arXiv subjects

Jian Cheng

Publications and source records attributed to Jian Cheng.

At least 19 recordsLinked to original sources

arXiv2026

Learning to Model Optimization Problems from Solver Outcomes

Automating optimization modeling with LLMs is a promising path toward scalable decision intelligence, but existing approaches either rely on agentic pipelines built on closed-source LLMs with high inference latency or fine-tune smaller LLMs using costly process supervision that often overfits to a single solver API. We study how reinforcement learning from inexpensive outcome labels can improve optimization modeling and support reuse across heterogeneous solver environments. To this end, we propose Optimization Modeling from Solver Outcomes (OMSO), which uses reinforcement learning (RL) with a mathematical programming solver as a deterministic, interactive environment. Given a natural-language problem and a target solver, OMSO generates solver-specific code, executes it in a sandboxed harness, and converts execution outcomes into scalar rewards in a closed-loop generate-execute-feedback-update process. This RL formulation learns optimization modeling without process-level supervision and supports cross-solver generalization by switching the solver environment rather than reconstructing solver-specific datasets. Experiments on NL4OPT, MAMO, IndustryOR, and OptiBench across Gurobi, OR-Tools, and COPT show that OMSO matches or outperforms process-supervised SFT, supports zero-shot training-free transfer, direct RL training, and SFT-cold-started RL, and improves modeling reasoning and semantic correctness beyond solver API usage.

cs.AI↗

arXiv2026

Let the Gate Decide: Per-Lookup Precision for Quantized Conditional Memory in LLMs

Conditional memory layers, such as DeepSeek's Engram, attach large hashed n-gram tables to a language model, and removing them from DeepSeek-V4.1-Flash costs 36.8 PopQA points. Low-bit quantization loses long-tail facts first, and the standard remedy, giving more bits to frequently accessed rows, backfires: refining lookups by access frequency is worse than refining them at random. We trace this to the gate that scales each retrieved row. A lookup's quantization error is weighted by its gate, which is small for frequent n-grams, large for the rare n-grams of long-tail entities, and computable from a 1-bit code. GateRQ therefore stores each row as a ladder of 1-bit residual codes, keeps level 1 in GPU memory to compute the gate, and reads deeper levels only for the lookups the gate ranks highest. Without retraining, GateRQ cuts GPU table memory by 7.8x on DeepSeek-V4.1-Flash, which raises vLLM throughput by 11%-31% over FP8 tables, while losing about 1 PopQA point at 1.50 bits read per lookup and 0.2 points at 2.10 bits across two models.

cs.LG↗

arXiv2026

RAQ-MLA: Role-Aware Quantization for Multi-Head Latent Attention Caches

Multi-Head Latent Attention (MLA) has been adopted by recent frontier large language models. Instead of per-head keys and values, MLA caches a compressed KV latent that is up-projected into keys and values at decode time, together with a narrow positional key that carries position information into the attention score. Existing KV-cache quantization methods are designed around the keys and values of standard attention and do not account for what MLA actually caches, which costs accuracy at low bit widths. We analyze the MLA cache both in activation distribution and in how quantization error propagates to the attention output, and show that the two cached tensors play different roles: the latent mainly supplies the values, whereas the positional key dominates the attention score. We therefore propose RAQ-MLA (Role-Aware Quantization for Multi-Head Latent Attention Caches), which fits an invertible transform to each cached tensor against the error that tensor delivers to the attention output: attention-weighted value error for the latent and score perturbation for the positional key. Without training, RAQ-MLA quantizes the full MLA cache to about 2.5 bits per element, a 6.5x increase in theoretical KV-cache capacity, while remaining nearly lossless on reasoning and coding tasks across four MLA models.

cs.AI↗

arXiv2026

NormTune: Norm-Guided Parameter-Efficient Fine-Tuning for Reasoning-Intensive Tasks

Full fine-tuning has achieved remarkable success on reasoning-intensive tasks, but it often requires substantial GPU resource consumption. Parameter-efficient fine-tuning (PEFT) provides a promising alternative for reducing this cost, yet we observe a substantial performance gap on reasoning-intensive tasks. In this paper, we present three key observations showing that this gap is closely associated with norm-related under-adaptation, where PEFT produces markedly smaller weight-update norms than full fine-tuning. Based on these observations, we propose Norm-Guided Parameter-Efficient Fine-Tuning (NormTune), a simple yet effective strategy that adaptively strengthens insufficiently updated layers according to their weight-update norms. Extensive experiments show that NormTune consistently enhances multiple PEFT methods across challenging reasoning benchmarks. Specifically, when trained on S1K-1.1, integrating NormTune into rank-8 LoRA, Bone, and MoRA yields substantial average accuracy improvements of 24.24\%, 11.33\%, and 16.21\% , respectively, across AIME 2024, MATH 500, GPQA, and AIME 2025. Moreover, NormTune-enhanced rank-256 LoRA, Bone, and MoRA achieve performance comparable to S1 full fine-tuning, with respective average gains of 2.17\%, 0.93\%, and 3.00\% over this baseline.

cs.AI↗

arXiv2026

Benchmarking the Limits of In-Context Reinforcement Learning for Ad-Hoc Teamwork

In-Context Reinforcement Learning (ICRL) has enabled foundation agents to adapt instantaneously to novel tasks, yet its efficacy in Ad-Hoc Teamwork (AHT)-where coordination with unknown partners is required-remains unexplored. To rigorously evaluate this, we introduce a large-scale benchmark ICRL4AHT, built upon a high-throughput JAX implementation of Overcooked-V2. Our benchmark includes a large, diverse teammate suite spanning both RL and heuristic policies, enabling controlled train-test shifts, and provides a reproducible end-to-end pipeline for teammate generation, learning-history collection, dataset construction, and online multi-episode evaluation. We evaluate representative history-conditioned ICRL algorithms, including Algorithm Distillation (AD) and Decision-Pretrained Transformer (DPT), across millions of transitions. Results reveal notable limitations: contrary to their success in single-agent domains, these baselines fail to exhibit robust test-time adaptation in multi-agent settings. Specifically, these methods frequently underperform random baselines across both unseen teammate and unseen layout tracks, with no clear in-context improvement over long horizons. These findings highlight the challenges of strategic inference under partial observability within the OvercookedV2 AHT protocol, establishing our benchmark as a critical testbed for next-generation coordination algorithms.

cs.AI↗

arXiv2026

TernaryLLM: Ternarized Large Language Model

Large language models (LLMs) have achieved remarkable performance on Natural Language Processing (NLP) tasks, but they are hindered by high computational costs and memory requirements. Ternarization, an extreme form of quantization, offers a solution by reducing memory usage and enabling energy-efficient floating-point additions. However, applying ternarization to LLMs faces challenges stemming from outliers in both weights and activations. In this work, observing asymmetric outliers and non-zero means in weights, we introduce Dual Learnable Ternarization (DLT), which enables both scales and shifts to be learnable. We also propose Outlier-Friendly Feature Knowledge Distillation (OFF) to recover the information lost in extremely low-bit quantization. The proposed OFF can incorporate semantic information and is insensitive to outliers. At the core of OFF is maximizing the mutual information between features in ternarized and floating-point models using cosine similarity. Extensive experiments demonstrate that our TernaryLLM surpasses previous low-bit quantization methods on the standard text generation and zero-shot benchmarks for different LLM families. Specifically, for one of the most powerful open-source models, LLaMA-3, our approach (W1.58A16) outperforms the previous state-of-the-art method (W2A16) by 5.8 in terms of perplexity on C4 and by 8.2% in terms of average accuracy on zero-shot tasks.

cs.LG↗

arXiv2026

GeoScaffold: Learning Compact Geometric Latents via Reconstruction for Efficient Vision-Language Navigation

Recent vision-and-language navigation (VLN) systems increasingly adopt streaming Video-LLM policies that map egocentric RGB observations and instructions directly to low-level actions. Yet these policies inherit weak 3D geometric priors from 2D pretraining. Existing geometry-aware extensions charge a persistent inference-time price: depth sensors, 3D encoders, or per-step perception tool calls. We propose GeoScaffold, a geometric supervision framework that pays this price once, at training time, by internalizing geometry into the policy itself. It first learns a compact depth tokenizer on depth maps from the training trajectories and freezes it. It then fine-tunes the policy with a handful of learnable geometry query tokens, training their hidden states to reconstruct navigation-critical geometry such as depth, connectivity, and traversability. This supervision turns the query states into compact geometric latents for action decoding, and through the shared weights also internalizes geometry into the backbone's own representations. Like a scaffold, the tokenizer, target generators, and reconstruction heads are discarded after training, leaving the backbone and action interface unchanged. Extensive experiments show that GeoScaffold consistently outperforms leading vision-only navigators on continuous VLN benchmarks, offering a practical paradigm for lightweight edge deployment of spatially aware embodied navigation models.

cs.RO↗

arXiv2026

SymbolicArena: A Unified Infrastructure for Benchmark Distillation and Dynamic Evaluation in Symbolic Regression

Symbolic regression (SR) seeks concise and interpretable mathematical expressions from data for scientific equation discovery. Existing SR benchmarks face a tradeoff between evaluation cost and benchmark validity. Repeated evaluation of large task pools is expensive, and compact benchmarks lack systematic evidence of preserved task diversity and algorithm discriminability. SymbolicArena provides a unified infrastructure for benchmark distillation and dynamic evaluation. The framework standardizes 664 heterogeneous tasks with executable ground truth expressions and distills the Full Task Set into Core50, a validated benchmark of 50 tasks. The distillation process preserves task coverage and algorithm discrimination under explicit balance constraints. SymbolicArena applies a unified execution protocol to heterogeneous SR algorithms and produces comparable outputs and search trajectories. Multi Axis Evaluation characterizes numerical quality, symbolic quality, and search behavior. Core50 reduces evaluation workload by 92.5% and maintains agreement with Full Task Set evaluations. Experiments show that SymbolicArena achieves 72.6% to 86.7% lower approximation error than alternative selectors, further supporting its fidelity to the Full Task Set. Evaluation reveals a substantial gap between numerical fitting and symbolic recovery across current SR methods, suggesting that reliable equation recovery remains an open challenge.

cs.LG↗

arXiv2026

Not Every Term Adds New Structure: Sobolev Novelty for Symbolic Regression

Symbolic regression (SR) aims to discover compact and meaningful mathematical equations from data, but searching the vast combinatorial space of symbolic structures remains challenging. Existing methods typically guide this process using expression-level objectives, such as fitting error, which assess a candidate equation as a whole but provide little information about whether an individual term contributes genuinely new structure or is largely redundant with the rest of the expression. We introduce \textbf{Sobolev Novelty}, a term-level measure of structural independence for symbolic equations. For each term, we construct an empirical Sobolev signature from its function values and exact derivatives over the observed inputs, and quantify how much of this behavior cannot be reconstructed by the remaining terms. We further derive a theory-calibrated threshold, yielding a principled and tuning-free criterion for identifying structurally novel terms. Using this threshold, 92.6\% of terms in benchmark ground-truth equations exhibit sufficient structural novelty, compared with only 38.3\% on average for expressions produced by 15 SR methods, revealing a substantial gap between scientific equations and current SR solutions. As a lightweight plug-in, Sobolev Novelty can be incorporated into diverse SR paradigms to support term pruning, search guidance, LLM feedback, and data selection, yielding consistent performance gains and demonstrating broad applicability.

cs.LG↗

arXiv2026

SPO: Discovering Adaptive Large Neighborhood Search Operators via Stackelberg Program Optimization

Large neighborhood search (LNS) relies critically on destroy and repair operators, whose effectiveness depends on both adaptation to the evolving LNS state and interaction between the two roles. We introduce Stackelberg Program Optimization (SPO), an LLM-based framework for discovering adaptive executable destroy-repair programs. SPO conditions operator decisions on a compact LNS state, allowing state-dependent behavior to emerge through program discovery, and organizes destroy-repair discovery as a Stackelberg interaction over program space that reflects their asymmetric dependency. Role-specific credits evaluate destroy programs as leaders and repair programs as conditional follower responses, guiding a coupled optimization process that combines LLM generator learning with population-based evolutionary search over programs. Experiments on the traveling salesperson problem and capacitated vehicle routing problem show that SPO outperforms strong baselines across a broad range of settings and generalizes beyond the discovery scale to larger instances and benchmark sets. Behavioral analyses further demonstrate state-dependent operator behavior and coupled destroy-repair improvement during discovery.

cs.AI↗

arXiv2026

You Only Need 2/3 of the Chosen Experts: An Empirical Study of Dynamic Expert Pruning in Fine-Grained MoE LLMs

Fine-grained mixture-of-experts (MoE) architectures have become a mainstream design for open-weight LLMs, with hundreds of experts and increasingly many selected per token. This shift makes dynamic expert pruning an attractive route to cheaper inference. Yet existing evidence comes largely from coarser architectures and likelihood-scored multiple-choice benchmarks, leaving three central questions open in the fine-grained regime: how redundant per-token expert selection is, how effectively existing pruning methods exploit that redundancy, and what governs a model's sensitivity to pruning. We fill this gap with a systematic empirical study of twelve fine-grained MoE checkpoints spanning nine architecture families, with a core suite of eleven benchmarks covering knowledge QA, mathematics, code generation, and general reasoning. We find that expert selection is far more redundant than the field's operating points assume: uniformly retaining about two thirds of the selected experts preserves 98.8% of unpruned performance on average, requiring only a one-integer change and delivering 1.2-1.7x measured speedup across two serving backends. This simple baseline leaves little room for dynamic allocation at conservative budgets: even the best published rules differ from it by under 1% at matched expert budgets. Their value emerges under aggressive pruning, where the best rules recover up to 3.0% over uniform truncation, with gains concentrated in the generative tasks that suffer the sharpest degradation. Sensitivity to aggressive pruning also depends on the model: larger and thinking models are more resilient, whereas multimodal models are more vulnerable. Together, these findings reveal how much expert computation fine-grained MoEs can dispense with, and establish when dynamic allocation earns its complexity, informing both practical deployment and future pruning methods.

cs.LG↗

arXiv2026

Train Where the Quantized Model Goes: On-Policy Distillation for Low-Bit Reasoning

Quantization-aware distillation (QAD) restores much of the short-form question-answering performance lost to sub-3-bit quantization, yet leaves mathematical and code reasoning substantially impaired. Long generations often degenerate into repetitive loops, exhausting the decoding budget without completing a solution. We trace this gap to quantization-amplified exposure bias: QAD trains on fixed corpus prefixes, while quantization-induced deviations compound along the model's own autoregressive trajectories. To address this mismatch, we introduce an on-policy distillation (OPD) stage that places teacher supervision where the quantized model actually goes. Starting from a QAD checkpoint, the student generates through the quantized forward path used at deployment and receives feedback from a frozen full-precision teacher on its own prefixes, combining dense token-level guidance with task-verifier rewards. Across four models at 2.79 and 1.88 effective bits, OPD raises average BF16 performance retention from 35% to 70% on MATH-500 and from 66% to 91% on HumanEval while preserving short-form performance, with reasoning gains substantially exceeding those of continued teacher-forced QAD in matched-budget comparisons. By coupling QAD's stable low-bit initialization with OPD's on-policy reasoning recovery, our framework provides a comprehensive sub-3-bit solution that preserves broad capabilities while restoring long-form reasoning.

cs.LG↗

arXiv2026

From Splats to Silicon: Rethinking Computational Efficiency of 3DGS

3D Gaussian splatting (3DGS) represents scenes with explicit primitives and supports real-time novel-view synthesis, yet its system efficiency varies substantially across scenes, viewpoints, rendering paths, and platform constraints. Existing studies pursue efficiency through representation and algorithm design, GPU runtime optimization, and architectural support, but their reported gains correspond to different points along the rendering and update paths. Connecting these indicators to end-to-end system benefit requires tracing how each optimization changes Gaussian selection, screen-space work, data movement, and stage or frame time. We therefore use a workload-centric framework to connect representation and algorithm research, GPU runtimes, and hardware architectures and to identify recurring workload patterns. We complement literature analysis with reproduced measurements and controlled GPU profiling of selected implementations, relating workload counts to stage time and memory traffic. Together, these comparisons show that system gains depend on workload reductions reaching downstream execution, granularity matching each stage, and the cost of data transfers, synchronization, and cached results, gradients, and optimizer data. Building on these findings, we discuss more consistent evaluation under rendering-quality constraints and identify key directions for future system design.

cs.AR↗

arXiv2026

Mind-VLA: Instruction-Aware Spatial Representation Alignment for Vision-Language-Action Models

Recent Vision-Language-Action (VLA) methods improve generalization by aligning their representations with 3D scene geometry. However, these methods are fundamentally instruction-agnostic: the representations align the entire scene uniformly, neglecting the 3D geometry of the specific target object designated by the language instruction. This causes failures on fine-grained manipulation and target occlusion tasks, where success depends on accurate 3D understanding of the target object rather than the entire scene. To address this, we present Mind-VLA, an instruction-aware spatial representation alignment method for VLA models. Specifically, Mind-VLA first obtains the target object specified by the language instruction, then prepares its canonical target views and extracts the corresponding VAE and VGGT features. Finally, the latent representation of the VLA model is aligned with these features to enable instruction-aware 3D understanding. Mind-VLA reaches 94.4% on LIBERO and 4.47 on CALVIN with a compact 345M-parameter backbone. On real-robot tasks with target occlusion, Mind-VLA reaches 54% average success, outperforming the matched scene-VGGT control by 26 percentage points.

cs.RO↗

arXiv2026

Ban&Pick: Enhancing Performance and Efficiency of MoE-LLMs via Smarter Routing

Sparse Mixture-of-Experts (MoE) has become a key architecture for scaling large language models (LLMs) efficiently. Recent fine-grained MoE designs introduce hundreds of experts per layer, with multiple experts activated per token, enabling stronger specialization. However, during pre-training, routers are optimized mainly for stability and robustness: they converge prematurely and enforce balanced usage, limiting the full potential of model performance and efficiency at inference. In this work, we uncover two overlooked issues: (i) a few highly influential experts are underutilized due to premature and balanced routing decisions; and (ii) enforcing a fixed number of active experts per token introduces substantial redundancy. Instead of retraining models or redesigning MoE architectures, we introduce Ban&Pick, a post-training, plug-and-play strategy for smarter routing. Pick discovers and reinforces key experts-a small group with outsized impact on performance-leading to notable accuracy gains across domains. Ban further dynamically prunes redundant experts based on layer and token sensitivity, delivering faster inference with minimal accuracy loss. Experiments on fine-grained MoE-LLMs (DeepSeek, Qwen3) across math, code, and general reasoning benchmarks demonstrate that Ban\&Pick delivers free performance gains and inference acceleration without retraining or architectural changes. For instance, on Qwen3-30B-A3B, it improves accuracy from 80.67 to 84.66 on AIME2024 and from 65.66 to 68.18 on GPQA-Diamond, while accelerating inference by 1.25x under the vLLM.

cs.LG↗

arXiv2026

MODE: Modality-Decomposed Expert-Level Mixed-Precision Quantization for MoE Multimodal LLMs

Mixture-of-Experts Multimodal Large Language Models (MoE-MLLMs) offer remarkable performance but incur prohibitive GPU memory costs, making compression essential. Among PTQ methods, expert-level mixed-precision quantization has proven effective for MoE-LLMs, yet suffers notable degradation on MoE-MLLMs due to two overlooked biases in expert importance estimation. (1) At the cross-modal level, the numerical dominance of vision tokens causes expert selection frequency to be dominated by vision tokens, masking experts that are critical to the text modality; (2) at the intra-vision level, the large proportion of redundant vision tokens further skew frequency statistics, obscuring experts critical for informative visual content. To bridge gaps, we propose MODE, a modality-decomposed expert-level mixed-precision quantization framework for MoE-MLLMs that decomposes expert selection frequency by modality, filters redundant vision tokens to obtain denoised visual frequency, and further evaluates quantization sensitivity per modality as a complementary signal to frequency-based estimation. These signals are integrated into an Integer Linear Programming formulation to assign per-expert bit-widths under a given budget. Extensive experiments show that MODE is particularly well-suited for MoE-MLLMs, limiting average performance loss to within 2.9% at W3A16, with larger gains at the extreme 2-bit setting.

cs.LG↗

arXiv2026

HiViS: Hiding Visual Tokens from the Drafter for Speculative Decoding in Vision-Language Models

Speculative decoding has proven effective for accelerating inference in Large Language Models (LLMs), yet its extension to Vision-Language Models (VLMs) remains limited by the computational burden and semantic inconsistency introduced by visual tokens. Recent studies reveal that visual tokens in large VLMs are highly redundant, and most of them can be removed without compromising generation quality. Motivated by this observation, we propose HiViS (Hiding Visual Tokens from the Drafter for Speculative Decoding in Vision-Language Models), a framework that utilizes the target VLM as a semantic fusion model, allowing the drafter to obtain visual information without explicitly processing visual tokens, ensuring that the drafter's prefill sequence length matches that of the textual tokens. Furthermore, HiViS employs a time-step-aware aligned training scheme that allows the drafter to autonomously propagate and refine instructive visual-textual semantics during independent drafting, guided by step-dependent bias-correction residuals. Extensive experiments across representative VLMs and benchmarks demonstrate that HiViS achieves significant improvements in average acceptance length and speedup ratio.

cs.LG↗

arXiv2026

DeGS: A Scalable 3DGS Architecture via Decoupled Workload Parsing and Reorganization

3D Gaussian Splatting (3DGS) has emerged as a leading technique for real-time novel view synthesis, yet existing 3DGS accelerators suffer from poor architectural scalability: increasing the number of PEs leads to marginal performance improvement during rendering. We identify that the root cause is the tightly coupled ``checking-while-blending'' dataflow, which exacerbates PE underutilization caused by spatial redundancy from irregular Gaussian coverage and temporal redundancy from asynchronous pixel-wise termination under parallel execution. To address this issue, we propose DeGS, a scalable architecture for efficient 3DGS inference. To systematically eliminate the redundancies inherent in rendering, DeGS exploits a decoupled dataflow, restructuring the coupled $α$-checking, transmittance checking, and $α$-blending of the standard rendering process into consecutive workload parsing, reorganization, and blending stages. This allows the fragmented, length-variable, and temporal-dependent workloads to be reorganized into compact, conflict-free, and dense workloads prior to blending, thereby significantly improving PE utilization during parallel blending. Implemented in 28 nm technology, DeGS achieves 2.36$\times$--7.25$\times$ throughput, 1.82$\times$--6.02$\times$ end-to-end speedup, and 1.59$\times$--4.42$\times$ energy efficiency over state-of-the-art 3DGS accelerators (GSCore, GBU, GCC) across diverse scenes and resolutions (720p to 8K). Moreover, scaling from 16 to 1024 PEs, DeGS maintains over 80\% PE utilization at high resolutions, significantly outperforming existing accelerators.

cs.AR↗