Skip to main content
Search

Pennsylvania State University

15 related articles

Subscribe to updates
  1. Coding agents stop rewriting history: ObsDrop drops stale KV instead, reaching 5.37× throughput at 16 concurrent agents

    Coding agents mostly append to their history, so prefix-cache reuse reaches 98–99%, yet text-level context compaction rewrites earlier tokens and forces the whole suffix to be recomputed. ObsDrop moves compaction from the prompt into the KV cache: it keeps the logical token prefix unchanged while releasing the KV of stale tool observations and excluding those positions from attention, implemented in SGLang and Sparse-vLLM. At 16 concurrent agents it reaches 2.45–2.85× throughput with FlashInfer and a 5.37× peak with Triton relative to no compaction, with task quality close to no compaction.

  2. arXiv

    Software agents no longer get only pass or fail: the policy itself ranks debugging traces within the same outcome

    Reinforcement learning rewards for software engineering agents no longer stop at pass or fail: the current policy compares debugging traces from the same task and verifier outcome, turning those preferences into bounded auxiliary rewards and lifting Qwen3-14B average accuracy on SWE-bench Verified from 32.1% to 34.7%.

  3. arXiv

    One character changed, the decision changed: 30 text models and 17 vision models take the same attack candidates

    DecisionRobust evaluates JEV-style decision models under text and image perturbations and prompt injection, and finds that 30 text models and 17 vision models take the same candidate set, with held-out models excluded from candidate selection also incurring errors while label-consistency checks sharply reduce apparent attack success.

  4. arXiv

    A 27B model learns to propose language-model upgrades: 200 real training runs beat GPT-6-sol on held-out tasks

    Deciding which change is worth trying when a language model improves a language model has depended on costly training experiments. LM2LM uses a frozen LLM as a research world model that predicts a proposal's gain from real experimental records in context, turning those predictions into reinforcement-learning rewards, and periodically executes the current policy's top-rated proposals to write measured outcomes back into the records. Across eight autoregressive and diffusion language-model development environments, Qwen3.8-27B trained this way outperforms GPT-6-sol on held-out tasks, even when GPT-6-sol is given 300 real experimental records in context, at about 33% lower data-generation cost.

  5. arXiv

    Routing stops ranking answers by one score: security queries get their own quality, closing nearly 33 points of the security gap at the same models and budgets

    SCRouter splits routing quality into utility for ordinary queries and security for security-sensitive ones, weights the two per query, and recovers 60.9% of the gap between random selection and the per-query oracle on its own benchmark, 12.9 points above the previous best router R2-Router.

  6. arXiv

    Geo-distributed LLM serving stops reading only grid carbon: counting batching and queueing cuts emissions up to 60% at about 3% SLO violation

    When geo-distributed LLM serving schedulers account for an engine's batching and queueing state, they cut carbon emissions by up to 60% relative to round-robin under MLPerf per-token latency targets, with an average SLO violation of about 3%.

  7. arXiv

    Tool descriptions become security contracts: 253 bugs in 273 MCP servers, most without a crash

    When a large language model calls an external tool, the orchestration layer checks that the call is well formed, not that its arguments are safe. ToFu turns the argument semantics already documented in tool metadata into executable security checks, then generates fuzzing inputs and oracles from them, confirming 253 bugs across 343 tools in 118 of 273 runnable MCP servers, most of which returned normally.

  8. arXiv

    PAPER2LLM++ lets models continually self-evolve from a stream of papers, absorbing new findings on sequential failures while retaining earlier gains

    The work introduces PAPER2LLM++, a framework for continual self-evolution of LLMs from research papers: it treats the growing literature as a stream of evidence and supervision, extracts evidence-grounded findings from each incoming paper, tests whether the reported limitation persists in the current model, converts findings into candidate learning signals when needed, and uses a try-evaluate-commit procedure to integrate an update only when it improves the targeted behavior without substantially forgetting prior improvements or degrading general capabilities; across a sequential stream of research-discovered LLM failures, models progressively incorporate new findings while retaining earlier gains.

  9. arXiv

    LinkedIn and university collaborators find that 7B-to-72B students trained on rejects from smaller frozen models do better on code and math while using less compute

    Holding prompts, chosen responses, SeqKD initialization, DPO objective, and training budget fixed while varying only the reject source, the study finds that across Qwen2.5 students from 7B to 72B, rejects generated by smaller frozen Base models with less inference compute train stronger students than the students' own rejects on code generation and mathematical reasoning, and it derives a finite-horizon utility bound in a linearized feature model that motivates three interventions: source mixing, prompt reassignment with lexical permutation, and reference-likelihood reselection.

  10. arXiv

    ATLAS preserves relational geometry while calibrating the latent distribution in LeWM, improving goal-reaching success on PushT, TwoRoom, and OGBench-Cube, with the largest gain on higher-novelty TwoRoom episodes

    The authors introduce ATLAS, a training objective that transfers normalized pairwise structure from an encoder's mean-pooled patch features to the planning latent and uses Wasserstein embedding matching (WEMReg) based on one-dimensional Wasserstein-2 transport to calibrate its marginal; instantiated in LeWM, ATLAS improves mean goal-reaching success across PushT, TwoRoom, and OGBench-Cube on both lower- and higher-novelty evaluation subsets, with the largest gain on higher-novelty TwoRoom episodes, while diagnostics show stronger novelty-related structure in the planning latent, improved marginal calibration, and lower multi-step prediction error.