Multilingual Fine-Tuning via Localized Gradient Conflict Resolution
Long P. Hoang, Yiran Zhao, Wei Lu, Wenxuan Zhang
Abstract
The rapid evolution of Large Language Models (LLMs) has established cross-lingual versatility as a defining feature of modern systems. However, fine-tuning these models frequently induces negative interference across languages. To address this, we reformulate multilingual fine-tuning as a multi-objective optimization (MOO) problem. Specifically, we introduce Bucket-Level MOO, a scalable distributed framework that applies gradient-based MOO algorithms locally on parameter buckets. This enables conflict-aware updates without the prohibitive communication overhead of reconstructing full gradient vectors. Theoretically, we prove this localized resolution natively enforces Refined Pareto Stationarity, a strictly tighter necessary condition for Pareto optimality. Empirically, Bucket-Level MOO mitigates interference by driving LLMs to construct distinct language-specific dimensions, improving representational separability. Extensive experiments across four base LLMs demonstrate that our method significantly improves both seen and unseen multilingual performance over standard fine-tuning paradigms.
AI Impact Assessments
(1 models)Scientific Impact Assessment: Multilingual Fine-Tuning via Localized Gradient Conflict Resolution
1. Core Contribution
This paper addresses negative interference during multilingual LLM fine-tuning by reformulating the problem as multi-objective optimization (MOO) and introducing Bucket-Level MOO, a distributed framework that applies gradient-based MOO algorithms (MGDA, CAGrad, PCGrad) locally within parameter memory buckets rather than across the full model. The key insight is twofold: (1) gradient conflicts are structurally heterogeneous across layers, so localized resolution is more appropriate than global aggregation; and (2) distributed training systems (ZeRO, FSDP) already partition gradients into buckets, making local MOO a natural fit that avoids the prohibitive communication overhead of reconstructing full gradient vectors.
The paper provides a theoretical contribution by proving that bucket-level resolution enforces Refined Pareto Stationarity (RPS), a strictly tighter necessary condition for Pareto optimality than the standard Pareto stationarity guaranteed by global MOO methods. This is a notable result because the practical engineering decision (localizing computation for efficiency) turns out to be theoretically superior, not merely an approximation.
2. Methodological Rigor
Theoretical Analysis: The convergence proofs (Theorems 3.1 and 3.2) are cleanly structured and leverage the additive decomposition of inner products over disjoint parameter blocks. The connection to Refined Pareto Stationarity from Hu and Yu (2025) is well-motivated and correctly applied. However, the theoretical guarantees rely on the partition structure being fixed (determined by the distributed training framework's bucket size), and there is no analysis of how partition granularity affects convergence rates or solution quality—only that finer partitions yield tighter stationarity conditions.
Experimental Design: The experimental setup is reasonable, covering four base LLMs (Meta-Llama-3-8B, Llama-3.1-8B, Qwen3-4B-Base, Qwen3-8B-Base), eight seen and five unseen languages, and four diverse benchmarks. The use of both seen and unseen language evaluation is a strength for assessing generalization. However, several concerns arise:
3. Potential Impact
Practical Relevance: The framework elegantly integrates into existing distributed training pipelines without requiring architectural modifications, making it immediately deployable. The minimal computational overhead (0.6 → 0.7-0.8 hours) is a significant practical advantage. The 41% reduction in peak VRAM (123 GB → 72 GB) compared to global MOO makes MOO feasible for LLM-scale training for the first time.
Broader Applicability: While framed for multilingual fine-tuning, the Bucket-Level MOO framework could generalize to any multi-task learning scenario with LLMs—multi-domain fine-tuning, instruction tuning across diverse task types, or RLHF with multiple reward objectives. This broadens potential impact considerably.
Mechanistic Insights: The analysis showing that MOO drives models to construct language-specific neural dimensions (increased Silhouette scores, higher language-specific neuron ratios) provides interpretable evidence for *why* the method works, going beyond pure performance metrics.
4. Timeliness & Relevance
The paper addresses a genuine bottleneck in the field. As multilingual LLM deployment accelerates globally, the tension between serving diverse languages with shared parameters is increasingly acute. The observation that even strong multilingual models (Qwen3-8B-Base) still exhibit gradient conflicts validates the problem's relevance. The framework's compatibility with standard distributed training infrastructure (DeepSpeed ZeRO, PyTorch FSDP) makes it immediately relevant to practitioners.
The connection to Refined Pareto Stationarity is timely, building on very recent theoretical work (Hu and Yu, 2025, ICLR) and demonstrating its practical utility in a high-impact application.
5. Strengths & Limitations
Key Strengths:
Notable Limitations:
Summary
This is a well-executed paper that identifies a genuine practical problem (gradient conflicts in multilingual LLM fine-tuning), proposes a clean solution that leverages distributed training infrastructure, and provides both theoretical justification and empirical validation. The insight that localized MOO is both more efficient *and* theoretically superior is compelling. The main limitations are the relatively small experimental scale and the narrow baseline comparisons. The work represents a meaningful methodological advance at the intersection of multi-objective optimization and large-scale LLM training.
Generated Jun 5, 2026
Comparison History (23)
Paper 1 addresses a fundamental and practically important problem in multilingual LLM fine-tuning with both theoretical contributions (proving Refined Pareto Stationarity) and empirical validation across multiple models. Its scalable distributed framework for gradient conflict resolution has broad applicability beyond multilingual settings to any multi-objective fine-tuning scenario. Paper 2, while interesting as a benchmark for evaluating inductive reasoning, is primarily diagnostic—it characterizes LLM limitations but doesn't provide solutions. Paper 1's methodological innovation with immediate practical utility for the growing multilingual AI community gives it higher potential impact.
Paper 1 addresses a critical bottleneck in the development of autonomous agents: step-level verification in open-ended tool use. By introducing a human-annotated benchmark for process supervision beyond mathematics, it directly enables future research into Process Reward Models (PRMs) for agents, a highly active and impactful area. While Paper 2 offers a rigorous optimization technique for multilingual fine-tuning, Paper 1 provides foundational infrastructure (data and evaluation) that will likely be widely adopted to advance agentic reasoning, leading to broader scientific impact and real-world application.
Paper 2 has higher likely impact due to a concrete, scalable method for a widely encountered problem (multilingual fine-tuning interference) with clear real-world applicability and broad relevance to LLM training. It offers methodological rigor via a formal MOO reformulation, a distributed bucket-level algorithm addressing communication constraints, theoretical guarantees (Refined Pareto Stationarity), and extensive empirical validation across multiple base models with gains on seen/unseen languages. Paper 1 is conceptually interesting for disagreement-aware routing in value-laden tasks but appears more framework/proposal-oriented with narrower application scope and less demonstrated empirical rigor.
Paper 1 addresses the critical and timely problem of security for computer-use agents, proposing a novel architectural solution (NOVA) that reconciles security isolation with practical utility. It identifies a new attack vector (Branch Steering) and demonstrates that single-shot planning can provide control flow integrity against prompt injection. Given the rapid deployment of AI agents in real-world settings, this work has immediate practical relevance and broad impact across AI safety, security, and HCI. Paper 2, while technically solid in multilingual optimization, addresses a more incremental improvement in a well-studied area.
Paper 1 is more novel and potentially higher-impact because it identifies and formally corrects a widely used but biased estimand in RLVR, providing an exact causal decomposition with pre-registered experimental validation and tooling that can immediately audit prior and future alignment results. Its methodological rigor (proofs, controlled simulator, factorial design, identification/bounding analysis, re-audits) and direct relevance to current RLHF/RLVR practice give it broad influence across alignment, evaluation methodology, and causal inference. Paper 2 is useful and timely for multilingual tuning, but the bucketed MOO idea is a more incremental systems/optimization contribution with narrower conceptual spillover.
Paper 1 addresses a fundamental bottleneck in LLM training (multilingual interference) with a theoretically rigorous and scalable framework. Its broad applicability across foundational AI models gives it a wider, horizontal scientific impact across NLP and machine learning compared to the highly domain-specific, albeit important, public health application of Paper 2.
Paper 2 likely has higher impact due to timeliness and breadth: controlling safety drift in self-evolving agents is a central open problem for autonomous AI, relevant to alignment, agentic systems, and deployment governance. It offers actionable guidance (where/when oversight matters) and evaluates across multiple domains (coding, math, safety), increasing real-world applicability. Paper 1 is methodologically strong and novel for multilingual fine-tuning, but its impact is narrower (primarily multilingual adaptation) and competes with many existing gradient-conflict/MTL methods, whereas Paper 2 addresses a rapidly emerging capability with high societal stakes.
Paper 1 addresses a fundamental challenge in multilingual LLM fine-tuning with both theoretical contributions (proving Refined Pareto Stationarity) and practical scalability via distributed bucket-level multi-objective optimization. It demonstrates broad applicability across four base LLMs with improvements on both seen and unseen languages. Paper 2, while well-structured, presents a domain-specific framework for Reddit community adaptation with narrower scope. Paper 1's combination of theoretical rigor, methodological novelty, and broad impact on the widely-studied multilingual NLP problem gives it significantly higher potential for citations and influence.
Paper 1 addresses a fundamental challenge in LLM multilingual fine-tuning with a novel theoretical framework (Bucket-Level MOO) that has broad applicability across all multilingual NLP tasks. It provides both theoretical guarantees (Refined Pareto Stationarity) and empirical validation across multiple LLMs. The breadth of impact is significantly larger given the ubiquity of multilingual LLMs. Paper 2, while practical, addresses a narrower application domain (traffic sign inspection) with a reformulation approach that, though useful, has more limited generalizability and audience.
Paper 1 addresses a fundamental challenge in multilingual LLM training with both theoretical contributions (proving Refined Pareto Stationarity) and practical scalability innovations. Its multi-objective optimization framework for distributed training has broad applicability beyond multilingual settings. Paper 2, while practically useful for inference efficiency through KV cache eviction, represents a more incremental contribution in the crowded space of efficient inference methods. Paper 1's combination of theoretical depth, novel distributed optimization framework, and impact across multiple languages and base models suggests broader and more lasting scientific influence.
Paper 1 proposes a novel, theoretically grounded optimization framework for LLM fine-tuning that tackles a fundamental challenge (negative interference) with broad applicability. In contrast, Paper 2 is an evaluation study of existing models on a specific task (Lean 4 formalization). The methodological innovation and theoretical contributions of Paper 1 provide a significantly higher potential for broad scientific impact.
Paper 1 addresses the highly impactful problem of multilingual LLM fine-tuning, proposing a scalable distributed framework (Bucket-Level MOO) with both theoretical guarantees (Refined Pareto Stationarity) and strong empirical results across multiple LLMs. It sits at the intersection of multi-objective optimization and large-scale NLP, fields with enormous current interest and broad applications. Paper 2 makes a solid contribution to combinatorial search algorithms for longest-path problems, but targets a narrower community with more specialized applications. The timeliness, breadth of impact, and relevance to the LLM ecosystem give Paper 1 substantially higher potential impact.
Paper 1 addresses a fundamental challenge in LLM development—cross-lingual interference—with a mathematically rigorous multi-objective optimization framework. Providing theoretical proofs for Refined Pareto Stationarity and empirical validation across multiple models, it offers deep methodological innovation. Paper 2, while highly relevant to the growing field of local agents, solves a more niche architectural and deployment problem. The scalable gradient conflict resolution proposed in Paper 1 has a wider potential to influence foundational model training paradigms across the broader NLP community.
Paper 1 addresses a foundational challenge in LLM training with high methodological rigor, including theoretical proofs for Pareto stationarity and extensive empirical evaluation. Its scalable optimization framework fundamentally improves model representations. In contrast, Paper 2 presents a practical engineering system for the application layer, but its evaluation is limited to a small sample of 21 sessions, making its broad scientific and theoretical impact comparatively lower.
Paper 1 presents a technically rigorous contribution with theoretical guarantees (Refined Pareto Stationarity), empirical validation across four LLMs, and addresses a practical, well-defined problem (multilingual interference in fine-tuning). It combines novel methodology (Bucket-Level MOO) with scalability considerations for distributed training. Paper 2 proposes a conceptual architecture for conversational AGI motivation but lacks empirical validation, remaining largely speculative with 'sketched' extensions. Paper 1's concrete results, theoretical foundations, and immediate applicability to the widely-used LLM fine-tuning pipeline give it substantially higher near-term and likely long-term scientific impact.
Paper 1 introduces a novel, scalable algorithm for LLM fine-tuning backed by theoretical proofs and extensive empirical validation, directly addressing a major technical bottleneck in a highly active field. Paper 2, while socially relevant, is primarily a synthesized review of existing data and policy frameworks, offering less methodological innovation and technical advancement compared to Paper 1.
Paper 1 presents a novel, theoretically grounded approach to a significant problem (multilingual interference in LLM fine-tuning) with rigorous mathematical guarantees (Refined Pareto Stationarity) and extensive empirical validation across multiple models. It offers both theoretical contributions and practical scalability through its bucket-level MOO framework. Paper 2 proposes an entropy-based evaluation framework that, while useful, is more incremental—applying well-known information-theoretic concepts to agent evaluation without deep theoretical novelty. Paper 1's methodological rigor, broader applicability, and stronger theoretical foundations give it substantially higher impact potential.
Paper 2 likely has higher impact: it tackles a broad, widely-relevant problem (multilingual fine-tuning interference) with a generally applicable optimization framework. The localized gradient conflict resolution is novel, scalable for distributed training, and comes with stronger theoretical guarantees (Refined Pareto Stationarity), increasing methodological rigor and credibility. Its applicability spans many LLMs, languages, and downstream tasks, giving broad cross-field and industry relevance. Paper 1 is innovative for agent memory/state management and shows strong efficiency gains, but its impact is more niche to long-horizon agent architectures and specific benchmarks.
Paper 1 offers stronger scientific impact through its novel theoretical contribution (proving Bucket-Level MOO enforces Refined Pareto Stationarity) combined with practical scalability for distributed multilingual fine-tuning. It addresses a fundamental problem—negative interference in multilingual LLMs—with rigorous methodology spanning theory, mechanistic analysis (language-specific dimensions), and extensive empirical validation across four base LLMs. Paper 2 provides a valuable empirical benchmark and practical baseline (AutoMEM) for agentic memory, but is primarily an empirical comparison with a relatively incremental architectural insight. Paper 1's broader theoretical and methodological contributions give it higher impact potential.
Paper 1 addresses a fundamental and broadly applicable challenge in multilingual LLM fine-tuning with both theoretical contributions (proving Refined Pareto Stationarity) and practical scalable solutions. Its impact spans NLP, optimization, and distributed systems, with demonstrated improvements across multiple base LLMs. Paper 2, while timely and socially important regarding AI deception and ethics, is primarily an observational content analysis of a single dataset from a discontinued experiment, limiting its methodological generalizability and broader technical impact. Paper 1's contributions are more likely to influence ongoing research directions in the rapidly growing multilingual AI field.