Overview

Instruction tuning studies how to turn pretrained models into systems that respond usefully to instructions, often through supervised fine-tuning on instruction-response pairs and sometimes preference optimization or reinforcement learning. Research develops ways to select, balance, and generate training data, including methods for multilingual, domain-specific, and multimodal tasks. Evaluations use instruction-following tests such as IFEval alongside reasoning, coding, dialogue, and specialized-domain benchmarks. A central challenge is that gains can depend on task and format: tuning may improve instruction compliance while harming other capabilities, and reliable generalization beyond training examples remains difficult.

Latest papers

89 papers16 papers in the last four weeks, up 129% on the four weeks before. 0.1% of all new papers.

Nov 8, 2026

Self-Indexing Latent Memories for Access-Aware Continual Instruction Tuning of LLMs

Continual instruction tuning often preserves old skills with task-private modules. Inference still requires selecting the right module from the prompt alone. We introduce Task-Isolated Latent Adapter Continual Training (TILA-CT), a replay-free method that makes each task module self-indexing. Each task receives a latent memory and adapter bank in a frozen LLM. During adaptation, the memory gathers causal-prefix information and injects task-conditioned context. For access, its persistent state provides the routing prototype used by a shared prompt-only router. On the 8-task CT-Pipeline stream, TILA-CT achieves 65.19 acquisition and 63.98 final macro score, with 1.38 points of average forgetting. A middle-reversed order gives 65.38 acquisition and 64.17 final macro score. Across 6 backbones, final scores remain within 1.5 points of acquisition. Ablations support memory-derived routing keys and causal gathering. Final macro routing accuracy is 93.9%, and oracle routing recovers 1.21 final macro points. These results separate preserved task computation from learned access to it.
Nov 6, 2026

Continual Instruction Tuning for LLM-based Management of Continuum Resources

Large language models (LLMs) are increasingly explored for resource management across the Cloud-Edge continuum. However, embedding LLMs directly inside latency-sensitive control loops places inference latency, execution jitter, and hallucination risks on the critical path. This study argues for a different use of LLMs: a continually instruction-tuned intent compiler that runs once per management intent or declared context change, above a fast, deterministic tactical controller that handles all low-latency runtime actions. We formulate intent compilation as a continual instruction tuning problem, because management intents arrive sequentially and accumulate while infrastructure and workload contexts drift. Unlike recent LLM-based management frameworks, our design scores the model's outputs by execution, so that catastrophic forgetting can be measured in the executed objective of the underlying controller. On a continuum simulator with a 2B backbone and parameter-efficient adapters, preliminary results show that naive sequential tuning forgets earlier intents in the executed objective, that replay of stored examples nearly prevents this forgetting, and that the tuned compact compiler performs on par with simpler translators while being more robust than a grammar parser to rephrased intents.
Nov 2, 2026

TrafficInstruct: A Multi-Task Network Traffic Dataset and Benchmark for Instruction Tuning

Instruction tuning for multi-task network traffic analysis requires data with clearly defined sources, supervision targets, and evaluation conditions. We introduce TrafficInstruct, a newly collected dataset comprising raw packet captures and 55,000 structured supervised fine-tuning records across ten tasks: malware traffic detection, botnet detection, malicious DNS-over-HTTPS detection, Web attack detection, advanced persistent threat detection, encrypted VPN detection, Tor behavior detection, encrypted application/service classification, website fingerprinting, and application classification under concept drift. An automated workflow captures traffic from controlled software executions, attack experiments, and application activities. Labels are supported by task-specific collection evidence, including process or socket attribution, activity records, and execution outcomes. Filtering, sanitization, and task-specific transformations produce instructions, traffic observations, and structured answers. We verify sample sources and partition the dataset by complete collection sessions, keeping all samples from the same session within one partition. The benchmark defines allowed inputs, deterministic label mappings, fixed training/validation/test partitions, and task-wise scoring rules. We evaluate majority-class, random-forest, and ET-BERT baselines, an unadapted 27B language model, and independently trained task-specific LoRA adapters on a shared frozen base. Adapters are selected using validation data and evaluated on the corresponding test sets. Task-specific adaptation improves classification performance, supporting instruction tuning across diverse traffic-analysis tasks.
Oct 26, 2026

3, 7, 14, 17, 42: Instruction Tuning Installs and Sharpens Number-Choice Heuristics in Language Models

We measure, without sampling error, the probability a language model assigns to each integer from 1 to N when asked for a random number. For a fixed prompt the model defines one probability vector, which we compute by summing over each integer's scored spellings rather than estimate by sampling, removing the noise that limits earlier studies [1, 4]. We evaluate 18 open-weight model lines at five ranges up to N = 100, under 29 decoding configurations, on two inference stacks; 15 of them have matched base and instruction-tuned checkpoints that both give usable answers. Instruction tuning makes the distribution less uniform in all 15 pairs. It installs one heuristic, round-number avoidance, and sharpens another, endpoint avoidance. Base checkpoints do not avoid round numbers once the endpoints are removed from those sets; tuned checkpoints avoid them, multiples of 10 almost completely. Base checkpoints return the endpoints 1 and N at 0.30 to 0.60 of the interior rate; tuned checkpoints almost never return them. A preference for odd numbers is consistent across tuned checkpoints only at N = 100, where base checkpoints also show it weakly. Endpoint avoidance and the odd preference fit the account Kubovy and Psotka [6] proposed for human number choice; round-number avoidance is not part of that account, and the midpoint avoidance it predicts is essentially absent, with one slight exception at N = 100. Tuning does not install a new favourite integer shared across models: above N = 10 no modal integer is shared by more than 42 percent of tuned conditions, and P (17) rises in 10 of 15 pairs and falls in 5. The concentration is present in the model's own distribution: at temperature 1.0, nucleus truncation at top-p 0.95 adds a median of 0.035 to the total variation distance.
Oct 26, 2026

The Geometry of Emotion Representations Across Scale, Depth, and Instruction Tuning

We measure the emotion geometry of six LLMs spanning Google's Gemma 2 family. Using an extraction pipeline established by Sofroniew et al. (2026) on Claude Sonnet 4.5 and replicated by Hollowell (2026) on Gemma 2 27B base, we compare emotion geometry across base and instruction-tuned variants of Gemma 2 2B, 9B, and 27B. First, all six model cells exhibit the core emotion geometry. A linear map fit between two cells places held-out emotions nearest their own vectors in the target model. Across all fifteen model pairs, fit in one direction each, this failed exactly once, and the median rank was 1 in every pair. Linear maps refit on cleaned embedding-table vectors recover every held-out emotion at rank 1 in all thirty ordered pairs. Second, scale spreads the geometry's variance across more directions without weakening its tie to valence. The leading component's share of variance falls from 31.6 percent (2B) to 29.8 (9B) to 26.7 (27B) in the base models and from 31.9 (2B) to 29.2 (9B) to 27.3 (27B) in the tuned ones, while its correlation with human valence ratings stays between .71 and .78 in all six models. Third, we project each emotion vector onto the output vocabulary and ask at what depth the vector's own emotion word enters the top 10,000 of the 256,000 tokens. In the 2B and 9B models the median emotion word is there from layer 0, while in the 27B models it begins near chance, stays outside the top 10,000 for half the network, and enters at layer 23 in one step. Fourth, tuning leaves every registered analysis-layer measure within its floor. At every size, valence correlation rises and arousal correlation falls by amounts within their floors, the latter against prediction. PC2's variance share also falls, with a pooled decrease significant after correction. Mid-late layer stability exceeds its floor and traces to base 27B's final layer. Code, data, and audit reports are released.
Oct 26, 2026

Improving Instruction Hierarchy by Learning from Goal-Goal Conflicts

The recent incident in which OpenAI agents autonomously compromised Hugging Face's infrastructure illustrates the importance of instruction hierarchy (IH): an LLM should follow lower-priority instructions, such as user requests, only when they do not conflict with higher-priority instructions, such as system policies. IH can be effectively improved by fine-tuning LLMs on simulated examples of instruction conflicts, whose diversity is key to teaching LLMs to properly handle conflicts at test time. However, our decomposition of a task into goals and constraints reveals a conflict type -- goal-goal conflicts -- that is absent from existing IH training sets from OpenAI and NVIDIA. We show the importance of this conflict type: fine-tuning Qwen3.5-9B and Gemma4-12B-it on examples of this conflict type significantly improves IH on a wide range of real-world tasks unseen in training; adding these examples to existing IH training sets for fine-tuning also gives better IH adherence. The IH gains come with trivial mean utility changes on no-conflict tasks across eight benchmarks. Our work advocates including this new conflict type in IH training, even if these examples do not reflect typical LLM use.
Oct 26, 2026

DocTalk: Enhancing Medical Knowledge Retrieval via Fine-tuned Large Language Models with Chain-of-Thought Reasoning for Robust Clinical Decision Support

Objective: The rapid advancement of large language models (LLMs) in healthcare has opened new possibilities for AI-assisted clinical decision-making; however, safe and effective deployment remains challenging because of hallucination and domain misalignment. This study introduces DocTalk, a doctor-assistive question-answering (QA) system designed to generate concise, chart-grounded responses from de-identified discharge summaries. Methods: We developed a high-fidelity data generation pipeline that produces 500 clinically relevant QA pairs through sentence-level decomposition, chain-of-thought-guided question formulation, and expert review. Six open-source LLMs, including three general-purpose and three biomedical models, were evaluated before and after parameter-efficient instruction tuning using Low-Rank Adaptation (LoRA). The framework integrates prompt engineering, in-context learning, and medically aligned evaluation to support factual and interpretable responses. Results: On DocTalk-QA, instruction tuning produced consistent improvements across lexical, semantic, and clinical fidelity metrics. The best-performing tuned model, Llama-3-Med42-8B, achieved ROUGE-L = 0.40, BERTScore-F1 = 0.76, and UMLS-F1 = 0.49. To examine external generalization beyond the DocTalk-QA dataset, the two best-performing models, Llama-3.1-8B-Instruct and Llama-3-Med42-8B, were evaluated on PubMedQA before and after instruction tuning. Accuracy improved from 0.931 to 0.977 for Llama-3.1-8B-Instruct and from 0.935 to 0.943 for Llama-3-Med42-8B. Conclusion: DocTalk was deployed as a secure web-based platform for interactive EHR querying. These findings suggest that clinically grounded instruction tuning can improve medical QA performance and support practical clinical decision support.
Oct 25, 2026

"High-Quality" Data for Early-Stage Instruction Tuning: An Empirical Study of Data Quality Metrics

"Data quality" is widely invoked in instruction tuning, but the term is often underspecified. We study it as an operational property by asking which simple example-level metrics, drawn from prior work on data selection, text quality, and instruction tuning, best predict training utility under a fixed finetuning budget. For each metric, we sort instruction-response examples by score, partition them into low, middle, and high bins with roughly equal token mass, finetune separately on each bin, and compare validation perplexity and LLM-as-judge scores for the models' generated answers as a function of cumulative training tokens. Our results suggest that answer-dense supervision, as captured by response ratio, is a stronger and more portable predictor of instruction-tuning utility than commonly used alternative proxies.
Oct 23, 2026

The Language of Instruction: Genre and Generative AI

This article argues that instruction-fine-tuned language models should be understood as implementing a genre, what is proposed as the instructional-conversational genre. This genre is an amalgamation of discursive language from prior conversations between generative models and humans recursively run through these massive neural networks. It is instantiated in applications through instruction fine-tuning, the inclusion of a pretext known as a system prompt, and the formalization of inputs in mark-up languages characterized by the use of special tokens and role assignment in turntaking structures.
Oct 21, 2026

Task-Preserving Multilingual Instruction Tuning for Open-Ended QA in WMT 2026 MIST

This paper describes our submission to the WMT 2026 Multilingual Instruction Shared Task. We adapt Qwen3-8B with LoRA using organiser-provided data, EuroAlpaca, a multilingual instruction dataset created through task-preserving localisation, and CrossEuroAlpaca, a cross-lingual augmentation that assigns different source and target languages to instructions, contexts and responses. Our system targets open-ended question answering (QA-OEG) in ten shared-task languages represented in the auxiliary data. Under matched decoding, adaptation improves the mean Gemma-4 judge score on target-language QA-OEG by 7.8% relative to the base model. On an additional evaluation using the Aya Evaluation Suite across all shared-task languages, it yields relative gains of 41.5% in macro-averaged ROUGE-L and 2.9% in BERTScore-F1, with larger improvements in target languages. However, these gains are task- and language-specific. Gemma-4 scores improve for QA-OEG and summarisation in the target languages but decline for context-grounded QA there and across all three tasks in the non-target languages.
Oct 20, 2026

The Instruction Tax on Factual Recall: Base Models Can Beat Their Instruction-Tuned Counterparts

We present our solution to the AKBC LM-KBC 2026 Shared Task, where we achieved a score of 0.6836 on the official leaderboard, more than double the baseline of 0.2964. This task requires generating all possible objects for subject-relation pairs using a language model with up to 32 billion parameters, with no retrieval or fine-tuning involved. We show that the prompting methodology is as important as the model architecture: while the instruction-tuned model version achieves 0.5681 with its chat interface, the base model achieves 0.6836 when prompted using text completion. Thus, instruction tuning decreases factual recall; we refer to this effect as the instruction tax, which is caused by the model's weights rather than the prompting method. Our final solution is fairly straightforward: we prompt the base model with some examples from the training data, generate 64--256 outputs, and retain only the repeated ones. Two specialized methods are then applied for each relation: one improves the recall of large answer sets, and the other verifies a person's vitality before predicting their place of death.
Oct 16, 2026

LongForge: Measuring and Improving Long-Context Instruction Data via Evidence-Assertion Reasoning Graphs

Existing quality control for long-context instruction-tuning data is a one-shot judgment inside the generation pipeline that leaves only a scalar, revealing neither which document parts an answer uses nor how many reasoning steps it writes down; the data can thus be neither audited nor controlled. We formalize data quality as a property of the alignment structure among document, question, and answer, and build LongForge, which pairs an analyzer that measures long-context data with a synthesizer that generates it. The analyzer reverse-reconstructs each sample into an evidence-assertion reasoning graph and programmatically computes four dimensions on it: breadth, depth, correctness, and completeness. Every reading traces back to specific evidence spans and reasoning steps, making data auditable. Measuring seven mainstream datasets reveals two previously unquantified defects: breadth barely grows with document length, and depth is generally low, the two being nearly orthogonal. The synthesizer then embeds the analyzer into generation, gating on correctness and completeness and deferring breadth and depth to a training-time mixture over measured values, which makes the training distribution controllable. On two backbones, its two mixtures, LongForge-1 and LongForge-N, outperform the strongest baseline by 2.2-4.7 points on average over six long-context benchmarks, and in a controlled comparison that holds the gate and the candidate pool fixed, raising breadth and depth alone is worth a further 1.16 points, making the structural distribution of training data a lever on downstream quality rather than only a description of it.
Oct 15, 2026

AutoSFT: An Agentic System for Automated Data Research in Supervised Fine-Tuning

Language model agents increasingly automate post-training, but when given the full training stack they converge on supervised fine-tuning and spend most effort revising training data. Existing systems give the agent one kind of control, selecting rows from a fixed pool or calling a generator, whereas a human data engineer adjusts proportions, prunes, repairs, replaces and synthesizes the data. We design a data action space for these changes and ask whether an automated system can make the full range rather than one of them, and what structure that requires. We first run this loop with a single agent, and it reaches only one action, adjusting category proportions, while its context fills with execution output and is discarded mid-search. We therefore introduce AutoSFT, a multi-agent system where a meta agent separates execution from analysis, gathers proposals from independent angles, and calls a synthesis subagent for capabilities the pool cannot cover, with the model, the trainer, and the budget fixed within each comparison. Adding each component in turn helps: the executor halves context loss, the analysis subagents lift the best IFEval by +6.5+6.5, and synthesis adds +3.4+3.4 IFEval and +2.9+2.9 IFBench. The same loop transfers to code, where analysis adds +3.7+3.7 MBPP+ over a single agent that does not improve. Every result comes from runs that retrain from the base model each round, so each component's effect is individually attributable. Our findings suggest that automating SFT data research depends more on covering diverse data actions than on a stronger single action.
Oct 15, 2026

Multi-SFT: Permissively-Licensed Multilingual Instruction-Following

Supervised fine-tuning datasets remain overwhelmingly English-centric and commercially restricted, as their content is frequently generated by proprietary models whose terms of service forbid training competing systems. We introduce Multi-SFT, a commercially permissive collection of 502,818 instruction-response pairs across five European languages, built via translation and regeneration pipelines for tasks where translation is insufficient. Human evaluation yields inter-annotator agreement of at least 0.87 across all datasets and languages. Fine-tuning nine models from three different families (EuroLLM, OLMo 2, Qwen3; 1B-32B) and evaluating them on 13 benchmarks shows that Multi-SFT alone is competitive with larger augmented mixes and official instruction-tuned releases, with decisive advantages on safety tasks.
Oct 12, 2026

The Format is the Filter: Scoring Format Fixes What an LLM Judge Can Express

When an LLM judge filters instruction-tuning data, its scoring format (binary verdict, 1--5 rating, 0--100 score, or comparison) is rarely justified or ablated. Holding the candidate pool, judge prompt, decoding and selection budget fixed, we score 15,406 UltraFeedback candidates under five formats with three judges, curate, fine-tune 66 models, and evaluate them with four judges. A 0--100 scale is worth only 7--12 equally used levels, two to three times a 1--5 scale rather than twenty, and a binary verdict leaves the top score tied in 77--78% of prompts, so most selections are made at random. These properties are set by the format almost independently of the judge. Kendall's τbτ_b hides them: it ranks binary best while τaτ_a and top-1 agreement rank it last, because it discounts the ties a selector must break at random. Both comparative formats inherit a slot bias whose sign differs across judges. Downstream, curated data beat random selection under every evaluation judge, robust to length and markdown adjustment, but no ordering among formats is detectable. We propose four quantities that judge-based curation pipelines should report.
Oct 12, 2026

PitchIn: Natural-Language Vocal Coaching from Unstructured Community Feedback

Effective vocal coaching connects performance issues with corrective advice, but recordings paired with expert diagnoses and advice are scarce. Online singing communities provide a potential source of supervision, but observations and advice are scattered across unstructured comments. We introduce PitchIn, audio-language models for natural-language vocal coaching, and the PeerCoach Dataset: 14,445 instruction examples covering Focused QA, Overall Feedback, and Recognition QA. We compare target combinations on Qwen2.5-Omni and train PitchIn on Focused QA and Overall Feedback using Qwen2.5-Omni and MOSS-Music. On VocalCoachBench, the proportion of requested diagnosis-correction pairs matching an expert issue with valid advice rises from 21.3% to 61.0% for Qwen2.5-Omni and from 31.4% to 51.2% for MOSS-Music. Direct training on source comments yields only 19.8% for Qwen2.5-Omni. For Qwen, Focused QA alone reaches 65.6%, but adding Overall Feedback improves alignment with expert reviews and human preference. Both models improve expert-review alignment over pretrained baselines and are significantly preferred for usefulness (92.0% and 67.3%, respectively). Demo examples are available at https://hayeonbang.github.io/PitchIn.
Oct 2, 2026

Enhancing Biomedical Named Entity Recognition via Multiple Programming Languages Instruction Tuning and Ensemble Method

Instruction tuning has become a common paradigm for applying large language models (LLMs) to biomedical named entity recognition (BioNER). However, existing instruction-tuning approaches still face two key challenges. First, conventional natural-language instructions typically serialize BioNER annotations as flat textual outputs, providing limited structural constraints for typed entity extraction. Second, high-quality biomedical annotations are limited, and learning from a single serialized output form may restrict structural diversity and reduce model robustness. Although external biomedical knowledge can be introduced to alleviate data scarcity, it often requires costly resource construction. To address these challenges, we propose MITE, a Multiple Programming Languages Instruction Tuning and Ensemble method for BioNER. MITE reformulates BioNER as a structure-to-structure generation task by representing both instructions and entity outputs in code-formatted representations. Specifically, each training instance is transformed into multiple programming-language formats, including Python, C++, and Java, while preserving the same underlying entity semantics. These language-specific representations provide structurally diverse supervision without requiring external biomedical knowledge or additional annotations. During inference, MITE aggregates predictions from different code formats through an entity-level voting strategy, reducing language-specific prediction variance and improving robustness. Experiments on three widely used BioNER datasets demonstrate that MITE achieves competitive performance against representative BERT-based and LLM-based baselines, attaining the best results on two benchmarks. Ablation studies further validate the effectiveness of multi-language representations.
Sep 30, 2026

Coverage Before Control: Route-Instruction Grounding and Steering for Controllable Retrosynthesis

Single-step retrosynthesis models are commonly evaluated by their ability to recover recorded reactions. In practice, chemists may need to choose among several precursor sets for the same product, for example to preserve a particular motif. Recovering a recorded answer alone does not establish this ability to follow a preference. Satisfying such requests requires both coverage of relevant alternatives and control over which alternatives are favored. We introduce Route-Instruction Grounding and Steering (RIGS), a two-stage framework for instruction-conditioned retrosynthesis. Stage A trains a language projector, teaching it which alternatives an instruction favors or discourages. Stage B uses the projector learned in Stage A to steer a frozen generative model through lightweight residual adapters. We construct nested one-to-many training supports by pairing each product with increasing numbers of candidate precursor sets. Extensive experiments demonstrate that broader support helps the model generate a wider range of alternatives, and RIGS can learn to guide generation according to instructions. The relationship between coverage and control is consistent across model scales but non-monotone.
Sep 30, 2026

A helps B while B hurts A: directed transfer in instruction-tuning mixture

Adapting a language model to a specialized corpus means choosing which instruction-tuning tasks to train on under a fixed budget, and testing one choice costs a fine-tuning run. Common heuristics add more source tasks or pick sources similar to the target. The first assumes transfer is never negative; the second, that it is symmetric. We show that both assumptions fail: task AA can help task BB while BB hurts AA, so helpfulness is a signed property of ordered source--target pairs. We introduce the transfer map, a signed estimate of how much each source helps or hurts each held-out target. We fit the map in hundreds of fine-tuning runs on Qwen3 and Mistral models from 0.6B to 32B parameters, with all sources drawn from one corpus and no training examples from the target. The map predicts a held-out target's accuracy on unseen mixtures: recorded before those runs, its predictions have less than half the error of a mixture-agnostic baseline. The map is specific to its target and corpus but transfers across model scale: a mixture selected in advance at one size beats training on all source tasks at every other size we tested. Transfer is thus a property of the data. The map selects the tasks that help and drops the one that interferes: accuracy on the reasoning targets (causal explanation, multi-hop questions and methodological critique) rises by up to 14 percentage points over training on all source tasks.
Sep 29, 2026

CompOrca: Corpus-Scale Compliance Labelling of Instruction-Tuning Data

Studying how fine-tuning shapes refusal and noncompliance behaviour requires knowing which training examples refuse or otherwise fail to fulfil the request. Existing annotations cover evaluation sets, which are far smaller than training corpora. We present CompOrca, compliance labels for all 4,233,923 examples of the OpenOrca corpus. Every example was classified as compliant or noncompliant by five passes of an open-weight LLM judge (LongCat-2.0, 1.6T parameters). The corpus is released as unanimous compliance (94.75%), unanimous noncompliance (1.28%), and nonunanimous rows (3.97%), with the raw vote counts. A single pass flags 2.7-3.2% of the corpus as noncompliant, while only 1.28% is flagged by all five, so the most ambiguous rows can be filtered out. Against 450 human-annotated examples (150 annotated twice; human-human κ=0.93κ=0.93), the unanimous compliance and noncompliance labels are 97.3% and 86.7% precise. The noncompliance label is a high-precision subset of the corpus's noncompliance. Published refusal-detection methods recall between 0.4% and 94.1% of human-labelled noncompliance. We release the full corpus with its per-row labels and vote counts at https://huggingface.co/datasets/cemiu/CompOrca
Sep 26, 2026

ScopeIF: Improving Scope-Aware Precise Instruction-Following in Large Language Models via Graded Reward Modeling

Precise instruction-following is a fundamental ability of large language models (LLMs), requiring their outputs to strictly satisfy objective constraints in input instructions. In complex application scenarios, these constraints often possess diverse scopes that govern specific response segments rather than the entire output. However, existing optimization methods often neglect constraint scope during data construction and rely on binary per-constraint rewards, yielding limited data diversity and sparse supervision for complex constraints. To this end, we propose ScopeIF, a novel training framework for scope-aware precise instruction-following. We first introduce a unified schema that factorizes objective constraints into three decoupled dimensions: Scope, Target, and Range. Grounded in this schema, we construct ScopeInstruct, a large-scale instruction dataset with diverse scope-aware constraints, and combine tool-grounded verification with graded reward modeling to quantify the violation degree of each constraint, providing dense supervision for policy optimization. Extensive experiments demonstrate that ScopeIF consistently outperforms existing methods, particularly on complex scope-aware constraints, while preserving general capabilities. Notably, it enables optimized Qwen3-4B and 8B models to rival or surpass strong frontier models such as Gemini-2.5-Pro and DeepSeek-V3.2, establishing an effective paradigm for advancing scope-aware instruction-following. Our code and data are available at https://github.com/thu-coai/ScopeIF.
Sep 25, 2026

TemporalGraphLLM: Temporal Graph Neural Networks with Large Language Models for Dynamic Text-Attributed Graphs

Dynamic text-attributed graphs (DTAGs), where nodes, edges, and textual attributes evolve over time, are crucial in applications such as social networks, citation graphs, and knowledge graphs. However, existing approaches struggle to jointly model the temporal evolution of graph structures and the semantic richness of textual attributes. While Temporal Graph Neural Networks (TGNNs) capture evolving node relationships, they often lack contextual text reasoning. Conversely, Large Language Models (LLMs) excel in textual understanding but struggle with structured graph reasoning in temporal settings. To bridge this gap, we propose TemporalGraphLLM, a novel framework that can integrate any temporal GNN with an LLM for enhanced reasoning in DTAGs. Our approach fine-tunes LLMs using graph-time-aware instruction tuning and novel temporal GNNs injection to replace dedicated added tokens with graph embeddings. TemporalGraphLLM effectively leverages pretrained TGNNs within an LLM framework to achieve state-of-the-art performance on edge classification, link prediction, and edge-based text generation tasks. Extensive evaluation on real-world dynamic graph datasets demonstrates state-of-the-art performance. Our findings highlight the synergistic potential of LLMs and TGNNs, opening new directions for learning on evolving graphs.
Sep 22, 2026

Rewired or Gated? How Instruction Tuning Shapes Knowledge-Conflict Circuits in LLMs

In language models, the choice between believing the prompt and believing the weights is made by a handful of identifiable attention heads. Instruction tuning changes how models behave under conflict, but whether it rewires the underlying circuit or merely gates/reweights already present components, remains unknown. We provide the first mechanistic base-vs-instruct comparison of conflict-resolution circuits, across three families (Llama-3.2-3B, Qwen-2.5-3B, Gemma-3-4B). Five independent methods, node and edge attribution, superposition role analysis, causal ablation, and path patching, converge on gating, with the same heads, in the same late-layers, are found to be reweighted rather than replaced with a high node overlap (0.60-0.82). Behaviorally, tuning shifts models toward parametric memory, making instruct models reject a terse counterfactual context far more than base ones, the opposite of a naive user-following expectation. Yet this added skepticism is a factor of framing since it disappears when the same false claim is delivered as a coherent, evidential passage. The robustness that instruction tuning buys against terse injection is therefore real but narrow. More broadly, we believe that because the conflict circuit is preserved rather than rebuilt, interpretability and control tools calibrated on base models should transfer directly to their deployed instruct siblings.
Sep 11, 2026

Clustering-Based Balanced Sampling and Allocation with Data Parallelism for High-Performance Fine-Tuning

Instruction-tuning datasets for large language models (LLMs) are often large, redundant, and imbalanced, limiting efficient adaptation. Naive large-batch fine-tuning repeatedly includes overrepresented sample groups while weakly covering underrepresented but informative ones, especially under data parallelism (DP) across multiple GPUs. We propose CluSTER, a Cluster-aware balanced Sampling framework for Training Efficient data Reduction in DP instruction tuning. CluSTER curates a representative reduced dataset through gradient-space clustering and DP-aware balanced allocation, ensuring dual-level coverage across clusters and workers, while preserving the original data distribution by weighted update. As a result, CluSTER reduces redundant computation and improves training stability without compromising model quality. Across multiple instruction-tuning datasets, CluSTER reduces training time by up to 69.6% with almost no accuracy loss compared to prior sampling and data reduction methods. Code is available at https://github.com/kaist-dmlab/CluSTER.
Sep 8, 2026

Stabilizing Instruction Supervision for Instruct-TTS via Controllable Diversification and Drift Filtering

Instruct-TTS systems expand structured style labels into natural-language training instructions through LLM rewriting, yet we find that over 40% of unconstrained rewrites contain semantic drift that corrupts supervision and weakens generalization. We formalize this problem as instruction supervision instability and propose a data-centric stabilization recipe that jointly improves coverage and fidelity through three mechanisms: controllable instruction diversification for systematic expansion, LLM-based drift filtering for quality control, and attribute-aligned supervision that grounds prosody control in acoustic perturbations. On the Chinese split of InstructTTSEval, our recipe raises instruction-following from 34.5% without fine-tuning and 51.0% with naive fine-tuning to 56.4%, while constrained rewriting reduces drift from 40.4% to 15.4%. Ablations confirm the three mechanisms are complementary, and the drift taxonomy may generalize to instruction-driven generation beyond TTS.
Sep 4, 2026

EuroAlpaca: Task-Preserving Localisation of Instruction Data for European Languages

Machine translation (MT) offers a scalable way to extend English instruction-tuning data to multiple languages, but it can distort task-critical constraints and required outputs, creating corrupted training examples and degrading models trained on such data. We introduce EuroAlpaca, a task-preserving localisation pipeline and near-parallel resource covering 50 European languages and regional varieties, together with European-IFEval, a multilingual benchmark for verifiable instruction following. Depending on the example, our pipeline applies field-wise MT while preserving task-critical content or reconstructs a task-equivalent target-language instance, followed by validation of cross-field coherence and target-language consistency. Across LoRA experiments with four LLMs, training on directly translated data improves ROUGE-L and F-BERT on the Aya Evaluation Suite, but reduces accuracy on European-IFEval by 29.8% relative to the unadapted baseline. In contrast, adaptation with EuroAlpaca improves accuracy by 12.9% over the same baseline, reversing the degradation caused by direct MT, while also achieving the highest ROUGE-L and F-BERT scores on Aya. These results show that preserving task semantics is essential for multilingual instruction tuning.
Sep 2, 2026

How Output Format Confounds Data Quality and Capability in Instruction Tuning

Instruction-tuning data are judged by quality metrics, and tuned models are judged by benchmarks, but both judgments pass through an output interface: the surface format in which an answer is written. Using gradient signatures across 12 tasks, four semantically equivalent interfaces, three model families, and controlled corruptions, we show that this interface confounds both measurements. Spectral statistics such as effective rank are provably invariant to interface rotation and empirically blind to semantic corruption, while the direction of the update carries the quality signal. The interface-varying residual is not noise: it identifies each unit's own target task perfectly across all three families. Capability itself is stored relative to the training interface: a skill that raises accuracy by more than 40 points under the training format can be nearly invisible under every other, and correcting a single generation budget flips the measured effect of fine-tuning on GSM8K from a gain into a large loss. Pre-registered interventions delimit where this geometry stops short of control. Data quality and model capability are interface-conditioned quantities, and current practice often reports the interface instead of the content.
Aug 26, 2026

From Specialization to Generalization: Instruction-tuned LLMs for Robust Harmful Content Mitigation

Large language models (LLMs) demonstrate impressive performance across a wide range of general NLP tasks; however, their effectiveness in sensitive domains, such as hate speech detection, remains less clear. Prior studies comparing prompted LLMs with state-of-the-art encoder-based models (e.g., BERT variants (Roy et al., 2023; Dönmez et al., 2024)) have shown only marginal gains, suggesting that LLMs may not excel in hate speech detection or mitigation. In this work, we revisit this question through the lens of instruction tuning. By thoroughly unifying 36 English hate speech datasets spanning multiple labeling schemes, we fine-tune a generalist LLM, based on Qwen3 (Qwen Team, 2025), specifically for hate speech mitigation. Our results demonstrate not only state-of-the-art performance on in-domain benchmarks but also substantial improvements in cross-domain and cross-lingual generalization--areas where encoder-based specialist classifiers often struggle.
Aug 24, 2026

Industrial-Instruction: An End-to-End Framework for Building Instruction-Tuning and Benchmark Datasets from Industrial Technical Reports

Industrial technical reports contain high-value knowledge for maintenance, troubleshooting, and product engineering, but their heterogeneous structure (dense prose, specifications, tables) makes them difficult to index and reason over with standard retrieval and QA pipelines, and no public instruction-tuning or benchmark datasets are built from such documents. We address this gap with Industrial-Instruction, contributing (i) two open QA datasets built from real industrial technical reports and (ii) the end-to-end pipeline that produces them. Using 906 public Panasonic documents (7,525 pages), we apply layout-aware extraction, build a semantic retrieval index, and synthesize multiple-choice QA grounded in retrieved evidence under five query-document relationships (irrelevant retrieval, single-/multi-document support, single-/multi-document answer). After filtering an initial 23.9k generated samples, each dataset provides approximately 13.6k QA pairs with source documents and a held-out benchmark split. Fine-tuning small open LLMs (under 10B parameters) improves Set-Match Accuracy from 28.5% to 42.0% and F1 from 46.6% to 63.5% on the Panasonic benchmark. We release two parallel versions built by the same pipeline: one generated with the open-weight Qwen3-30B-A3B-Instruct model and one with the closed, API-based Claude-Opus-4.6 model, enabling a direct comparison of open- versus frontier-model data generation. The Claude-Opus-4.6 dataset yields a cleaner raw corpus and larger fine-tuning gains, at roughly two orders of magnitude higher cost. MMLU evaluation shows models trained on the Claude-Opus-4.6 data retain essentially all general knowledge, versus a small but measurable forgetting effect for the Qwen-generated data. Together, these datasets and pipeline offer a practical, reproducible path toward scalable industrial benchmarks and training data from real-world documentation.
Aug 13, 2026

Are You Sure You're Sure? On the Impact of Instruction Tuning on Confidence and Lexical Diversity

Instruction-tuned language models achieve strong performance across a range of generation tasks but have recently been shown to exhibit verbalized overconfidence, which may manifest in less diverse supporting rationales for incorrect answers. However, whether such overconfidence is associated with rationale consistency remains an open question. In this paper, we study whether changes in the lexical diversity of generated answer rationales accompany changes in model confidence induced by instruction tuning. We evaluate three matched base and instruction-tuned models across question-answering benchmarks and find that instruction tuning consistently increases answer confidence, despite limited changes in predictive accuracy, while degrading likelihood-based calibration. Secondly, we observe a non-uniform effect of instruction tuning on rationale diversity: cross-rationale diversity consistently decreases, whereas surface-level lexical diversity varies in both direction and magnitude across models and benchmarks. Finally, we find that these differences persist after controlling for answer selection and rationale length, confirming that confidence and rationale diversity capture distinct effects of instruction tuning.