arXiv subjects

Sadiba Nusrat Nur

Publications and source records attributed to Sadiba Nusrat Nur.

3 recordsLinked to original sources

arXiv2026

Let Them Sleep: Recursive Self-Pruning Memory for Long-Horizon LLM Agents

Agents that run for a long time have a memory problem, and it is not the one most systems optimize for. Storing every trace and retrieving the nearest neighbors works early and quietly stops working later: the store fills up with failures the agent then repeats, and with facts that were true fifty turns ago. RSPM (Recursive Self-Pruning Memory) borrows the fix biology already found. At intervals the agent sleeps. It clusters its recent failures, distills each cluster into a short rule, and deletes the raw episodes it just learned from. Rules decay unless something reinforces them, and rules that recur across many contexts get promoted into general ones. We tested this on three settings with little in common: a difficulty-calibrated SWE-bench simulation, temporal consistency across 12 conversational benchmarks, and long-horizon QA on LoCoMo-10. All three improve, by 9.1, 11.9, and 7.6 points respectively. More interestingly, the two studies that instrument the store itself converge on nearly the same operating point without being tuned toward it: roughly 93% fewer retry loops and roughly 30% fewer stored entries. That hints that knowing what to forget is a general property of agent memory rather than a benchmark-specific trick.

cs.NE↗

arXiv2026

MIRAGE: A Polarity-Flipping Encoding Subspace in LLM Agents

When LLM agents are coerced into covertly encoding sensitive data (Base64, ROT13, acrostic, synonym chains, and beyond), the resulting outputs evade output-side detection but the underlying computation does not. Across nine encoding families and eight models from five architecture families, that computation is supported by a shared low-dimensional encoding subspace in the residual stream. A logistic-regression probe trained on eight encoding families recovers the held-out ninth at AUC 0.975-1.000, reading the computation rather than surface features. The same direction exhibits a second mechanistic signature at the planning token, flipping polarity to activate positively when the model will simulate the encoding inline and negatively when it will outsource it to a tool call, distinguishing two execution strategies before the encoded text exists. We build MIRAGE (Model-Internal Readout of Agentic Generation Exfiltration), a two-channel real-time monitor exploiting both signals. On 126 agentic exfiltration scenarios, it reaches AUC = 0.918, substantially outperforming output-only detection (AUC = 0.518). Monitor performance is fundamentally a property of the host model's geometry: benign-encoding false-positive rate ranges from 0% on Qwen-7B to 100% on Phi-3.5, revealing that the probe faithfully reads whether a model's geometry separates covert from overt encoding. Across all tested adversarial budgets, every attack suppressing the subspace also destroyed encoding fidelity, reported as an empirical regularity on the evaluated budgets, not a structural impossibility claim.

cs.CL↗

arXiv2026

UCSC-NLP at SemEval-2026 Task 13: Multi-View Generalization and Diagnostic Analysis of Machine-Generated Code Detection

With the rapid growth of large language models for code generation, distinguishing between human-written and AI-generated code has become increasingly critical for academic integrity, hiring evaluations, and software security. We present our system for SemEval-2026 Task 13: Multilingual Machine-Generated Code Detection, participating in Subtask A (binary detection) and Subtask B (multi-class attribution across 10 LLM families). For Subtask A, we fine-tune UniXcoder-base with a multi-view training framework that promotes generator-invariant representations. The framework combines domain-specific structural prefixes, delexicalization with symmetric KL consistency loss, token dropout, and mixed-content augmentation. Our system achieves 0.993 macro F1 on validation and 0.845 macro F1 on the test set, which spans unseen languages and domains. For Subtask B, we show that severe class imbalance (88.4% human code, 221:1 majority-to-minority ratio) causes catastrophic minority-class failure under standard fine-tuning, with macro F1 collapsing to 0.086 despite 88.4% accuracy. A class-weighted extension trained for 3 epochs recovers macro F1 to 0.345 (+301% relative), confirming that multi-class attribution requires imbalance-aware training strategies.

cs.SE↗