arXiv subjects

Jianfeng Xu

Publications and source records attributed to Jianfeng Xu.

At least 19 recordsLinked to original sources

arXiv2026

Value-Conditioned Proposal Acquisition for Test-Time Reasoning

Sampling and selection spend test-time compute inside a candidate pool, but they cannot recover a correct answer that never enters it. We introduce ValueFormation, a test-time policy that decides after each primary-model draw whether to stop, continue sampling, or acquire one proposal from a heterogeneous reasoning specialist. The three actions are compared on one cost-regularized value scale learned from vote-state features; the calibration is frozen across tasks and primary models, and acquired proposals are accepted only through a frozen posterior comparison. On held-out MATH-500, one Llama-fitted calibration improves over the strongest of eight identically evaluated baselines by more than fifteen points for each of three 8B primaries and transfers unchanged to Qwen3-14B. Removing acquisition lowers Llama MATH accuracy by 20.43 points, while same-trace and same-access replays attribute the gain to state-conditioned acquisition rather than specialist access alone. Where primary coverage is high, as on GSM8K, acquisition is rare and gains mainly come from cheaper allocation; on MBPP, a fixed code route attains the top accuracy at the lowest interaction cost. Varying the token price monotonically reduces calls without retraining.

cs.AI↗

arXiv2026

Partial Path Interception in Derivation DAGs: Structural Complexity, Parameterized Algorithms, and Approximation

Partial path interception (PPI) asks for a minimum-cost vertex set whose removal from a directed acyclic graph leaves at most a prescribed number of designated terminals able to reach a protected sink. We study how costs, depth, and fan-out affect its complexity. Unit-cost PPI is NP-complete at depth two with exactly two paths per terminal. Unique-path and individual-coverage instances admit polynomial-time algorithms, and bounded treewidth yields fixed-parameter tractability. Parameterized by the terminal threshold, weighted PPI is in XP and W-one-hard already at depth two, with no subexponential-time algorithm in that threshold under ETH. At depth two, undeletable terminals with bounded fan-out admit an FPT algorithm parameterized by the terminal threshold plus the fan-out bound. For unit costs, a dual-gammoid rank identity characterizes the optimum. Unit-cost depth-two instances with intermediate in-degree at most two are solvable in polynomial time for every terminal threshold, whereas depth-three instances are W-one-hard even with terminal out-degree one. On general directed acyclic graphs, minimum cuts and submodular optimization give a randomized approximation whose factor is the smaller of the coverage demand and roughly the square root of the number of initially exposed terminals, up to logarithmic factors. Reductions from smallest specified-edge subgraph, densest subgraph, and minimum union transfer conditional approximation lower bounds. PPI also gives an exact formulation of proof-valid caching under conjunctive derivation semantics: under independent premise erasures, the recovery probability equals one minus the erasure rate raised to the number of exposed premises. We extend this formulation to shared workloads, derive a sorting-based optimum under exact module routing, and give an AND-OR reliability envelope and a separate coded payload-recovery benchmark.

cs.IT↗

arXiv2026

Consistency Is the Attack Surface: Detecting Coordinated Retrieval Poisoning with Embedding Geometry

Consistency-based defenses for retrieval-augmented generation (RAG) rest on an assumption that a sparse adversary breaks: a poisoning campaign may modify a negligible fraction of the corpus, yet retrieval concentrates its passages into the reader's context. Across four Natural Questions retrieval seeds, at least one injection reaches the top-5 in 94.1% of targeted query--seed rows, and all three form a retrieved majority in 52.9%. Under our short-answer protocol, RobustRAG and ReliabilityRAG reach exposure-conditional attack success 0.512, against 0.576 for undefended RAG, while lowering clean accuracy from 0.328 to 0.216. Coordination, however, leaves a fingerprint: a dense, query-anchored component in the evidence embedding geometry. CoordGuard audits this structure from robustly normalized query-weighted affinities, without poison labels, correct answers, or reader gradients. Its always-filter mode suppresses the selected component whenever one forms; a benign-calibrated gate suppresses it only at unusually high scores and traces a robustness--utility frontier. Pooling four retrieval seeds, e5 always-filtering reduces $\mathrm{ASR}_{\mathrm{exp}}$ to 0.167 overall and 0.138 on parametrically unknown queries, at a 13.6-point clean cost, while a Contriever gate retains clean accuracy 0.315 at $\mathrm{ASR}_{\mathrm{exp}}=0.331$. The reduction transfers across two readers, three tasks, and multiple retriever--auditor pairs, and count-matched controls attribute it to component-targeted removal rather than deletion volume. These findings recast poison detection from judging truth to auditing the structure of retrieved evidence.

cs.CR↗

arXiv2026

Rethinking Calibration for High-Sparsity LLM Pruning: From Single Sources to Capability Coverage

Post-training LLM pruning relies on a small calibration set, usually chosen heuristically and judged by WikiText perplexity or by mean accuracy on general-domain commonsense benchmarks. This single-axis view hides capability-specific effects: at 60\% sparsity, strategies separated by $2.85$ points in averaged commonsense accuracy diverge by $51.9$ points in code retention, and across 15 sources no single one preserves all capabilities. We treat calibration as a capability-coverage problem. Balanced, a uniform four-way mixture under a fixed budget, tests the principle on real corpora: it outperforms every evaluated single source, and its margin over C4 grows from $+4.6$ points at 30\% sparsity to $+18.8$ at 60\%. The advantage holds across model scale and a second model family, and transfers to held-out benchmarks disjoint from calibration construction. When aligned corpora are unavailable, IGSP approximates that mixture from the base model and the evaluation taxonomy alone, outperforming prior self-generation methods by $7.23$--$12.79$ points on the LLaMA main suite. On Qwen2.5-7B, IGSP exceeds Balanced on both the main suite and held-out benchmarks. Capability coverage is a first-order design variable for high-sparsity pruning.

cs.LG↗

arXiv2026

CQDC-Warp: GPU-Accelerated Differentiable Quasi-Dynamic Simulation for Contact-Rich Manipulation

Efficient synthesis of contact-rich manipulation trajectories is essential for downstream robot learning. Gradient-based planning offers a promising approach, but jointly optimizing motion and contact decisions remains challenging. Meanwhile, many existing planners run on CPUs with limited throughput. We present CQDC-Warp, a GPU-accelerated differentiable quasi-dynamic simulator for batched model predictive control (MPC). We use it to study contact-guided trajectory optimization, in which intended robot--object contacts are specified before planning and encouraged through soft costs, while the simulated dynamics determine contact geometry and frictional response without a prescribed contact-mode schedule. Across $17{,}408$ trials on the $17$-object Allegro in-hand rotation benchmark, our framework achieves $79.8\%$ success, compared with $73.7\%$ for CFree and $65.2\%$ for IMPACT, and reduces mean trajectory-generation time by factors of $27.0$ and $50.2$, respectively. On a single NVIDIA RTX 4090 with $2{,}048$ parallel worlds, forward and backward throughput reach $2.06\times10^5$ and $2.25\times10^5$ world-steps per second, respectively. Further experiments cover manipulation with LEAP and Shadow hands, environmental contact exploitation, bimanual manipulation, whole-body control, and real-world trajectory replay on a Sharpa hand.

cs.RO↗

arXiv2026

SemRD-V2X: Closure-Guided Communication with Bounded Inference for Cooperative Perception

Vehicle-to-Everything (V2X) cooperative perception improves 3-D detection by sharing intermediate features, but dense remote features may repeat context that the ego agent can infer locally. Most communication-efficient designs optimize masks or codes empirically, leaving a more basic question open: which remote evidence is indispensable given the receiver's own observation? We introduce a closure-fidelity perspective on ego conditioned remote perception. Under a finite deductive abstraction and explicit conditions, its rate--distortion function decomposes over an irredundant core, and the exact zero-distortion rate becomes $P_A H(π_A)$. This analysis suggests a concrete design principle: transmit compact evidence and recover derivable context with bounded receiver-side inference. Guided by this principle, SemRD-V2X is an operational neural proxy that combines exact-budget BEV support selection, pointwise channel compression, and masked shared-weight reconstruction before standard fusion. Experiments on simulated V2XSet and real-world DAIR-V2X validate the resulting design. In a controlled five-run V2XSet comparison against a locally reproduced V2X-ViT-v1 baseline on one Tesla V100, SemRD-V2X reduces the analytical feature payload by $26.6\times$ while improving AP@0.5/AP@0.7 by 4.13/8.57 points, with 3.81\% additional mean compute latency. These results position closure fidelity as both an analytical lens and an actionable design principle for communication-efficient cooperative perception.

cs.AI↗

arXiv2026

CPUNeSy: Controlling Model Writes for Reliable Neuro-Symbolic Reasoning

LLMs excel at recalling statistical patterns but degrade sharply when answers must be derived, especially on multi-hop chains. Delegating derivation to deterministic symbolic executors shifts reliability to whether model-generated premises are source-supported. We introduce CPUNeSy, a serving architecture that controls model writes to symbolic state via a task-defined predicate interface and certificate gate, abstaining when grounding passes disagree. Component analysis isolates deterministic execution, restricted grounding, agreement, and source rechecking. Experiments show deterministic execution drives most accuracy recovery on derivation-heavy tasks; controlled writes mainly improve selective reliability by withholding unsupported or inconsistent answers, at a coverage cost. On multi-hop tests in law and formal math, deterministic execution recovers most of the gap over chain-of-thought and retrieval baselines, with full-pool gains up to 35.0 points. Certification is selective-serving control, not accuracy mechanism: with grounding traces fixed on ContractNLI, source rechecking removes a quarter of DeepSeek's wrong answers surviving two-vote agreement, at measurable coverage cost. When abstention is costly, routing withheld cases to an uncertified same-model fallback raises full-pool accuracy on MedCalc-Bench Verified by 13.9 and 4.9 points for Seed and DeepSeek; these gains are not from the certified channel. On LeanDojo Benchmark 4, kernel-restricted pools match BM25 recall@15 (89.3%). Gains depend on the grounder's error regime: bias-dominated grounders benefit less, consistent with our voting bound. Certificates guarantee derivational validity relative to admitted premises; semantic faithfulness to natural-language sources remains conditional on the source checker, and prospective validation is future work.

cs.SC↗

arXiv2026

Attention-Scoped Guidance: Training-Free Spatial Control for Image Editing

Instruction-guided image editing should change what the instruction names and leave the rest of the image untouched. In dual classifier-free guidance (CFG), an editor combines two directions at every denoising step, one that pushes toward the instructed edit and one that pulls back toward the source image, using global weights. We introduce Attention-Scoped Guidance (ASG), a sampler wrapper that makes these weights spatial. It reads a soft support map from the instruction attention that the editor already computes, then weakens text guidance where support is low and strengthens image anchoring where support is high. The wrapper requires no training, no external mask, and no additional network evaluation. On the full MagicBrush and PIE-Bench++ splits, ASG improves preservation-oriented metrics, leading three of four MagicBrush metrics and PIE-Bench++ background PSNR. A dose-matched control that removes the spatial placement loses up to 0.73 CLIP on PIE-Bench++, confirming that the spatial allocation itself carries the gain.

cs.CV↗

arXiv2026

Backbone-Adaptive Evidence Routing for Robust Pairwise LLM Judging

Pairwise language-model judges can gather evidence through direct comparison, reasoning, or reference-based verification, but no single protocol is best across benchmarks and judge backbones. We introduce Backbone-Adaptive Evidence Routing (BAER), which adapts the evidence mechanism while preserving candidate symmetry: swapping the two responses may reverse the preference but cannot change its strength. BAER separates each expert's signed preference from candidate-invariant reliability and builds three symmetric heads: evidence stacking, reliability-based expert routing, and candidate-blind reference verification. Development data select one head for each benchmark--backbone condition, and that choice is frozen before testing. Across four benchmarks and two 8B judge backbones, BAER achieves the highest test accuracy among the compared methods in all eight conditions, with full prediction coverage and gains of 0.87--7.32 points over the strongest external baseline. The results show that adapting how evidence is gathered is more reliable than fixing one judging protocol everywhere.

cs.AI↗

arXiv2026

New LoRA Skills Should Read but Never Write

Low-rank adapters (LoRA) make it cheap to fine-tune a large language model once per task, but combining several independently trained adapters into one model remains difficult: merging the updates in weight space causes interference, retraining on all task data is expensive, and routing between separate adapters gives up the goal of a single combined model. We trace the difficulty to two choices that every composition method makes implicitly. A LoRA update admits infinitely many equivalent factorizations; the choice among them is invisible while an adapter serves alone, but it determines what a learned interaction between adapters can see. A coupling between an old skill and a new one can likewise point in either direction, and the direction decides whether the old skills keep computing what they computed before. We introduce READ (Read-only Expansion of Adapter Deltas), which fixes both choices: each adapter is rewritten into a balanced canonical form that preserves its update exactly, and the coupling grows in one direction only, so a new skill can read the input subspaces of old skills but cannot write into their output subspaces. The only trainable object at each append is the new skill's row of the coupling matrix, and the composed update folds into the base weights with no inference cost, routing, or task-specific rules. We evaluate READ across four benchmark suites and two model families, adding skills one at a time. Across several families, READ improves every suite average over the strongest published baselines built from the same adapters---by more than twenty points on SuperGLUE and more than seven points on the domain suite---and nearly all complete addition sequences end above every direct baseline. Factor coordinates and coupling direction, which a lone adapter never exposes, are what decide whether composed skills survive.

cs.LG↗

arXiv2026

Regime Boundary Alignment for Evidence-Gated Question Answering

Retrieval-augmented language models are expected to answer from the retrieved evidence, but in practice they often keep answering when that evidence is missing. We trace this behavior to the training signal: answer-focused fine-tuning assigns no target to unsupported contexts, so it cannot distinguish a reader that abstains from one that guesses, and unsupported answering stays near 100% even as supported accuracy improves. We introduce Regime Boundary Alignment (RBA), which trains a single reader on matched variants of the same question and gold answer. The reader is trained to produce the gold answer when the context supports it, including when conflicting evidence is also present, and to abstain when the correct support is removed; inference is ordinary decoding, with no verifier, threshold, or regime label. On three multi-hop QA datasets across three seeds, RBA reduces the unsupported-answer rate by more than sixty percentage points relative to conflict-focused training while matching its supported accuracy. On a held-out TriviaQA retrieval-miss slice, the same reader reduces unsupported answering from 100% to below 1% while also improving supported accuracy. These results indicate that evidence-gated answering must be learned on both sides of the support boundary.

cs.CL↗

arXiv2026

BridgeMem: Causal Dyadic Transition Residuals for Temporal Knowledge Graph Forecasting

Temporal knowledge graph forecasting aims to infer future relational facts from the temporal structure of observed events. Existing forecasters mainly summarize history through entity states, relation states, paths, or exact recurrence. These views often miss pair-specific transition evidence, that is, the way prior relations between the query actor and a candidate change the odds of the target relation. We introduce BridgeMem, which estimates this quantity as a residual added to the log scores of a frozen full-vocabulary forecaster. For each candidate, BridgeMem retrieves the pair's events that strictly precede t, encodes their relations, directions, and lags, and converts them into a likelihood-ratio correction. A support-adaptive empirical-Bayes reader trusts exact transition counts where they are abundant and backs off to a learned attention estimator where they are sparse. The backbone's own uncertainty gates the correction, so confident queries and candidates without dyadic history are left unchanged. On five benchmarks, BridgeMem improves on the strongest of nine baselines from 2021--2026 in all 20 filtered MRR and Hits@{1,3,10} comparisons, with MRR gains of 0.0213, 0.0164, 0.0216, 0.0112, and 0.0028 over the best prior result. These results show the value of explicit dyadic transition modeling.

cs.LG↗

arXiv2026

ALOE: Semantically Addressed Low-Rank Operators for Knowledge Editing

Knowledge editing changes what a model knows by modifying parameters so that a requested fact updates while unrelated behavior is preserved. This is usually treated as a write problem, but editing also involves an address problem: deciding which hidden states should receive the new residual. An update that activates too narrowly memorizes one prompt, while one that activates too broadly disrupts neighboring knowledge. Parametric editors encode this scope implicitly, whereas memory-based editors make the selection explicit but keep it outside the edited model. We propose ALOE (Addressed Low-rank Operator for Editing), which learns semantic addresses from paraphrases and hard same-subject negatives, aligns them with autoregressive hidden states through rollout refinement and gate calibration, and embeds the resulting gated low-rank operator within one MLP layer, so that the deployed model runs in a single forward pass with no external retriever or auxiliary router. Evaluated on CounterFact, ZSRE, and KnowEdit across three 7--8B model families, ALOE achieves efficacy between 0.955 and 0.999 and locality between 0.981 and 1.000; mechanistic analyses confirm that the learned geometry separates competing edits and that calibration suppresses out-of-scope activation. The remaining errors concentrate in paraphrase coverage and write fitting.

cs.AI↗

arXiv2026

Reasoning Instructions Can Break Answer Decoding in Vision--Language Models

Chain-of-thought (CoT) instructions can distort multiple-choice VLM evaluation when a scorer appends a reasoning cue but reads answer-label logits before the model generates any rationale. We call this CoT-prefix scoring. On ScienceQA, Qwen2.5-VL-7B drops from 80.76% to 45.48%, and across five option-content permutations 93.54% of CoT-prefix predictions select the first slot. Condition-matched linear probes recover 78.94% from the same hidden states, while free generation restores 75.24%, showing that the answer often survives the prefix and the immediate readout fails. Vocabulary and layer diagnostics explain the mismatch: probability mass moves toward continuation tokens, while answer information remains linearly accessible in late layers. The effect recurs with varying severity across datasets and models, though not universally. These results show that CoT-prefix scoring can confound model knowledge with an evaluation-interface mismatch and should be avoided unless the requested and scored output events are aligned.

cs.CL↗

arXiv2026

Closure-Preserving Rate-Distortion for Reversible Logging

We study semantic compression of reversible-execution evidence for rollback reasoning. A run is a finite fact base; rollback semantics are modeled by a monotone closure operator induced by function-free Horn rules. A single edit replaces one fact by another; fidelity is the Jaccard discrepancy of the resulting closures, yielding a finite-alphabet distortion for rate-distortion analysis. A deterministic deletion scan decomposes the log into an irredundant core--preserving the closure--and a redundant remainder. Under admissible reconstructions (facts entailed by the original log), redundant facts are distortion-invisible, reducing the semantic rate-distortion function to a core-only optimization scaled by the core probability mass. At zero distortion, the optimal rate is a hypergraph entropy induced by overlaps of zero-distortion reconstruction sets on the core. We introduce a rollback-task loss based on a rollback observable, deriving parallel endpoint and factorization laws. The framework is instantiated on reversible causal nets and discussed in the event-structure view, showing how reversing disciplines yield different cores and compression frontiers. Numerical evaluation uses Blahut-Arimoto to design single-letter test channels and Monte Carlo reconstruction to assess end-to-end degradation at the log level.

cs.IT↗

arXiv2026

Rate-Distortion Theory for Deductive Sources under Closure Fidelity

We study lossy compression of a finite statement source generated in a fixed deductive environment. The source symbols are statements in a knowledge base endowed with a shared proof system, and reconstruction fidelity is measured by preservation of deductive closure rather than by symbolwise equality. Fixing the proof system and a canonical scan order yields a decomposition of the source alphabet into an irredundant core and redundant stored consequences. At zero distortion, each core symbol induces a set of distortion-free reconstructions. In the nonconfusable (disjoint-core) regime, we show that the minimum zero-distortion rate equals the source mass of the core times the entropy of the source conditioned on that core. In the general confusable-core regime, we characterise the exact zero-distortion rate via a hypergraph-entropy quantity induced by jointly realisable core subsets, with a reduction to Korner-style graph entropy under a natural pairwise realisability condition. For reconstruction alphabets contained in the deductive closure of the source knowledge base, we further prove that the full rate-distortion function depends only on the core, so redundant states are invisible to both rate and distortion. Finally, when the decoder is limited to a bounded inference-depth budget (a bounded number of iterations of the immediate-consequence operator), we obtain an exact rate-depth-distortion characterisation. Under an additional order-robustness assumption identifying the chosen core with the order-free essential set, this characterisation interpolates between classical symbolwise compression and unconstrained deductive compression.

cs.IT↗

arXiv2026

Semantic Rate-Distortion Theory: Deductive Compression and Closure Fidelity

Shannon's rate-distortion theory treats source symbols as unstructured labels. When the source is a knowledge base equipped with a logical proof system, a natural fidelity criterion is closure fidelity: a reconstruction is acceptable if it preserves the deductive closure of the original. This paper develops a rate-distortion theory under this criterion. Central to the theory is the irredundant core-a canonical generating set extracted by a fixed-order deletion procedure, from which the full deductive closure can be rederived. We prove that the zero-distortion semantic rate equals a quantity that is strictly below the classical entropy rate whenever the knowledge base contains redundant states. More generally, the full semantic rate-distortion function depends only on the core; redundant states are invisible to both rate and distortion. We derive a semantic source-channel separation theorem showing a semantic leverage phenomenon: under closure fidelity, the required source rate is reduced by an asymptotic leverage factor greater than one, allowing the same knowledge base to be communicated with proportionally fewer channel uses-not by violating Shannon capacity, but because redundant states become free. We also prove a strengthened Fano inequality that exploits core structure. For heterogeneous multi-agent communication, an overlap decomposition gives necessary and sufficient conditions for closure-reliable transmission and identifies a semantic bottleneck in broadcast settings that persists even over noiseless channels. All results are verified on Datalog instances with up to 24,000 base facts.

cs.IT↗

arXiv2026

Semantic Channel Theory: Deductive Compression and Structural Fidelity for Multi-Agent Communication

Shannon's information theory deliberately excludes message semantics. This paper develops a rigorous framework for semantic communication that integrates formal proof systems with Shannon-theoretic tools. We introduce an axiomatic information model comprising Lsem-definable state sets linked by computable enabling maps, and define the semantic channel as a composition of Markov kernels whose supports respect the enabling structure. A fixed proof system induces an irredundant semantic core and a derivation-depth stratification, enabling four distortion measures of increasing semantic depth: Hamming, closure, depth, and a parameterized composite. Six families of computable semantic channel invariants are defined and their inter-relationships established, including a data processing bound, a semantic Fano bound, and an ideal-channel collapse theorem. The central quantitative result is a deductive compression gain: under closure-based fidelity, the minimum block length is determined by the irredundant core size rather than the full knowledge-base size. We instantiate the framework for heterogeneous multi-agent communication, introducing an overlap decomposition that yields necessary and sufficient conditions for closure-reliable communication. A semantic bottleneck phenomenon is identified in broadcast settings: vocabulary mismatch imposes irreducible fidelity limitations even over noiseless carriers. All results are verified on an explicit Datalog instance.

cs.LO↗