leads to improved alignment to safety-critical scenarios beyond the training distribution, exhibiting increased robustness against jailbreak and prompt injection, reducing absolute attack success rates by up to 20%. Our results establish reasoning over instruction hierarchies as a practical mechanism for improving AI reliability, where targeted updates to system prompts produce predictable, controllable, and robust changes in model behavior.

Autonomous agents executing human instructions must operate reliably even when instructions are incomplete. While recent approaches improve detection of missing information, detection alone is insufficient: agents often proceed to execution even after recognizing underspecification, leading to incorrect or unsafe actions. We identify this failure as arising from a lack of coupling between detection and execution, and propose that reliable behavior requires enforcing missing information as a precondition for action. We instantiate this principle in Cocoreli, a modular architecture that represents task structure, tracks missing information, and blocks execution until required details are resolved through targeted clarification. In Cocoreli, detection and prevention are structurally coupled: detecting a missing parameter simultaneously blocks execution. We evaluate Cocoreli in a controlled construction environment isolating underspecification and sequential execution. Cocoreli blocks execution under unresolved specifications by construction, eliminating hallucinated actions. In contrast, chain-of-thought, prompt-chaining, and ReAct-style reasoning may still execute under incomplete specifications despite high detection rates. The same representation supports abstraction and reuse, and generalizes to API workflow tasks on ToolBench. These results show that reliable collaborative execution requires architectural enforcement, not just model capability

Large language models (LLMs) are increasingly expected to tackle complex tasks, driven by their expanding applications and users' growing proficiency in crafting sophisticated prompts. However, as the number of explicitly stated requirements increases (particularly more than 10 constraints), LLMs often struggle to accurately follow such complex instructions, which limits their applicability in complex real-world scenarios. To the best of our knowledge, existing datasets do not exceed 10 constraints per instance. To address this challenge, we propose RECAST, an efficient and scalable framework for synthesizing datasets where each example incorporates far more constraints than those in existing benchmarks, aiming to challenge and extend the boundaries of models' ability to follow complex instructions. These constraints are extracted from real-world prompt-response pairs to ensure practical relevance. Using this framework, we construct RECAST-30K, a large-scale, high-quality dataset comprising 30k instances spanning 19 constraint types. Experimental results demonstrate that models finetuned on RECAST-30K substantially improve in following complex instructions while maintaining their general capabilities without degradation. Moreover, RECAST enables automatic verification of constraint satisfaction via rule-based validators for quantitative constraints and LLM-based validators for qualitative ones; the verifiability provided by RECAST enables the design of reward functions for reinforcement learning, which further boosts model performance on complex and challenging tasks.