arXiv subjects

Heng Ji

Publications and source records attributed to Heng Ji.

At least 19 recordsLinked to original sources

arXiv2026

Energy-Based Reasoners: Your Recurrent Reasoner Should Be an Energy-Based Model

Looping is a powerful inductive bias for reasoning. By repeatedly iterating a recurrent block, looped models can scale their effective depth and enable multi-step reasoning in latent space. However, existing looped models are not inherently guided by an explicit measure of progress, making it difficult to determine when to stop looping and how to improve upon the current state. To enable this, we introduce Energy-Based Reasoners (EBRs), which learn an energy function over latent reasoning states and formulate looped updates as gradient descent on this energy. The energy scores latent states, while its gradient guides refinement and provides a signal for stopping. To train EBRs effectively, we further introduce exploration-based training to discover additional modes and to better shape the energy landscape over a broader set of states. Across four algorithmic reasoning benchmarks, EBRs match or outperform prior recurrent reasoning models, achieving 99.9% accuracy on Sudoku-Extreme while using 345$\times$ less inference compute than the next-best model and maintaining comparable training compute. On all multi-solution benchmarks, EBRs achieve state-of-the-art coverage of distinct valid solutions while maintaining near-100% single-sample accuracy. These results suggest that energy-based models are a promising approach for strong reasoning performance.

cs.LG↗

arXiv2026

Mixture of Layers: Dynamic Layer Routing for Visual Reasoning

Pre-trained vision encoders contain layer-wise visual representations that differ in spatial granularity, semantic abstraction, and sensitivity to local details. However, most Multimodal Large Language Models (MLLMs) rely on only the final or penultimate vision encoder representations or fixed aggregation rules, making visual abstraction largely query-agnostic and limiting access to fine-grained cues such as small objects, spatial details, text, and subtle visual attributes. In this work, we propose Mixture of Layers (MoL), an instruction-conditioned layer routing approach at the visual patch level that dynamically aggregates query-relevant latent representations from intermediate vision encoder layers. Given a text query, MoL predicts routing probabilities over vision encoder layers and performs a top-k sparse aggregation over selected hidden states at either the image level, patch level, or through a hybrid routing mechanism. In doing so, MoL enables query-adaptive access to layer-specific visual features for fine-grained visual reasoning. Our experiments across 7 fine-grained visual reasoning tasks demonstrate substantial performance improvements, especially across fine-grained visual grounding and understanding tasks such as +18.9% improvement on V* in overall accuracy, +4.5% on HRBench4K, and +16.3% on CharXiv compared to the baseline MLLMs, without resorting to multi-resolution inputs, simple interleaving of multiple vision encoders, or increasing the number of patch tokens. We study vision encoders' receptive field scales across different layers and their sampling behaviors to provide an in-depth analysis of why layer-wise sampling is helpful, demonstrating that conditional visual representations are a key step towards better visual perception and reasoning in MLLMs. Our project page is available at https://wjdghks950.github.io/mol.github.io/.

cs.CV↗

arXiv2026

Generalizable VLA Finetuning via Representation Anchoring and Language-Action Alignment

Finetuning a pretrained vision-language model (VLM) on robot demonstrations via behavior cloning (BC) has become the standard recipe for vision-language-action (VLA) policies. However, BC finetuning progressively overwrites the pretrained representations that support visual and semantic generalization. Co-training on web image-text data, a common remedy, applies language and action losses to separate observations, leaving VLAs with language-action misalignment that standard manipulation benchmarks do not expose. We propose Anchor-Align, which augments BC with two objectives: Vision-Language Anchoring distills layer-wise representations from a frozen VLM copy to prevent this drift, while Language-Action Alignment converts each action target into a discrete motion-direction label and jointly trains language and action prediction on the same robot observation. We conduct real-world evaluations across eight manipulation settings on single-arm xArm7 and bimanual YAM robots, using two VLA architectures with regression and flow-matching action heads. Across these settings, Anchor-Align consistently improves over BC on novel targets, layouts, and motion-sensitive bimanual tasks requiring coordinated control. At scale in simulation, we demonstrate consistent improvements on OOD perturbations, perceptual robustness, and long-horizon control across LIBERO-PRO, LIBERO-Plus, and CALVIN, respectively, suggesting that preserving pretrained representations and effective action learning are not fundamentally at odds. Project page: anchoralignvla.github.io

cs.RO↗

arXiv2026

Enhancing Software Engineering Through Closed-Loop Memory Optimization

Large language models (LLMs) have enabled powerful software engineering (SE) agents capable of navigating complex codebases and resolving real-world issues. However, these agents remain fundamentally episodic: they fail to retain, refine, and reuse experiences across tasks, repeatedly reconstructing context from scratch and reproducing similar mistakes. Even with memory support, they offer no remedy for the absence of a principled, task-agnostic \textit{memory utility}, making them difficult to evaluate rigorously or generalize across agents and settings. To tackle these limitations, we introduce MemOp, a closed-loop framework for memory augmentation in SE agents. MemOp grounds memory utility in \textit{validated downstream impact}, establishing utility as both a task-agnostic \textbf{evaluation benchmark} and an annotation-free \textbf{optimization signal}. Evaluating MemOp across four benchmarks, three programming languages, and two SE agents, under complementary \textit{single-episode} and \textit{cross-episode} memory augmentation, results demonstrate the \textbf{effectiveness} of MemOp, achieving absolute gains of up to $\uparrow5.25\%$ in success rate and $\uparrow4.63\%$ in resolve efficiency, while notably reducing deployment cost by $\geq10.33\%$. MemOp further exhibits \textbf{generalizability} across SE agents, memory backbones, and optimization algorithms, together with \textbf{transferability} of MemOp-trained memory model to unseen programming languages and SE activities. Our project page: https://xhguo7.github.io/MemOp/.

cs.SE↗

arXiv2026

Free Lunch for LLMs Confidence: The RL Critic Already Knows

Reliable confidence estimation is essential for trustworthy deployment of large language models (LLMs), but dedicated confidence training stage requires additional data and compute. We ask a fundamental question: what prediction problem does confidence training actually solve? Inspired by prior work on external confidence estimation, we compare self- and external estimators of comparable capacity trained on the same (question, response, correctness) triples. We show that an external model trained solely on a policy's behavioral traces can estimate the policy's confidence as accurately as the policy itself, without accessing the policy's internal representations. This points to a deeper connection with value function learning in reinforcement learning (RL): under a correctness-based outcome, we prove that confidence estimation and critic training are equivalent, both predicting the likelihood that a policy's response will be correct from its trajectory. A simple consequence follows: for LLMs trained with value-based RL, the existing critic can be reused as an off-the-shelf confidence estimator with no additional training or data. The policy remains unchanged, so its task performance is preserved, and the critic provides confidence throughout generation, including intermediate reasoning steps. Across math, code, tool calling, and general reasoning spanning a dozen domains, the reused critic is the best confidence estimator on every in-distribution split and the best-calibrated on 14 of 16, without hurting the task performance.

cs.CL↗

arXiv2026

JevAgentHarness: Benchmarking, Deploying and Training Decision Models That Share Control with LLMs

Large language model (LLM) agents now fix software, operate computers and serve customers, but their harnesses, the programs that run them, call a large LLM at every step, which makes long tasks slow and costly. Most of these calls end in a multiple-choice decision, such as which tool to call next. Decision models such as Jev answer these questions directly, at a fraction of an LLM call's time and price. We ask how a harness can delegate these decisions to a decision model with little loss of success. To evaluate decision models in agentic tasks, we build AgentDecisionBench, which tests them from isolated questions up to control of a running agent, on public benchmarks and our DecisionGym environment. The decision model most accurate on isolated questions solved only 69.5% of agent tasks, and without a fitted confidence threshold, delegating tool choices lowered success by 19.0 points. Building on these findings, we then design JevAgentHarness, which fits this threshold for each choice on the agent's own trajectories and lets the LLM decide below it. Because existing decision models never see the states an agent reaches after its own mistakes, we train AgentDecision-4B, an open 4B model, on such states with Jev's answers as labels. On tasks unlike its training data, the agent solved 91.5% of them with AgentDecision-4B and 82.0% with Jev. On tasks like their training data, 4B models trained this way made agents 81.1-89.0% cheaper and up to 2.5 times faster, without losing 3 or more points of task success. The models, data, and harness will be released at https://github.com/JevAgentHarness/JevAgentHarness.

cs.AI↗

arXiv2026

Evolutionary Informed Translation from Protein to Nucleic Acid Sequence

The choice of nucleic acid sequence for a given protein strongly impacts function and is species-specific. Current protein-to-nucleic-acid translation models treat sequences in isolation, often over-favoring frequently used codons. To generate more natural nucleic acid sequences, we introduce PINATA (Phylogenetically Informed protein to Nucleic Acid Translation with sequence Alignment), a model that leverages sequences from related species encoded in multiple sequence alignments (MSAs) to guide prediction. Our architecture leverages evolutionary context and learns species-and protein specific context to make biologically-informed predictions. Evaluated across 61,934 gene trees and 882 species, PINATA generates sequences that mirror the naturally-occurring sequence substantially better than existing language models. PINATA leverages MSA data to improve exact native codon recovery by 17.95 percentage points over existing language models, effectively preserves natural rare-codon nuances, and more closely matches native sequence properties. PINATA maintains strong predictive performance even with limited evolutionary data. While previous methods largely treated sequences from different species in isolation, our method leverages the evolutionary relationships between nucleic acids and proteins across species to create a more biologically-informed translation. We demonstrate that leveraging rich evolutionary data through an biologically-designed method strongly aids in translating protein sequences to nucleic acid sequences across species. We will make data and code publicly available upon publication.

q-bio.BM↗

arXiv2026

PixSearch: Region-Grounded Retrieval for Knowledgeable Large Multimodal Models

Multimodal Retrieval-Augmented Generation (MM-RAG) augments visual question answering (VQA) with external knowledge, but many systems rely on modular pipelines that separate visual grounding, query construction, and answer generation. This separation can propagate errors from intermediate detections, masks, or captions into retrieval while adding latency and limiting coordination with ongoing reasoning. These limitations are especially consequential for egocentric images, where questions often concern small, off-center, or long-tail entities that global retrieval can overlook. We propose PixSearch, an end-to-end segmenting Large Multimodal Model (LMM) that unifies region-level visual grounding with retrieval-augmented reasoning. During autoregressive generation, PixSearch learns when to retrieve, how to route queries across text, whole-image, and region-level retrieval, and how to crop the target entity as a visual query using a jointly trained mask decoder. A two-stage supervised fine-tuning regimen with search-interleaved supervision teaches retrieval triggering, query selection, and evidence-grounded answer generation while preserving segmentation ability. On egocentric and entity-centric VQA benchmarks, PixSearch achieves an 18.3% relative accuracy gain on CRAG-MM over whole-image retrieval, with particularly strong gains on egocentric images, while remaining competitive on broader VQA and text-only QA tasks. It also reduces inference latency by 36.4% compared with a cached GroundedSAM-LLaVA pipeline.

cs.CV↗

arXiv2026

When Do Language Models Add Value? Relative Competence in Hybrid Event Forecasting

When a forecasting system already has a prediction from a market, crowd, or statistical model, when should it use a language model? Standalone accuracy cannot answer this question because a less accurate model may still improve the existing forecast. We study this distinction through relative competence, separating a model's replacement advantage from its complementary value. Under Brier loss (squared error between predicted probabilities and binary outcomes), we characterize when model disagreement can correct external forecast errors and derive the population gain from regime-specific affine pooling. We introduce a competence gate that estimates source weights from resolved outcomes, shrinks regime-specific estimates toward a global solution, and chronologically recalibrates the pooled forecast. Across 4,974 resolved binary questions and four language models, the best standalone model has 24.94% higher Brier loss than the calibrated external forecast, yet the gate improves that reference by 3.35% (95% CI, 1.80 to 4.90%). Model contributions vary across forecasting regimes, with larger corrections in thinner markets and forecasting crowds and smaller corrections in higher-participation settings. The gate also shows no statistically detectable performance difference from more flexible contextual aggregators. These results demonstrate that standalone superiority is not necessary for complementary predictive value and that language models should be evaluated relative to the forecasts already available.

cs.AI↗

arXiv2026

Measuring Combinatorial Creativity in Large Language Models and Humans

Combinatorial creativity, the process of combining familiar ideas in unfamiliar ways, is one of the foremost engines of scientific, technological, and artistic progress. This work introduces Kombine, an LLM benchmark for combinatorial creativity tasking models to discover surprising, original, and useful associations, analogies, and blends between unrelated entities, and to use those analogies and blends to invent novel concepts by re-combining the parts of existing ones. In a large-scale study across 35 LLMs, we find that: (1) LLMs are 10-16 percentage points better at discovering associations and analogies than blends, and (2) exhibit an abstraction bottleneck on blending, in which even frontier LLMs struggle to find a shared abstraction between inputs and instantiate a coherent blended space. (3) LLMs also exhibit inventive homogeneity, in which one in five concepts invented for a given item reuses properties invented by another model. Moreover, (4) models are $14\times$ more likely to reuse other invented properties on blends than analogies, and (5) models from the same provider are $2.3\times$ more likely to reuse properties than those from different providers. Next, in a comparative study involving 75 humans on the same tasks, we find that (6) humans score below nearly every LLM on each task and (7) both humans and LLMs find blending most difficult, but association is easier than analogy for LLMs, while analogy is easier than association for humans. Lastly, (8) humans and LLMs access distinct regions of semantic space on each task, suggesting that combinatorial creativity could benefit from systematic human-AI collaboration to enhance conceptual diversity. In summary, we provide a framework to assess progress towards LLMs, AI systems, and human-AI collaborations more broadly, that can elicit combinatorial creativity in a setting relevant to scientific and technological innovation.

cs.AI↗

arXiv2026

LMBuild: Evaluating LLM Agents for Generating Buildable and Functional Structures

LLM-based agents are increasingly capable of generating complex 3D structures, with the potential to reshape how objects are designed and realized in the physical world. Yet, producing elegant geometry is fundamentally different from producing objects that can be built and perform their intended functions. Existing evaluations largely focus on geometric quality while overlooking physical realizability. We introduce LMBuild, a benchmark for evaluating LLM agents on generating buildable and functional structures. LMBuild represents generated objects as assembled structures comprising part decompositions, joints, materials, and sequences. To support reproducible evaluation, we provide a unified framework consisting of: (1) an interactive environment in which agents can use tools to retrieve, create, and place components to construct objects; (2) a curated benchmark that repurposes established CAD datasets and augments them with knowledge from Wikipedia; and (3) a evaluation framework covering structural soundness, functional affordance, design quality, and physical realization. Evaluations across 30 systems reveal several intriguing findings: (a) Soundness and alignment are no longer the primary bottlenecks for frontier closed-source models, while functional affordance and physical operability remain substantially more challenging; (b) stronger models more effectively create new components, whereas weaker models tend to rely on retrieval; and (c) providing functional specifications substantially improves part completeness, kinematics, and physical operability. These results show that generating real-world structures requires deeper reasoning about functional affordances, mechanics, and designing and creating novel components. We expect LMBuild to provide a foundation for measuring progress and incentivizing research toward agents that generate buildable and functional structures.

cs.AI↗

arXiv2026

VISPA: Pluralistic Alignment via Automatic Value Selection and Activation

As large language models are increasingly used in high-stakes domains, it is essential that their outputs reflect not average} human preference, rather range of varying perspectives. Achieving such pluralism, however, remains challenging. Existing approaches consider limited values or rely on prompt-level interventions, lacking value control and representation. To address this, we introduce VISPA, a training-free pluralistic alignment framework, that enables direct control over value expression by dynamic selection and internal model activation steering. Across extensive empirical studies spanning multiple models and evaluation settings, we show VISPA is performant across all pluralistic alignment modes in healthcare and beyond. Further analysis reveals VISPA is adaptable with different steering initiations, model, and/or values. These results suggest that pluralistic alignment can be achieved through internal activation mechanisms, offering a scalable path toward language models that serves all.

cs.CL↗

arXiv2026

Learning Meta-Skills for Agent Harness Design in Test-Time AI4AI

Agent performance depends on both reasoning ability and the environment in which it acts. We study test-time AI-for-AI, asking how a Builder can learn to construct better execution environments for a Target while both models' weights remain fixed. To make the Builder's experience reusable, we introduce Meta-Skill: principles specifying when support is needed and what resources to provide. The Builder learns these principles from Target's execution feedback on the development set, then uses the frozen skill bank to construct harnesses for unseen tasks. Across Harness-Bench and NewtonBench, full-bank meta-skills improve macro-average performance by 8.95 percentage points over no-skill construction, and 12.02 points over direct delivery of the same bank to the Target. These results highlight the value of translating experience into executable support. Gains when the same model serves both roles further suggest a path to system level self-improvement through learning to build better environments.

cs.AI↗

arXiv2026

Agent Error Dataset: Scaling 50,000 Error--Diagnosis Pairs for Failure Analysis and Error-Aware Post-Training

An unsuccessful LLM agent rollout contains more information than its final reward: the observations available to the agent, the actions it chose, and the environment's responses. Reusing this experience for learning requires identifying a decision to revise and testing a concrete alternative. We introduce the Agent Error Dataset (AED), comprising 50,228 error-diagnosis pairs from 9,961 source tasks across 33 environments, 19 harness families, and 23 policy models in text-based agent systems. We retain source traces and execution metadata to support cross-setting failure analysis and re-diagnosis without repeating the original rollout. Our five-stage Agentic Error-to-Training (AET) pipeline collects natural failures, generates diagnoses and proposed corrections, and checks them against recorded evidence. Where replay is supported, we compare corrections with original-action retries from the same checkpoint under matched execution settings. We then construct separate training views for diagnosis and actor recovery. Across 3,062 matched replay pairs, first-proposal corrections raise verifier pass rates from 18.4% to 51.1%, a gain of 32.7 percentage points. Using a separately frozen diagnosis release, full-diagnosis fine-tuning on 1,656 source tasks raises Qwen3-8B's exact-step agreement with internal teacher labels from 47.2% to 63.6%, averaged over three seeds on a 943-case holdout. The strongest prompted reference in this comparison scores 54.7%, and mean agreement improves at each of four increasing training-set sizes. In a single-seed comparison of actor-training recipes, action-only repair training scores 6.67 percentage points higher on WebShop-lite than success-only training.

cs.AI↗

arXiv2026

MemGuard: Preventing Memory Contamination in Long-Term Memory-Augmented Large Language Models

Memory-augmented large language models extend reasoning beyond a fixed context window by maintaining long-term memory across interactions. However, existing memory systems often collapse stable user facts, episodic events, and behavioral rules into a shared space, allowing functionally distinct memories to be retrieved and used as interchangeable evidence. We identify this failure mode as heterogeneous memory contamination, where context-specific events become overgeneralized claims, or semantically relevant but functionally incompatible memories mislead generation. To this end, we introduce MemGuard, a type-aware memory framework that preserves functional memory boundaries during memory construction and retrieval. It assigns each memory an explicit functional role at write time, maintains relations across type-isolated memories, and selectively composes evidence only from necessary memory types, reducing contamination from irrelevant or functionally incompatible evidence. Across hallucination and long-horizon conversation benchmarks, MemGuard improves memory reliability by up to 28.27% while retrieving up to 5.8x fewer memory tokens than prior methods. These results suggest that reliable long-term reasoning depends on principled organization and selective use of heterogeneous memory.

cs.CL↗

arXiv2026

Constructive Distortion: Improving MLLMs with Attention-Guided Image Warping

Multimodal large language models (MLLMs) often miss small details and spatial relations in cluttered scenes, leading to errors in fine-grained perceptual grounding. We introduce AttWarp, a lightweight method that allocates more resolution to query-relevant content while compressing less informative areas, all while preserving global context. At test time, the approach uses an MLLM's cross-modal attention to perform rectilinear warping of the input image, reallocating spatial resolution toward regions the model deems important, without changing model weights or architecture. This attention-guided warping preserves all original image information but redistributes it non-uniformly, so small objects and subtle relationships become easier for the same model to read while the global layout remains intact. Across five benchmarks (TextVQA, GQA, DocVQA, POPE, MMMU) and four MLLMs (LLaVA, Qwen-VL, InternVL, and InstructBLIP), AttWarp consistently improves accuracy, strengthens compositional reasoning, and reduces hallucinations, outperforming four competitive baselines that manipulate raw images at test time. Together, these results show that attention-guided warping prioritizes information relevant to the query while preserving context, and that the same MLLMs perform better when given such warped inputs.

cs.CV↗

arXiv2026

Learning to Predict Future-Aligned Research Proposals with Language Models

Large language models (LLMs) are increasingly used to assist ideation in research, but evaluating the quality of LLM-generated research proposals remains difficult: novelty and soundness are hard to measure automatically, and large-scale human evaluation is costly. We propose a verifiable alternative by reframing proposal generation as a time-sliced scientific forecasting problem. Given a research question and inspiring papers available before a cutoff time, the model generates a structured proposal and is evaluated by whether it anticipates research directions that appear in papers published after the time. We operationalize this objective with the Future Alignment Score (FAS), computed via retrieval and LLM-based semantic scoring against a held-out future corpus. To train models, we build a time-consistent dataset of 21,835 paper occurrences across 3,642 instances from targets and their pre-cutoff citations, and synthesize reasoning traces that teach gap identification and inspiration borrowing. Across Llama-3.1 and Qwen2.5 models, future-aligned tuning improves future alignment over unaligned baselines (up to +10.6% overall FAS), and domain-expert human evaluation corroborates improved proposal quality. Finally, we demonstrate practical impact by implementing two model-generated proposals with a code agent, obtaining 4.17% accuracy gain on MATH from a new prompting strategy and consistent improvements for a novel model-merging method. Our code and data are publicly available at https://github.com/Arthur-Heng/future-aligned-proposals.

cs.CL↗

arXiv2026

Random Attention: Rethinking KV Cache Eviction for Efficient Reasoning

Large language models achieve superior performance on tasks that require extended reasoning, but long chains of thought make the KV cache a severe memory bottleneck. Existing KV cache compression methods share one paradigm: score each cached token by some estimate of how much it will matter later, and keep the top-scoring ones. We show that the selection signal contributes almost nothing. Random Attention keeps the prompt and evicts uniformly at random within each attention head, computing no score at all; across four models and six reasoning tasks, it matches the strongest baseline in task performance while delivering 32-43% higher throughput than that method when deployed with vLLM. Controlled experiments explain this by showing that 1) the prompt is the fragile part of the cache, and most of the gap between selectors is just whether their selection signal happened to keep it; 2) the reasoning trace protects itself against eviction with redundancy at two levels, in the text (the model restates what it still needs as it works) and across attention heads (each keeps its own copy of the trace), so once the prompt is safe, a random draw retains enough copies of what the model still needs, and no score is required to pick them. Our code is publicly available at https://github.com/SalesforceAIResearch/Random-Attention.

cs.CL↗