TodayNewsResearchFieldsStacksMarketAnalytics
⌘K

Research

New papers ranked by attention, with their code, models and discussion linked. Pick a field on the left.

WeekMonthQuarterAll
NewestMost relevant

Papers

Open problems

Rising research topics

Ask the research

Sample
ByteTrail · AI and developer-tools intelligence
NewsResearchStacksFor AI agentsAPI
Fields · this week
All fields2209Multimodal297Frontier LLMs225Recursive self improvement200Inference & serving145Evals142Agents & indirect prompt injection138Robotics126Neural operators122Clinical reasoning114Agents97Forecasting time85Training & fine-tuning62Computer vision60Linear bandits57Voice AI55All 32 fields

Rising research topics

  • Spiking neural networks11+175.0%
  • Quantum machine learning34+25.9%
  • Character recognition OCR44+25.7%
  • Speech emotion23+21.1%
  • Voice AI76+13.4%
  • Forecasting time133−2.9%
  • Robotics195−3.0%
  • Channel state information18−5.3%

Open problems across fields

  • Comprehensive and reliable benchmarking for AI capabilities3 papers raise it · 3 address it

    Current benchmarks often lack the depth to fully characterize model capabilities, especially for complex, real-world, or long-horizon tasks.

    • SO-Arena: Benchmarking the Capability Spectrum and Reliability of System One Modelsraises · addresses
    • Infrathon: Benchmarking Language Agents for Real-World, Long-Horizon AI Infra Optimizationraises · addresses
    • One Table, Many Queries: Benchmarking Generalist Tabular Models Across Generation, Imputation, and Predictionraises · addresses
  • Effective AI agents in complex, dynamic real-world environments3 papers raise it · 3 address it

    AI agents struggle with real-world tasks requiring adaptation, learning from limited data, or debugging live systems, limiting practical deployment.

    • Where LLM Agents Get Stuck Debugging Databases: Access More Than Reasoningraises
    • World In-Context Learning for Agentic Robot Manipulationraises · addresses
    • Infrathon: Benchmarking Language Agents for Real-World, Long-Horizon AI Infra Optimizationraises · addresses
    • Sage-Taiyi: A Metacognitive Architecture for Scientific Discoveryaddresses
  • Efficient and scalable generation of complex data2 papers raise it · 2 address it

    Generating high-quality, complex data like streaming video or diverse tabular data efficiently and with fewer computational resources is crucial.

    • Few-Step Streaming Video Generation via Condition Shiftingraises · addresses
    • One Table, Many Queries: Benchmarking Generalist Tabular Models Across Generation, Imputation, and Predictionraises · addresses
  • Unified models for diverse data modalities and tasks2 papers raise it · 2 address it

    Creating single models that handle diverse tasks or data variations within a complex domain reduces complexity and improves generalization.

    • Prompt2Echo (P2Echo): Text-Prompted Multi-View Echocardiography Segmentation with One Modelraises · addresses
    • One Table, Many Queries: Benchmarking Generalist Tabular Models Across Generation, Imputation, and Predictionraises · addresses
  • Reducing computational cost of language model attribution1 paper raises it · 1 addresses it

    Understanding how prompt parts influence language model output is crucial for interpretability, but direct methods are computationally expensive.

    • Sketching Language Models for Context Attributionraises · addresses

Named by an AI from these papers' abstracts.

Papers

  • arXiv cs.LGSep 24

    Canopy: Exploiting Piecewise Smooth Tree Priors for Multi-Fidelity Bandits

    Michael Jerge, Suman Jana

    Many LLM inference problems, including model routing, prefix-cache management, prompt trimming, and test-time search, can be viewed as optimization over a tree. This structure arises naturally from autoregressive generation: every prefix defines a node, and its continuations form a subtree below it. Internal nodes of the tree provide cheap but biased estimates of a region's value, while leaf evaluations are expensive but accurate. Hierarchical bandit methods can exploit this structure, but typically require a specific smoothness schedule to be specified in advance, even though real objectives are often only piecewise smooth and their optima may lie near sharp boundaries. We introduce CANOPY, a multi-fidelity tree bandit that learns where the smoothness prior is valid rather than assuming it globally. CANOPY uses cheap random-path probes to construct an online certificate of local aggregation bias, then directs expensive leaf evaluations toward cells where the certificate detects a smoothness violation. We prove fixed-budget and regret guarantees whose additional cost is additive in the number of discontinuities, recovering the smooth-tree rate when no violations are present and…

    PDF
  • arXiv cs.ROSep 24

    Timed Rule-Based Supervision of an End-to-End Autonomous Parking Policy

    Kejia Gao, Liguo Zhou, Lei Yu, Alois Knoll

    We study whether a manually specified runtime supervisor can correct recurring failures of an existing end-to-end parking policy in a fixed CARLA parking lot. The vision-based Transformer architecture is inherited from Yang et al.; our contribution is a timed, rule-based Parametric Safety Shield (PSS) applied to its control outputs. The PSS uses hand-calibrated speed, position, and duration thresholds to intervene in observed failure modes, including boundary exits, delayed braking, and stalled or oscillatory control. In the reported closed-loop evaluation, 16 held-out target slots and six initial poses are each evaluated in four rounds (384 attempts per configuration). Target success increases from 327/384 (85.16%) for the retrained policy to 375/384 (97.66%) with the PSS; mean position and orientation errors among successful attempts are 0.21m and 0.33 degrees. These results show an improvement within this simulator setup. The repeated attempts share one map, vehicle, and sensor configuration, and the PSS uses simulator world coordinates; thus the results do not establish generalization to other lots or real vehicles, or a formal safety guarantee.

    PDF
  • arXiv cs.CLSep 24

    Low-Cost Assays for Measuring Model Behavior Across Vendors and Releases

    Tapan Parikh

    Language models advise people, keep them company, and write software while they sleep. Measuring what they do is hard: behavior has to be sampled repeatedly across models, prompts and releases, most of it lives in unstructured text that has to be coded before it can be counted, and the result has to be legible and rigorous enough to meaningfully compare models and vendors. To address these constraints, we present a simple, cheap, scalable, and replicable model for studying model behavior. Each study is a frozen, public stimulus run identically on a cross-vendor panel, at a few dollars per model or less. Each reads its transcripts one of three ways, chosen by how much interpretation the behavior needs: exact match on a clamped reply, a codebook applied by LLM judges whose agreement with a human coder is reported per code, and an instrumented environment that records what an agent did independently of what it said. Run across four years of model releases from both frontier and open-source labs, these instruments find four things. Convergence: asked to pick a word, 27 of 44 models answer serendipity at least once in four tries. Resistance: a trailing "right?" moves endorsement by up…

    PDF
  • arXiv cs.CLRAGSep 24

    Automated Regulatory Compliance Question Answering in Financial Services with Domain-Adapted Retrieval-Augmented Generation

    Tobias Deußer, Abhishek Pillai, Aurelio F. Bariviera, Dhananjay Bhardwaj, Lorenz Sparrenberg +3

    Financial institutions operate under dense, frequently amended rulebooks, and answering a compliance question correctly requires not only fluency but verifiable grounding in the authoritative text. Large language models are attractive for this task, yet the models that firms can realistically deploy on-premise are compact ones, and compact models hallucinate obligations. We study whether a carefully domain-adapted retrieval-augmented generation pipeline closes that gap. Our retriever is built in three stages on top of LegalBERT: entailment tuning that recasts question--passage matching as premise--hypothesis reconstruction, contrastive tuning with in-batch negatives, and score-level fusion with BM25. Our generator is a compact model (2B--12B parameters) served under 4-bit quantization, either prompted or adapted with retrieval-aware fine-tuning (RAFT) through LoRA. On ObliQA, a question-answering benchmark built from the Abu Dhabi Global Market rulebooks, the staged retriever raises Recall@10 from 0.256 to 0.774 and outperforms BM25 (0.678) and E5-large-v2 (0.758), the strongest general-purpose dense encoder we tested. RAFT-LoRA then improves the composite RePASs answer-quality…

    AISmall LLMs80PDF
  • arXiv cs.CLVoice AISep 24

    VietPrism: A large-scale Vietnamese speech and deepfake corpus with diverse dialects and code-switching

    Minh Hoang, Thai Le

    Vietnamese speech research is constrained by resources that isolate automatic speech recognition from speaker, dialect, code-switching, and deepfake analysis. We introduce VietPrism, an open, multi-domain corpus that brings these dimensions together at scale: 993.4 hours and 403,941 bona fide utterances from 1,262 verified speakers across 8,388 real-world videos. To our knowledge, it is the first large-scale Vietnamese corpus to jointly provide transcripts, consistent speaker identities, five dialect groups, and naturally occurring Vietnamese--English code-switching, which constitutes nearly half of the corpus by duration. We further create over 3.1K hours of spoof speech with four open-source and commercial synthesis systems. Every spoof is conditioned on a verified speaker reference and paired with a transcript- and speaker-matched bona fide utterance, enabling unique controlled evaluation with reduced lexical and identity confounds. Zero-shot evaluation of five pretrained multilingual detectors reveals striking brittleness: EER greatly varies across detector--generator pairings, while recent multilingual detector DFA-1B degrades from 16.3% to 33.6% as speaker similarity…

    AIEvals80PDF
  • arXiv cs.AIAgentsSep 24

    Advancing Model Research in AgentX: Long-Horizon Autonomy for Industrial Recommender Systems

    Shuang Yang, Zijie Zhuang, Changxin Lao, Pengbo Xu, Hanwen Xu +18

    Sustaining industrial recommendation research requires using the results of one experiment to decide what to investigate next. We present AgentX-Model, the next generation of AgentX's model research framework, which connects proposal development and model experimentation within sandboxes defined by business inputs and prediction tasks. AgentX-Model adopts a dual-agent architecture comprising a Research Agent and a Model Agent. The Research Agent develops independently reviewed proposals from papers and experimental findings, while the Model Agent conducts multi-round investigations and returns code, measurements, and unresolved questions. Using the returned results, the Research Agent selects a starting implementation and formulates the next research question, allowing subsequent experiments to build on earlier findings. We organize this continuing research around four actions: Reproduce, Follow-up, Composition, and Diagnose. The first three actions drive routine research, while Diagnose acquires the evidence needed to choose a repair, including for issues raised by business feedback and online evaluation, such as prediction bias measured by PCOC. Across the production…

    AI70PDF
  • arXiv cs.CVMultimodalSep 24

    GHOST-Q: Towards Studying Grounding Hallucinations Overlooked Under Same-score TradeOffs in Quantized VLMS

    Saim Rehman, Muhammad Shafique

    Post-training quantization of vision--language models (VLMs) is typically assessed through aggregate task accuracy and memory savings, but preserving a headline score does not guarantee preservation of visual grounding behavior. We present GHOST-Q, a cross-precision controlled evaluation of three 8B VLM families under FP16, INT8, and NF4 across utility and hallucination-sensitive benchmarks. Rather than comparing only aggregate accuracy, we pair FP16 and quantized predictions item by-item to quantify how compression redistributes grounding successes and failures. Five of six quantized variants preserve MMStar accuracy within $\pm2$ percentage points, yet 10 of 36 paired effects remain significant after false-discovery-rate correction, nine on hallucination-sensitive conditions. Same-device A100 profiling further demonstrates that substantial memory reduction does not necessarily mean lower inference latency. Finally, an open-ended AMBER audit reveals strong generation budget censoring whose severity varies by architecture and precision. These results show that quantized VLMs should be evaluated jointly for aggregate utility, grounding reliability, generation behavior, and…

    AIInference & serving80PDF
  • arXiv cs.HCSep 24

    Guardrails or Roadblocks? Effects of Pedagogical Style and Context Awareness in AI Teaching Assistants for Programming

    Madeleine Eastwood, Harshith Narne, Joseph Hilby, Paul Denny, Ashish Aggarwal +1

    AI teaching assistants (AI TAs) backed by large language models (LLMs) and pedagogical guardrails are increasingly being integrated into programming courses, providing students with scalable access to hints, conceptual explanations, and code-level feedback. However, guardrails may also create friction. If students feel that the support provided is overly restrictive or poorly contextualized to their current progress, they may bypass approved tools for general-purpose LLMs. To investigate how AI TA design affects students' learning experiences, we conducted a randomized controlled trial with 132 students in an introductory programming course. Students completed three tasks related to code-writing and debugging and were randomly assigned to one of four AI TAs varied across two dimensions: pedagogical guidance style (Socratic vs. Direct instruction) and context awareness (no context vs. full context of the problem and student solution). We examined students' perceptions, interaction behaviors, and evidence of post-task comprehension. Students rated the Socratic AI TA with full context least favorably, reporting significantly lower perceived support for task completion.…

    PDF
  • arXiv cs.LGSep 24

    Let Training Guide Selection: Online Synthetic Data Filtering via Real-Anchored Utility

    Yanran Wu, Sana Lakdawala, Renzo Tassara Miller, Chongyang Bai, Sharath Ciddu +4

    Synthetic data can scale training supervision when real-world data are limited, but noise and distribution mismatch can reduce its value. Existing synthetic data selection methods often emphasize fidelity or diversity rather than the learner's evolving needs. We propose FROST, an online framework that estimates synthetic-data utility through gradient feedback anchored in real training data. It calibrates batch utility against recent history to determine when filtering is needed and filters samples only in out-of-band batches to determine what to retain, without an external verifier or held-out validation set. Experiments on two public benchmarks for image classification and LLM fine-tuning for text-to-SQL show that FROST filters out around 20--30% of the synthetic data while improving real-task performance compared with training on the full synthetic data pool. We further apply FROST during training in a large-scale industrial ads re-ranking system, achieving significant performance gains over a highly optimized production baseline, demonstrating its effectiveness and generalizability.

    PDF
  • arXiv cs.CVSep 24

    OceanXL: Large-scale Underwater 3D Gaussian Splatting via Block Partitioning and Adaptive Pruning

    Haoran Wang, Shaoyu Cai, Adrian Azzarelli, Zhuodong Jiang, Guoxi Huang +5

    Underwater 3D reconstruction is critical for marine exploration, ecological monitoring, and subsea infrastructure inspection, yet remains challenging at large scale due to light attenuation, scattering, and limited capture coverage. While 3D Gaussian Splatting (3DGS) enables high-quality real-time rendering, its application to large underwater scenes is constrained by high memory consumption and inefficient optimization over extensive areas. We propose OceanXL, a fast and scalable 3DGS-based framework for large-scale underwater reconstruction. OceanXL adopts a divide-and-conquer strategy, partitioning scenes into spatially coherent blocks to enable efficient optimization while preserving global geometric consistency. We further introduce an adaptive pruning scheme tailored to underwater conditions that removes redundant primitives, producing compact representations without sacrificing visual fidelity. Together, these components improve training efficiency and rendering performance for large scenes. We also introduce a large-scale underwater dataset covering diverse marine environments. Experiments on five large-scale scenes demonstrate favorable scalability, compactness, and…

    PDF
  • arXiv cs.IRSep 24

    From Interests to Semantic IDs: Retrieval-Grounded Credit Assignment for Generative Recommendation

    Mengdan Zhu, Yufan Zhao, Yao Zhao, Sophie Di, Tao Di +3

    Semantic IDs (SIDs) encode each catalog item as a short token sequence, enabling generative recommenders to predict the next item autoregressively. Reasoning-enhanced variants, an increasingly common extension, first generate a textual trace and then decode a next-item SID by beam search. Such recommenders are commonly trained with group-relative policy optimization under an exact-match SID reward, which is sparse in large catalogs. Two failure modes follow. When all rollouts in a group miss the target, the group yields zero advantage and no learning signal. Rollouts sharing the same SID reward receive identical advantages, however much their traces differ. In both cases the reward reflects only the decoded SID, never the reasoning that produced it. This creates a credit-assignment gap. We address this gap with retrieval-grounded query attribution. Each trace is structured into a history summary, a set of interest hypotheses, and a final SID. A frozen retriever executes every hypothesis as a catalog query, so that each hypothesis becomes independently verifiable rather than judged only through the final SID. A rollout is rewarded when any of its queries retrieves the target…

    PDF
  • arXiv cs.LGSep 24

    Strategic Self-Consistency

    Tori Qiu, Ander Artola Velasco, Manuel Gomez-Rodriguez

    Self-consistency has become a popular technique for enhancing the reasoning abilities of large language models by generating multiple reasoning paths and selecting the final answer through a majority vote. However, because model providers typically charge users in proportion to the number of reasoning paths generated, they have a financial incentive to artificially increase the path count. In this work, we show that an unfaithful provider can exploit this incentive using a simple, efficient algorithm while avoiding detection by an auditor: by generating and strategically reordering additional reasoning paths, the algorithm makes every path appear necessary to reach the majority. To validate our algorithm, we conduct experiments with multiple instruct models from the Llama and Qwen families, as well as reasoning models distilled from DeepSeek-R1, on benchmark datasets spanning mathematics, science, and question answering. Our results suggest that the distribution of additional reasoning paths generated by our algorithm is heavy-tailed and that substantial capacity to overcharge remains even under the best possible audit designed to keep the false-positive rate below $α= 0.1$.

    PDF
  • arXiv cs.AISep 24

    Faithfulness Without Anchors: A Position Paper on the Structural Misalignments of Explainable AI Faithfulness Metrics for Feature Importance

    Marco Zullich, Emily Schiller, Ivan Gentile, David Dembinsky, Adriano Lucieri +2

    Explainable Artificial Intelligence (XAI) tools have often been presented as methods to increase trust in machine learning models by making aspects of their algorithms more human-understandable. However, evidence accumulated over the last decade points to an overall failure of this objective, instead promoting overreliance on incorrect predictions, which can subsequently lead to algorithmic aversion. Rather, XAI appears to function primarily as a debugging tool, useful for obtaining anecdotal insights into, for example, features learned by the model. One of the main issues underlying this limitation is that XAI methods often oversimplify the inner workings of the model, resulting in misalignment between the explanation and the model itself - a behavior referred to as unfaithfulness or infidelity. Faithfulness can be interpreted as the degree of approximation between an explanation and the model. Computing faithfulness in a rigorous manner would thus make it possible to quantify one of the various elements of trustworthiness of an explanation. However, existing metrics present significant limitations, making this quantification either infeasible or computationally expensive. In…

    PDF
  • arXiv cs.LGSep 24

    GeoACE: Geometry-Diverse Anchor-Correction Ensembles for Heterogeneous Treatment Effect Estimation

    Ali Haghpanah Jahromi, Mohammad Taheri, Zohreh Azimifar

    Predicting heterogeneous causal effects supports informed decision-making, yet suitable inductive biases vary with treatment overlap, covariate imbalance, prognostic structure, and sample size. The Geometry-Diverse Anchor--Correction Expert Ensemble (GeoACE) addresses this variation through five complementary experts whose predictions are combined using weights estimated from validation data. Compared with ensemble designs that derive diversity primarily from sample or feature subsets, reweighting, or feature transformations, its central contribution is the coordinated design of training objectives and anchor geometries that target different prognostic and overlap structures. Following validation, each expert is refitted on the combined training and validation data, while retaining its distinctive structure. The shared architecture couples a structured outcome anchor with a learned neural correction. ACE serves as the reference estimator, while OW-ACE emphasizes well-supported covariate regions through overlap-weighted factual training. G$Φ$-ACE constructs a global outcome-guided geometry from pooled prognostic associations, whereas A$Φ$-ACE incorporates arm-specific outcome…

    PDF
  • arXiv eess.ASSep 24

    PRIME-ANC: Path-Ratio-Informed Modeling for Efficient Neural Filter Synthesis in Active Noise Control

    Yaokun Huang, Chunyang Xu, Haowen Hua, Sen Lin, Shichao Hu +1

    Changes in listener acoustics require active noise control (ANC) filters to be redesigned for new acoustic paths. We introduce PRIME-ANC, a shared neural synthesizer that learns a bounded, path-dependent logmagnitude correction to a regularized path-ratio base. Minimum-phase reconstruction and truncation produce finite-impulse-response (FIR) filters; training optimizes their noise-control performance. Across ten random training/test splits per dataset, PRIME-ANC achieves average held-out one-third-octave reductions of 18.81 and 17.76 dB over 50 Hz-5 kHz on a ten-path dataset and a public earphone database, respectively. On original earphone measurements, it improves upon the path-ratio base by 7.89 dB. Ablation studies support the contributions of both the analytic base and the path-dependent correction. Given calibrated paths for a held-out listener condition, PRIME-ANC generates a path-specific FIR without iterative optimization. With three Gauss-Newton updates, it reaches 21.43 dB reduction, approaching direct weighted least-squares design while producing lower amplification and root-mean-square control output.

    AI20PDF
  • arXiv cs.IRSep 24

    Learning Better Reasoning for Generative Recommendation with Semantic IDs

    Mengdan Zhu, Yufan Zhao, Sophie Di, Yao Zhao, Tao Di +3

    Generative recommendation reformulates item retrieval as sequence generation, allowing a unified model to directly generate the next item from a user's interaction history. Semantic IDs further make this paradigm effective and scalable by representing each item as discrete codes, enabling knowledge sharing among semantically related items. Recent studies introduce explicit reasoning before Semantic-ID generation, helping models summarize user interests and infer possible preference transitions. However, reasoning is not inherently beneficial: Inaccurate or uninformative reasoning may mislead subsequent item generation and ultimately degrade recommendation performance. This raises a central challenge: how can a recommender select and learn effective reasoning traces and progressively evolve toward better reasoning from its own generations? In this work, we propose Evo-Rec, a three-stage framework for learning better reasoning and further enhancing it through reinforcement learning. First, we align Semantic IDs with their textual and behavioral contexts, enabling the model to understand and generate item identifiers. Second, we sample multiple candidate reasoning traces and retain…

    PDF
  • arXiv cs.ROSep 24

    World Action Agent: Harnessing VLMs for Robot Manipulation via World Action Rehearsal

    Yehang Zhang, Haojian Huang, Yifan Chang, Jianchong Su, Bohan Zhou +11

    General-purpose vision-language models (VLMs) bring broad knowledge and spatial reasoning to robot manipulation, yet existing systems either use them indirectly, to predict constraints or write programs, or give them a view of the scene rather than a world in which to act. We present World Action Agent (WAA), a multi-agent harness through which VLMs pilot robots with basic tools, making every decision within a visual action workspace. The workspace has three properties. Contact views, selected automatically from the scene geometry, present the scene around the current interaction. Action rehearsal turns each action into an editable proposal that the agent, alone or through an Imagination Agent, previews and revises against planning feedback before execution. In-view correction closes the loop between observation, rehearsal, and low-level execution, letting the agent remove residual offsets in the view where it observes them. Through the same workspace, WAA acquires embodied procedural knowledge in two ways: it evolves multimodal skills from expert videos and human teaching under evidence-based review and consults them through a Skill Agent, and its interaction traces train…

    Code
  • arXiv cs.CVSep 24

    ADATEX4D: adaptive texture capacity allocation for 4D gaussian splatting

    De Jiang, Peiqiang Wang, Kehong Yuan, Shaohua Ma

    Textured Gaussians improve local appearance capacity, but assigning the same texture resolution to every primitive wastes storage on low-detail or weakly visible regions. We introduce AdaTex4D, an adaptive texture-capacity module for deformation-based 4D Gaussian Splatting. Each Gaussian carries packed RGBA triplanes whose two axes grow independently according to visibility normalized screen-space gradients and deformed local scales. Experiments on N3DV and PanopticSports show that AdaTex4D reduces texture storage by more than half while preserving reconstruction quality. Under fixed memory budgets, adaptive allocation also improves quality over uniform texture assignment and reduces overall model and peak memory. These results show that dynamic, anisotropic texture allocation provides a more efficient way to distribute local appearance capacity in 4D Gaussian representations.

    PDF
  • arXiv cs.LGSep 24

    A Contraction Framework for Stochastic Operators with Bootstrapping: Application to TD Learning

    Ids van der Werf, Sergio Rozada, Antonio G. Marques

    Many iterative algorithms rely on bootstrapping. A variable is updated using a second, frozen copy as a target, which is periodically replaced with the updated variable. Majorize-minimize and inexact proximal-point methods share this structure, as does temporal-difference (TD) learning. However, existing convergence guarantees for scenarios that combine sampled updates with targets refreshed only every $K$ steps rely on the specific structure of the update, such as linear approximation or gradient-based inner steps, and on uniformly bounded sampling error. We instead model the sampled update as a stochastic operator on the parameter space, which reduces the analysis to a contraction argument that needs no gradient structure and allows the sampling error to grow with the iterates. Within this framework, we derive a finite-time bound for i.i.d. samples and any target-update period $K$. We show that the iterates converge geometrically in root mean square to a ball around the fixed point, provided the sensitivity to the frozen target is smaller than the contraction slack of the inner map. Existing deterministic frozen-target contraction and stochastic-gradient-type bounds follow as…

    PDF
  • arXiv cs.LGSep 24

    Beyond Average Safety: Chance-Constrained LLM Fine-tuning

    Taha Entesari, Mahyar Fazlyab

    Fine-tuning large language models on new objectives can improve helpfulness, instruction following, or domain-specific performance, but it can also induce regressions on safety-critical prompts. Existing safety-preserving fine-tuning methods typically control average safety loss or use weighted auxiliary penalties, which can obscure rare but severe failures. We propose a chance-constrained formulation for safety-preserving fine-tuning that limits the fraction of safety examples whose degradation relative to a reference model exceeds a prescribed threshold. Because the resulting empirical chance constraint contains a discontinuous indicator, we introduce a differentiable majorization of the violation rate, yielding a tractable conservative constraint. We then develop a constraint-aware gradient descent method that treats the majorized constraint as a safe set in parameter space and minimally modifies the fine-tuning direction to preserve feasibility. The resulting update admits a closed form and produces a tail-aware safety correction that emphasizes examples near or above the degradation threshold. We conduct an extensive set of experiments on harmful fine-tuning across three…

    PDF
Load more
PDF