arXiv subjects

Xiangchen Guan

Publications and source records attributed to Xiangchen Guan.

10 recordsLinked to original sources

arXiv2026

MemPrism: Task-Conditioned Relational Memory Views for Long-Horizon Agents

Long-horizon agents performing multi-step Web tasks must reuse earlier page observations and interactions to support subsequent actions. However, preserving relevant history does not ensure that it exposes the relations needed for the current decision. We call this gap representation mismatch. To this end, we propose MemPrism, a task-conditioned relational memory framework that separates persistent experience storage from decision-time working memory. MemPrism records interactions as the event stream and dynamically constructs relational views according to the current task context. A lightweight view policy selects the relation structure, evidence range, outcome condition, and granularity, while a deterministic composer and renderer transform historical facts into a temporary optical working-memory view for a frozen task policy. Experiments on long-horizon embodied and web-agent benchmarks show that MemPrism consistently improves the task performance, especially as trajectories become longer, while reducing memory token consumption. Furthermore, the learned view policy transfers across different VLMs without additional adaptation, demonstrating the effectiveness of task-conditioned relational views as a general memory interface for agents.

cs.AI↗

arXiv2026

DuplexLedger: Benchmarking Transactional Memory in Full-Duplex Speech Agents

Full-duplex speech agents must update persistent memory while listening and speaking, yet it remains unclear which events should be stored, revised, or excluded. Existing benchmarks isolate interaction timing from memory outcomes, obscuring the state transitions behind a correct answer. We introduce DuplexLedger, a bilingual benchmark of 1,094 persistent sessions aligning four full-duplex scenarios with event-level Write/Update/Ignore/Retain supervision, explicit pre/post memory states, and 6,564 same-session questions and checks. Evaluating five full-duplex models reveals an interaction-memory gap: no system combines strong boundary control with coherent final-state memory, and the best final-state fact coverage is 2.01%. We propose ACT-Mem, an 8.92M-parameter sidecar comprising an event controller, a versioned transactional ledger, and a typed retriever coupled with a frozen MiniCPM-o 4.5 backbone. On 219 held-out sessions, ACT-Mem increases latest-value accuracy from 4.11% to 8.68% while reducing stale assertions and distractor adoption. It improves interaction boundaries but slows interruption stopping. These results motivate selective, auditable full-duplex memory rather than indiscriminate transcript retention. Code is available at https://github.com/lyx-researcher/DuplexLedger.

cs.AI↗

arXiv2026

Depth, Not Breadth: Best-of-N Jailbreaking Beyond Surface Noise

Best-of-N jailbreaking spends a query budget on surface variation, scrambling and recasing a request until one draw lands. We ask what the same search buys when it wraps a structural encoding instead. Against SAGE, a published self-check defense, best-of-N over a code-completion encoding reaches 67, 22 and 15% of behaviors on three open-weight targets, where that encoding fired once reaches at most 4.7% and the published character search at full budget at most 3.0%: 9 to 68 times the sum of the parts, with bootstrap intervals clearing both ingredients on every target. We report the operative figure beside the headline rather than the headline alone: at the actionable severity threshold those cells read 24, 8 and 1 behaviors (95% interval [13, 28], [3, 13], [0, 3]). A 2x2 on Llama-3.1-8B holding encoding and variation apart shows that SAGE and a classifier gate fail to different factors: SAGE is broken by the depth of the encoding (7 -> 67 behaviors at fixed variation) and the gate by the breadth of the variation (13 -> 57 at fixed encoding). Per-draw reporting also overstates robustness, because an attacker who may try N times experiences the maximum over draws while safety results are reported as means: on one target SAGE stops 99.8% of individual draws yet loses at least as many behaviors to a repeat attacker (12) as a classifier gate that stops 95.6% (10). Removing the target's sampling shrinks SAGE's coverage, relative to its own level, by 59, 76, 82 and 29 percentage points more than an undefended control's, against 25, -5, 8 and 2 for a defense whose verdict comes from a fixed shadow model. The design that loses is the one fusing screening and answering into a single generation, so every attacker draw redraws the safety decision as well. The prescription is architectural, not free: do not fuse screening with generation.

cs.CR↗

arXiv2026

MemoWM: How World Models Change What Agents Need to Remember

Long-term agents face growing storage demands as they accumulate experience. World models capture reusable regularities that can reduce the information stored for each experience. We formulate the problem of memory allocation conditioned on a world model and introduce MemoWM, a framework that uses shared predictions to compress retained information and reconstruct omitted content. Its task-aware allocation rule balances the expected impact of reconstruction errors against storage cost, retaining information with downstream value beyond the predictive prior. Across five long-term agent-memory benchmarks, MemoWM achieves 42.42\% average answer accuracy, exceeding the strongest baseline by 2.62 percentage points, while reducing average experience-specific storage by 53.9\% relative to MIRIX, the most storage-efficient baseline. Further analysis shows that stronger world models reduce per-experience storage at comparable task quality. Accounting for model parameters reveals a trade-off between shared model capacity and recurring storage costs, with the capacity that minimizes total storage increasing as more interactions are retained. Our code is available at https://github.com/Feld-maxiu/MemoWM.

cs.LG↗

arXiv2026

Blind, Not Weak: A Best-of-Suite Safety-Utility Frontier for Recover-and-Reguard Defenses Against Encoded VLM Jailbreaks

Safety classifiers ("guards") are the dominant black-box defense for vision-language models, yet a guard judges an input's surface form, not its meaning: a harmful request re-encoded as set theory, formal logic, a classical language, code, or text rendered inside an image slips past a guard that would block it in plain language - the decode gap. The standard fix is a preprocessor that recovers image content and decodes the encoding before the guard. We build one and evaluate it against an ensemble of eleven encoding attacks - six published implementations, one standard encoding baseline, one adapted and three author-constructed renders - counting a behavior as broken if any attack succeeds. Restoring a view the guard never had is what buys coverage - block rates on image renders go from exactly zero to 67-90% - and what it costs in benign traffic is set by the guard, not by the mechanism: one guard pays 9 benign blocking points for the same 70-point gain another pays 69 for. It still does not make the system safer: against an attacker free to choose among eleven encodings, closing one channel relocates the success rather than removing it, and no ensemble contrast for that step survives multiple-comparison correction. What does lower ensemble attack success is a reguard step that re-screens the recovered pre-decode surface, and it is the one every guard pays for: it raises benign over-refusal on all ten guard-target pairs, where restoring a single channel raises it on some and not others. Across the full guard x target x condition factorial, no configuration reaches an ensemble attack-success rate at or below 40% while holding benign over-refusal under 70%. That empty region is a property of the configurations we sample, not a bound on what recovery-based defenses can reach, and we breach its safety half ourselves.

cs.CR↗

arXiv2026

The Uncontrolled Variable: Vision-Language Refusal Is Conditioned on the Image-Attachment Interface, and Not Robust to Irrelevant Image Properties

We show that aligned vision-language models also condition refusal on a property of a request's form: whether an image is attached, holding everything the request asks fixed. Attaching a blank canvas, an image that cannot be read, cannot relate to the request, and is byte-identical across every prompt in its condition, shifts benign refusal by tens of points. The shift is not blanket caution but a threshold shift: genuinely neutral instructions are almost unaffected (<=2 percentage points on three of four hosted models) while borderline-benign prompts move +23 to +51 points, so the cost falls on sensitivity-adjacent traffic, meaning benign questions about privacy, self-harm, violence and illegal activity. There is a benign reading of such a threshold, namely that attachment correlates with risk in real traffic, and we take it seriously; a black-box study cannot measure that correlation and we do not claim to. What it can test is whether the response to attachment is robust to variation carrying no information about the request, and on four independent measurements it is not. It varies with canvas colour and pixel count. Its sign inverts across checkpoints. It survives an explicit instruction to disregard the image. And on one model it fires on a bare assertion that an attachment exists, with nothing attached and the modality word contributing none of it. Finally we price it. On a matched harmful set the same canvas does lower attack success, so the cue buys something. But the charge is decoupled from the purchase: the checkpoint with the least harmful headroom we measure, completing only 2% of plain harmful requests, still pays the benign cost in full, and across our models the harmful-side denominator falls as alignment improves while the benign cost does not track it down. Image presence is not a conservative default that a deployer chose and priced. It is an uncontrolled variable.

cs.CR↗

arXiv2026

0%, 45%, or 99%: A Guardrail's Own Share of the Refusals It Is Credited With

A defended pipeline's refusals have two producers: the guardrail bolted in front of the model, and the model's own alignment. Recovering the split costs nothing, because a guard block replaces the model's response and the two counts are therefore disjoint. Holding the defense, the targets, the corpus and the judge fixed, the guardrail's own share of the refusals credited to it is 0%, 41-45%, or 99% across three settings that a results table would describe identically. Two choices move it, and neither belongs to the deployer who bought the guardrail. The attacker drives the share to zero by choosing which channel carries the payload: a plainly written request rendered as pixels, with nothing obfuscated, leaves a text guard's read covering none of it. The evaluator drives the share to 99% by choosing what text fills a defense's internal slots: fill them with the unencoded request behind an encoded attack, a read no deployed defender possesses, and the same guard blocks almost everything. The two consequences differ, and only the attacker's can happen to a running system. The evaluator's choice is an artifact carried by the literature, and its size is set by where the granted text lands: substantial at a guard gate, smaller in a caption-mediated re-check, absent in a majority-vote smoother, an ordering reproduced in an independent replicate. Isolating the grant inside the caption-mediated defense refutes the prediction we registered, since the harm-verdict stage contributes nothing while the stage that regenerates the answer carries the whole effect. The reference implementation builds every stage from a single prompt field that cannot represent the difference between what the attacker sent and what the benchmark records, and an audit of four further released harnesses and of the benchmark itself finds the same structural gap, so faithful porting supplies the grant silently.

cs.CR↗

arXiv2026

Refusing Everything Looks Safe: Restoring the Benign Arm to Encoded-Prompt Evaluation

Encoded-prompt attacks are evaluated almost entirely on their harmful arm: a benchmark sends obfuscated harmful requests and reports how often the model complied. A high refusal rate there is reported as safety, and it is equally consistent with a model that has stopped telling the request apart from anything else in the same format. We run the benign arm through the same transformation, and the two cases are far apart. Across four 7-8B models spanning three base families and four post-training recipes, refusal of harmful homoglyph-encoded prompts spans 0.08 while the same four span 0.57 on the identical requests in plaintext. What the encoding destroys is not refusal but the harm gap: on one model the gap between harmful and benign refusal falls from +0.82 in plaintext to exactly 0.00 under the encoding, and a benchmark reading only the harmful arm scores that model and one retaining a +0.61 gap identically. Running the cell such benchmarks leave out (plaintext content wearing the attack template, with nothing obfuscated) shows that on two of the four models the loss is caused by the protocol rather than by the character transformation, and on a third by the characters. Across a full SFT -> DPO -> RLVR pipeline the harm gap rises by +0.26 with a paired interval excluding zero while the standard harmful-arm metric registers no resolved change at all. We report twelve instrument defects, each with the control that caught it, including a binary jailbreak judge that fires on 0.61-0.70 of responses to plaintext benign prompts; six of the twelve inflate apparent safety, which is the direction a broken safety evaluation fails in by default.

cs.CR↗

arXiv2026

Trust Is Not Enough: Influence Calibration for On-Policy Self-Distillation in Agentic RL

On-policy self-distillation (OPSD) gives language agents dense token-level supervision from a privileged self-teacher on the policy's own trajectories. Existing methods allocate this supervision mainly by teacher trust, but trust does not reveal whether emphasizing a token supports the current policy objective. We call this the trust-utility mismatch and introduce Influence Calibration for Self-Distillation (ICSD). For each supervised token, ICSD measures the first-order response of its importance-weighted RL surrogate contribution to a teacher-directed output perturbation. Batch-adaptive calibration converts this non-stationary signal into a bounded allocation weight while preserving the original auxiliary-loss mass within each action turn. These detached weights affect only the distillation loss and require no additional model pass. Across ALFWorld, WebShop, and Search-QA, ICSD improves all matched aggregate metrics over trust-only allocation under Group Relative Policy Optimization (GRPO) and Group-in-Group Policy Optimization (GiGPO), across two model families spanning 1.5B to 7B. At 7B, it reaches 96.1% ALFWorld success and a WebShop score of 93.1. Frozen-batch analyses show that ICSD reduces teacher-supported mass assigned to objective-opposed tokens from 60.1% to 37.8% and raises cosine compatibility with the RL gradient by 0.192. A companion repository is avail- able at https://github.com/lanqz7766/Influence-Calibration-for-On-Policy-Self-Distillation-in-Agentic-RL.

cs.AI↗

arXiv2026

Decoy Images Amplify Caption-Mediated Defenses Against Encoded Jailbreaks

We report a counter-intuitive interaction between image inputs and existing black-box defenses on Vision--Language Models (VLMs): pairing an encoded jailbreak prompt with an unrelated decoy image can sharply lower attack success rate (ASR). The operative change is in the defense pipeline, not in the image. Across five frontier VLMs, two encoded-attack families, and three black-box defenses, a caption-mediated defense (ECSO) that leaves ASR essentially unchanged on text-only encoded input drops it by up to $73$pp once a content-free decoy is attached; every non-saturated contrast is significant under exact McNemar tests. We advance two hypotheses for this pattern, supported by indirect evidence rather than pipeline introspection, since a black-box threat model precludes inspecting vendor internals: caption-mediated defenses branch on image presence, and intrinsic image-side safety engages on image-resident content. Three controls constrain the explanation. Blank-canvas and natural-photograph decoys reproduce the effect on every model, implicating image presence rather than content; the effect replicates on three open-weight VLMs served with no moderation layer, so it is not a vendor-filtering artifact; and a non-symbolic, meaning-based encoder reproduces it, so it is not specific to symbolic obfuscation. Attaching a decoy unconditionally is not deployable --- it raises benign refusal to $20$--$79\%$, an inflation of $+10$ to $+67$pp --- but gating attachment on a lightweight encoded-input detector returns benign refusal to the text baseline while preserving the safety gain wherever the detector fires, making detector recall the binding constraint. Under adaptive attacks that target the caption-mediated re-check, the effect degrades but holds. We frame this as an observation about pipeline interaction, not as a robust defense.

cs.CR↗