Searcharxiv⌕ Search

arXiv subjects

Song Bian

Publications and source records attributed to Song Bian.

At least 19 recordsLinked to original sources

Training Reasoning Models in Simulation to Natively Use Real Robots

Robot-use agents (RUAs) have emerged as a promising approach to general-purpose robot control. However, strong performance has so far been achieved only by frontier proprietary models such as GPT-6 Astra and Claude Fable 5.2, whereas open-weight models remain substantially less capable. We show that an open-weight 27B reasoning model can be trained to use real robots natively through reinforcement learning, entirely in simulation and without demonstrations. Like a computer-use agent (CUA), the model observes the environment, reasons in a chain of thought, and generates the next end-effector pose, without any harness. We build a Post-training Engine for Agentic Robotics (PEAR) and use it to train the model on simulated robot tasks at scale with robot-use reinforcement learning (RURL), mirroring how frontier models acquire computer use. We find that the skills acquired in simulation generalize well beyond the training tasks: they transfer to an unseen robot, to the real world, and to novel instructions. After training on LIBERO and RoboDojo in simulation, the model succeeds in 60% of our trials on a real dual-arm robot, whereas the base model succeeds in none. On LIBERO, training raises the success rate from 1.6% to 74.4%, approaching GPT-6 Astra and GPT-6.1 Sol at maximum reasoning effort. Compared with $π_{0.5}$, a vision-language-action model fine-tuned on demonstrations, our model is substantially more robust to changes in the task itself. On LIBERO-Pro, $π_{0.5}$ succeeds in only 0.75% of episodes when assigned a new task in a familiar scene whereas our model succeeds in 57.4%. Finally, the model retains its general capabilities, with no measurable degradation on nine reasoning, coding and vision benchmarks. These results highlight robot-use reinforcement learning as a promising component of foundation-model post-training.

cs.RO↗

Evolving Infra: Agent-Adaptive LLM Serving at Scale

Agentic workloads are a major source of LLM inference demand. Existing prefill--decode (PD) disaggregated serving systems offer limited flexibility for runtime adaptation of compute allocation and KV cache management, and rely on coarse signals that only partly capture the diverse and dynamic agent workload. These limitations lead to load imbalance between prefill and decode nodes and eviction of KV caches that agents will reuse. We present Evolving Infra, an agent-adaptive serving framework that adapts compute allocation to agent workloads and KV cache management to agent context management at runtime. We decouple compute allocation from key--value (KV) cache management, allowing nodes to switch efficiently between prefill and decode. Building on this design, an online planner matches the PD split to workload composition, while an agent-aware distributed KV manager guides cache eviction and transfer with signals from agent context management. Across coding, math, and computer-use tasks with GLM-5.2 and Kimi K2.6 on up to 256 GB300 GPUs, Evolving Infra reduces mean end-to-end request latency and P95 TTFT by up to 4.1x and 14x over Dynamo, NVIDIA's production framework. We open-source the system implementation of Evolving Infra at https://github.com/mit-han-lab/EvolvingInfra.

cs.DC↗

Deep Personalization: Continuous Context Observation for Local Proactive Agents

As AI agents become increasingly capable, users must still recognize when and how to ask for help, leaving many opportunities for useful support unrealized. Proactive agents hold the promise to close this gap, but effective proactivity requires understanding users' real-time activities and personalizing intervention timing so that assistance is offered when it is likely to be helpful rather than disruptive. We introduce deep personalization, a closed-loop process that connects continuous context observation with continual agent adaptation: an agent observes user activity, selectively retains long-term context, and adapts intervention decisions from implicit feedback. We initiate deep personalization in Coco, a proactive desktop agent, by having a separate Observer Agent that monitors desktop activity, maintains compact long-term user modeling context, and learns when assistance is desirable from implicit signals such as engagement, dismissal, explicit requests, and repeated work patterns. We further develop efficient on-device strategies to enable the Observer Agent to run on local AI compute platforms without streaming raw activity data off-device. For evaluation at scale without collecting sensitive real-user activity, we introduce a data-synthesis pipeline that transforms existing computer-use trajectories into OSWorld-Persona, a benchmark with persona-conditioned intervention labels. On OSWorld-Persona, deep personalization improves macro F1 across three model backbones from 0.12-0.19 under direct prompting to 0.35-0.48 after adaptation, a 2.4-3.0x relative gain. In a week-long user study, we observe that Coco improves its macro-averaged F1 score from 0.199 to 0.515. These results suggest that on-device deep personalization can make proactive agents more timely and increasingly aligned with each user's everyday work.

cs.LG↗

Reaction Attack on CKKS Deployments with Asymptotically Optimal Query Complexity

We study a protocol-level reaction leakage in CKKS deployments where public coefficient-to-slot (C2S) evaluation keys are combined with a client's 1-bit validity response. Under this deployment surface, a malicious evaluator can recover a ternary secret key using asymptotically optimal $O(H)$ reaction queries, where $H$ is the entropy of the secret key. The attack does not break RLWE or CKKS IND-CPA security. Instead, it exploits the interaction between client-side validity reactions and public homomorphic linear transformations. A shift of a public encryption key yields a ciphertext whose plaintext is a function of the secret key; the public ${\sf C2S}$ layer then repacks this information to construct a ciphertext whose validity depends strictly on whether a targeted secret coefficient matches a guessed value. Furthermore, each submitted query is independently randomized with a fresh encryption of zero, so that each resulting malicious ciphertext appears entirely distinct, thereby allowing the attack to simply evade trivial detection mechanisms on the client side. We implemented the full end-to-end attack in Lattigo v6.2.0 and recovered the full secret key across four parameter sets drawn from standard CKKS presets. For uniform ternary secrets, the attack uses $O(N)$ queries and is within a factor $2/\log_2 3 < 2$ of the oracle-only entropy lower bound. For sparse ternary secrets, a tailored variant further reduces the number of queries in practice. Our results show that observable validity reactions form a practical key-recovery channel in CKKS deployments exposing this surface, even when the underlying cryptographic assumptions remain intact.

cs.CR↗

Humanize: Judgement Engineering for Agentic Coding

Agentic coding makes code generation cheap, but reliable completion remains difficult: the agent that writes the code is a weak judge of whether it is done. We present Humanize, a multi-agent orchestration workflow for agentic coding built around judgement engineering: explicit, mechanically enforced decisions at the boundaries between planning, implementation, review, and learning. A human approves a plan contract, a builder agent implements it in rounds, and a reviewer agent from another vendor decides completion; deterministic hooks, not a model, route work between these roles and enforce 72 mechanical gates. Viewed as a Markov chain over repository states, alternating builder and reviewer samples jointly from two models, so a defect survives only if both miss it. We study Humanize through its deployment, 118 public postmortems of real loops, and its applications. Over 68 versions in 108 days, it gathered 1,468 GitHub stars. Applications include a 567-file gem5 build-system migration under upstream review; Kernel Design Agents, which extend the loop with a kernel knowledge base and profiling feedback and placed in the top three of all three Full-Agent tracks of the MLSys 2026 FlashInfer contest; and, through Humanize Olympiad Agents (HOA), full scores in IOI 2026, IMO 2026, IPhO 2026, and IBO 2024, 418.5/437 in IChO 2026 (gold-medal). Humanize also achieves 672/672 on PutnamBench and ranks first (251/303) on Lean-Eval's leaderboard even competiting with professional mathematicians. The postmortems show that independent review catches unsupported builder claims, but stopping remains a key weakness. In reports that separate rounds by phase, two thirds of rounds occurred after implementation was accepted. This evidence is observational, not a controlled comparison of workflows.

cs.AI↗

When Do We Need On-Policy Distillation? Distilling on Offline Student Rollouts Is Often Better

On-policy distillation (OPD) has become increasingly popular for transferring teacher capabilities to student models. In this work, we ask a critical research question: Is on-policy sampling always beneficial for distilling arbitrary teacher-student pairs? We show that a simple alternative, Semi-OPD, which distills from offline rollouts generated by the initial student, can often outperform OPD in both accuracy and training efficiency. Across 17 teacher-student pairs ranging from 1.5B to 235B parameters, Semi-OPD outperforms OPD in 14 cases, with up to +13.6% accuracy and 11.4x training speedup. We further find that the choice between OPD and Semi-OPD depends on the alignment between the initial teacher and student, quantified by an output-token overlap ratio: OPD is beneficial only when the two are highly aligned with high overlap ratios. Our deeper investigation suggests that effective distillation requires on-policyness w.r.t. both the student and the teacher. For misaligned pairs, student rollouts can become increasingly off-policy w.r.t. the teacher as context length grows, weakening the distillation signal. In contrast, Semi-OPD is often more stable, as it distills on shorter contexts while covering full trajectories and exposing the student to more teacher-preferred tokens. Beyond proposing Semi-OPD as an efficient alternative, our work motivates the community to rethink when to use OPD and to study stronger OPD variants with meaningful teacher-student pairs.

cs.CL↗

Precise Debugging Benchmark: Is Your Model Debugging or Regenerating?

Unlike code completion, debugging requires localizing faults and applying targeted edits. We observe that frontier LLMs often regenerate correct but over-edited solutions during debugging. To evaluate how far LLMs are from precise debugging, we introduce the Precise Debugging Benchmark (PDB) framework, which automatically converts any coding dataset into a debugging benchmark with precision-aware evaluation. PDB generates buggy programs by synthesizing verified atomic bugs and composing them into multi-bug programs. We define two novel metrics, edit-level precision and bug-level recall, which measure how many necessary edits are made and how many bugs are resolved. We release two evaluation benchmarks: PDB-Single-Hard on single-line bugs, and PDB-Multi on multi-line bugs. Experiments show that frontier models, such as GPT-5.1-Codex and DeepSeek-V3.2-Thinking, achieve unit-test pass rates above 76% but exhibit precision below 45%, even when explicitly instructed to perform minimal debugging. Finally, we show that iterative and agentic debugging strategies do not substantially improve precision or recall, highlighting the need to rethink post-training pipelines for coding models.

cs.SE↗

Harbor Adapters and Harbor-Index: Infrastructure and a Curated Meta-Dataset for Large-Scale Agentic Evaluation

Evaluating agents on the growing number of agentic benchmarks is challenging because they often require complex environments and agent integrations. We introduce Harbor Adapters, a unified evaluation infrastructure for agentic benchmarks. Our work makes three contributions. First, we develop benchmark adapters that port more than 80 benchmarks to evaluate arbitrary agents, and validate them through rigorous code review and parity experiments. Second, we conduct a large-scale evaluation of 8 models spanning capability tiers across 54 benchmarks; every model is run with Terminus-2 and with one of 3 native harnesses. This enables a broader analysis of agent capabilities and failure modes than was previously possible. Third, we introduce Harbor-Index, a curated set of 82 difficult, diverse, and high-quality tasks spanning 29 benchmarks, refined from the adapted suite through difficulty filtering, AI and human audit, and an audit-and-fix loop. Harbor-Index preserves the challenge and breadth of large-scale agentic evaluations while being affordable to run; no evaluated model-harness configuration exceeds 30% pass rate, and the strongest (GPT-5.5 with Codex) reaches 28.0%. We release the adapters, evaluation results, in-depth analysis, and Harbor-Index as open-source artifacts to support more reliable and comprehensive evaluation of language-model agents.

cs.AI↗

Scaling Laws Meet Model Architecture: Toward Inference-Efficient LLMs

Scaling the number of parameters and the size of training data has proven to be an effective strategy for improving large language model (LLM) performance. Yet, as these models grow increasingly powerful and widely deployed, the cost of inference has become a pressing concern. Despite its importance, the trade-off between model accuracy and inference efficiency remains underexplored. In this work, we examine how key architectural factors, hidden size, the allocation of parameters between MLP and attention (mlp-to-attention ratio), and grouped-query attention (GQA), influence both inference cost and accuracy. We introduce a conditional scaling law that augments the Chinchilla framework with architectural information, along with a search framework for identifying architectures that are simultaneously inference-efficient and accurate. To validate our approach, we train more than 200 models spanning 80M to 3B parameters and 8B to 100B training tokens, and fit the proposed conditional scaling law. Our results show that the conditional scaling law reliably predicts optimal architectural choices and that the resulting models outperform existing open-source baselines. Under the same training budget, optimized architectures achieve up to 2.1% higher accuracy and 42% greater inference throughput compared to LLaMA-3.2.

cs.LG↗

On the Vulnerability of FHE Computation to Silent Data Corruption

Fully Homomorphic Encryption (FHE) is rapidly emerging as a promising foundation for privacy-preserving cloud services, enabling computation directly on encrypted data. As FHE implementations mature and begin moving toward practical deployment in domains such as secure finance, biomedical analytics, and privacy-preserving AI, a critical question remains insufficiently explored: how reliable is FHE computation on real hardware? This question is especially important because, compared with plaintext computation, FHE incurs much higher computational overhead, making it more susceptible to transient hardware faults. Moreover, data corruptions are likely to remain silent: the FHE service has no access to the underlying plaintext, causing unawareness even though the corresponding decrypted result has already been corrupted. To this end, we conduct a comprehensive evaluation of SDCs in FHE ciphertext computation. Through large-scale fault-injection experiments, we characterize the vulnerability of FHE to transient faults, and through a theoretical analysis of error-propagation behaviors, we gain deeper algorithmic insight into the mechanisms underlying this vulnerability. We further assess the effectiveness of different fault-tolerance mechanisms for mitigating these faults.

cs.CR↗

SoK: Trust-Authorization Mismatch in LLM Agent Interactions

Large Language Models (LLMs) are evolving into autonomous agents capable of executing complex workflows via standardized protocols (e.g., MCP). However, this paradigm shifts control from deterministic code to probabilistic inference, creating a fundamental Trust-Authorization Mismatch: static permissions are structurally decoupled from the agent's fluctuating runtime trustworthiness. In this Systematization of Knowledge (SoK), we survey more than 200 representative papers to categorize the emerging landscape of agent security. We propose the Belief-Intention-Permission (B-I-P) framework as a unifying formal lens. By decomposing agent execution into three distinct stages-Belief Formation, Intent Generation, and Permission Grant-we demonstrate that diverse threats, from prompt injection to tool poisoning, share a common root cause: the desynchronization between dynamic trust states and static authorization boundaries. Using the B-I-P lens, we systematically map existing attacks and defenses and identify critical gaps where current mechanisms fail to bridge this mismatch. Finally, we outline a research agenda for shifting from static Role-Based Access Control (RBAC) to dynamic, risk-adaptive authorization.

cs.CR↗

Terminal-Bench: Benchmarking Agents on Hard, Realistic Tasks in Command Line Interfaces

AI agents may soon become capable of autonomously completing valuable, long-horizon tasks in diverse domains. Current benchmarks either do not measure real-world tasks, or are not sufficiently difficult to meaningfully measure frontier models. To this end, we present Terminal-Bench 2.0: a carefully curated hard benchmark composed of 89 tasks in computer terminal environments inspired by problems from real workflows. Each task features a unique environment, human-written solution, and comprehensive tests for verification. We show that frontier models and agents score less than 65\% on the benchmark and conduct an error analysis to identify areas for model and agent improvement. We publish the dataset and evaluation harness to assist developers and researchers in future work at https://www.tbench.ai/ .

cs.SE↗

What Limits Agentic Systems Efficiency?

Large Language Models (LLMs), such as OpenAI-o1 and DeepSeek-R1, have demonstrated strong reasoning capabilities. To further enhance LLM capabilities, recent agentic systems, such as Deep Research, incorporate web interactions into LLM reasoning to mitigate uncertainties and reduce potential errors. However, existing research predominantly focuses on reasoning performance, often neglecting the efficiency of agentic systems. In this work, we present a comprehensive empirical study that identifies efficiency bottlenecks in web-interactive agentic systems. We decompose end-to-end latency into two primary components: LLM API latency and web environment latency. We conduct a comprehensive empirical study across 15 models and 5 providers to demonstrate high variability in API-based agentic systems. We observe that web environment latency can contribute as much as 53.7% to the overall latency in a web-based agentic system. To improve latency, we propose SpecCache, a caching framework augmented with speculative execution that can reduce web environment overhead. Extensive evaluations on two standard benchmarks show that our approach improves the cache hit rate by up to 58x compared to a random caching strategy, while reducing web environment overhead by up to 3.2x, without degrading agentic system performance.

cs.AI↗

Kangaroo: A Private and Amortized Inference Framework over WAN for Large-Scale Decision Tree Evaluation

With the rapid adoption of Models-as-a-Service, concerns about data and model privacy have become increasingly critical. To solve these problems, various privacy-preserving inference schemes have been proposed. In particular, due to the efficiency and interpretability of decision trees, private decision tree evaluation (PDTE) has garnered significant attention. However, existing PDTE schemes suffer from significant limitations: their communication and computation costs scale with the number of trees, the number of nodes, or the tree depth, which makes them inefficient for large-scale models, especially over WAN networks. To address these issues, we propose Kangaroo, a private and amortized decision tree inference framework build upon packed homomorphic encryption. Specifically, we design a novel model hiding and encoding scheme, together with secure feature selection, oblivious comparison, and secure path evaluation protocols, enabling full amortization of the overhead as the number of nodes or trees scales. Furthermore, we enhance the performance and functionality of the framework through optimizations, including same-sharing-for-same-model, latency-aware, and adaptive encoding adjustment strategies. Kangaroo achieves a $14\times$ to $59\times$ performance improvement over state-of-the-art (SOTA) one-round interactive schemes in WAN environments. For large-scale decision tree inference tasks, it delivers a $3\times$ to $44\times$ speedup compared to existing schemes. Notably, Kangaroo enables the evaluation of a random forest with $969$ trees and $411825$ nodes in approximately $60$ ms per tree (amortized) under WAN environments.

cs.CR↗

Tesserae: Scalable Placement Policies for Deep Learning Workloads

Training deep learning (DL) models has become a dominant workload in data-centers and improving resource utilization is a key goal of DL cluster schedulers. In order to do this, schedulers typically incorporate placement policies that govern where jobs are placed on the cluster. Existing placement policies are either designed as ad-hoc heuristics or incorporated as constraints within a complex optimization problem and thus either suffer from suboptimal performance or poor scalability. Our key insight is that many placement constraints can be formulated as graph matching problems and based on that we design novel placement policies for minimizing job migration overheads and job packing. We integrate these policies into Tesserae and describe how our design leads to a scalable and effective GPU cluster scheduler. Our experimental results show that Tesserae improves average JCT by up to 1.62x and the Makespan by up to 1.15x compared with the existing schedulers.

cs.DC↗

Scaling Inference-Efficient Language Models

Scaling laws are powerful tools to predict the performance of large language models. However, current scaling laws fall short of accounting for inference costs. In this work, we first show that model architecture affects inference latency, where models of the same size can have up to 3.5x difference in latency. To tackle this challenge, we modify the Chinchilla scaling laws to co-optimize the model parameter count, the number of training tokens, and the model architecture. Due to the reason that models of similar training loss exhibit gaps in downstream evaluation, we also propose a novel method to train inference-efficient models based on the revised scaling laws. We perform extensive empirical studies to fit and evaluate our inference-aware scaling laws. We vary model parameters from 80M to 1B, training tokens from 1.6B to 30B, and model shapes, training 63 models. Guided by our inference-efficient scaling law and model selection method, we release the Morph-1B model, which improves inference latency by 1.8x while maintaining accuracy on downstream tasks compared to open-source models, pushing the Pareto frontier of accuracy-latency tradeoff. Notably, our experiments reveal that wider and shallower models can yield efficiency gains while preserving accuracy.

cs.LG↗

APACHE: A Processing-Near-Memory Architecture for Multi-Scheme Fully Homomorphic Encryption

Fully Homomorphic Encryption (FHE) is known to be extremely computationally-intensive, application-specific accelerators emerged as a powerful solution to narrow the performance gap. Nonetheless, due to the increasing complexities in FHE schemes per se and multi-scheme FHE algorithm designs in end-to-end privacy-preserving tasks, existing FHE accelerators often face the challenges of low hardware utilization rates and insufficient memory bandwidth. In this work, we present \NAME, a layered near-memory computing hierarchy tailored for multi-scheme FHE acceleration. By closely inspecting the data flow across different FHE schemes, we propose a layered near-memory computing architecture with fine-grained functional unit design to significantly enhance the utilization rates of computational resources and memory bandwidth. The experimental results illustrate that APACHE outperforms state-of-the-art ASIC FHE accelerators by 10.63x to 35.47x over a variety of application benchmarks, e.g., Lola MNIST, HELR, VSP, and HE$^{3}$DB.

cs.AR↗

Oblivious Monitoring for Discrete-Time STL via Fully Homomorphic Encryption

When monitoring a cyber-physical system (CPS) from a remote server, keeping the monitored data secret is crucial, particularly when they contain sensitive information, e.g., biological or location data. Recently, Banno et al. (CAV'22) proposed a protocol for online LTL monitoring that keeps data concealed from the server using Fully Homomorphic Encryption (FHE). We build on this protocol to allow arithmetic operations over encrypted values, e.g., to compute a safety measurement combining distance, velocity, and so forth. Overall, our protocol enables oblivious online monitoring of discrete-time real-valued signals against signal temporal logic (STL) formulas. Our protocol combines two FHE schemes, CKKS and TFHE, leveraging their respective strengths. We employ CKKS to evaluate arithmetic predicates in STL formulas while utilizing TFHE to process them using a DFA derived from the STL formula. We conducted case studies on monitoring blood glucose levels and vehicles' behavior against the Responsibility-Sensitive Safety (RSS) rules. Our results suggest the practical relevance of our protocol.

cs.CR↗