Skip to main content
Search

Stony Brook University

7 related articles

Subscribe to updates
  1. arXiv

    Agent memory stops only looking back: Pro-Memory splits future intentions from the moment to act

    Most memory systems for large language model agents only store and retrieve past information, while proactive agents also need to hold future intentions and realize them when the right moment arrives. Pro-Memory separates keeping a future intention from deciding when and how to realize it: pending objectives stay linked to their supporting evidence, new observations are checked against the original goal and activation conditions, and contributions are coordinated with the agent's ongoing actions. On two benchmarks covering delayed-intention tasks and open-ended proactive assistance, proactive task performance improves consistently over representative memory approaches across multiple LLM backbones while unnecessary interventions are limited.

  2. arXiv

    MASA recasts sparse-attention selection as matrix approximation, lifting long-context accuracy across three sparse frameworks at no extra budget

    The authors argue that existing sparse-attention selectors keep large or high-mass entries of the attention matrix even though Transformers use that matrix through multiplication with value vectors, and propose MASA, which replaces attention-mass ranking with a closed-form score measuring each sparse unit's reduction in matrix-product approximation error; plugged into MInference, SeerAttention, and FlexPrefill without changing their sparse kernels, candidate units, or budgets, it yields consistent accuracy gains on RULER, InfiniteBench, and LongBench with negligible latency overhead.

  3. arXiv

    Pre-verdict residual stream activations predict LLM judge position flips with AUROC .621–.850 and transfer to .685–.853

    Using nested grouped cross-validation on 534 JudgeBench pairs and 1,802 MT-Bench comparisons, this work trains L2-regularized linear probes on residual stream activations recorded immediately before the verdict to predict position flips, where swapping candidate order changes the judge's selected response identity; AUROCs reach .621–.850 across four judges, exceeding a combined baseline of verbalized confidence, verdict logits, length differences, and the initial choice by .062–.113, and frozen probes transfer to MT-Bench at .685–.853.

  4. arXiv

    A safety-aligned but backdoored teacher passes its backdoor to a clean student through on-policy distillation, reaching up to about 70% trigger attack success at a 3% poisoning rate

    This work studies whether on-policy distillation (OPD) for safety can propagate backdoors: the authors build safety-aligned but backdoored teachers with bi-GRPO and distill initially clean students through OPD, finding that a 3% poisoning rate yields up to about 70% trigger attack success, that more training epochs (67% after 16 epochs with only 10 poisoned samples) and top-K KL both accelerate backdoor transfer, and that a KL-reward-clipping method, Lazy Defense, delays this transfer.

  5. arXiv

    HyperGuide guides LLM reasoning along a single trajectory via virtual tokens in hyperbolic space, matching or exceeding search and verifier baselines on competition math and code generation with far fewer tokens

    HyperGuide treats reasoning as movement through a tree of states embedded in hyperbolic space (the Poincaré ball), first training a state encoder and then a guidance head that predicts the direction of a minimum-cost transition from the current state, and inserts that direction into the decoder as a virtual token after each reasoning step so inference follows a single autoregressive trajectory without expanding or reranking candidates; on competition mathematics and code generation experiments it matches or exceeds the accuracy of search- and verifier-based baselines while generating substantially fewer tokens than those baselines and about as many as few-shot prompting.

  6. arXiv

    Kuaishou team's DARA weights rewards by active-group density, cutting tool-calling format steps by up to 26% and math length-compliance steps by up to 65%

    The work characterizes each reward's contribution to policy updates in multi-reward RL through "advantage energy," shows that under idealized GDPO normalization a reward's advantage energy is proportional to its active-group density (the fraction of rollout groups in which it yields nonzero relative advantages), and proposes DARA, an inverse-square-root density weighting that reaches target behaviors faster than GDPO: up to 26% fewer training steps for format compliance in tool calling and up to 65% fewer steps for near-saturated length compliance in mathematical reasoning, while remaining competitive in final performance.

  7. arXiv

    Dynamic Adaptation of the LLM Context for Generating Routines with Coupled Semantics

    The work proposes dynamic context adaptation: a validation-generation loop in which a validation agent extracts structured diagnostic feedback from execution traces, a generation agent proposes multiple candidates per iteration, a knowledge graph supplies semantic constraints, and simulated annealing performs non-greedy selection, targeting the LLM code-generation limitation the authors call static binding; across eight problems the method outperforms zero-shot, Reflexion, and OpenEvolve on seven of eight problems at both 300 and 600 evaluations (p < 0.01), achieves the best score at 1000 evaluations on the primary motivating problem of cross-coupled optimization (0.694 vs. 0.681, p = 0.019, d = 0.52), and ablations identify structured execution feedback as the primary driver.

You’re all caught up