Searcharxiv⌕ Search

arXiv subjects

Chuanjie Li

Publications and source records attributed to Chuanjie Li.

3 recordsLinked to original sources

Staged Semantic Distillation: Preserving Free-Form Semantics in a Low-Compute Non-LLM Surrogate of a Staged LLM Agent

Staged LLM agents make market and social simulators realistic, but one 102-day run of a 200-agent, ten-index market simulator issues 203,100 language-model calls and returns one realization. Conventional low-compute distillation drops their free-form intermediate state. We present staged semantic distillation, a low-compute method that preserves non-structured semantics in a non-LLM student trained from one recorded trace. A frozen encoder maps each teacher output to a 1024-d vector; the student sees typed stage-tagged vectors with stage-specific visibility, is jointly supervised, and trains with scheduled self-history. One TwinMarket trace yields a 13.2M-parameter deterministic student that trains in 1,658 seconds on one 8-GB GPU and runs with zero teacher or decoder calls and no text generation (0.14 ms per decision). Its analysis stays close to the teacher (cosine 0.943), and swapping semantic inputs between agents moves 64.7% of decisions against 12.8% for a matched donor, so the channel is preserved and live. We use the surrogate to study what the transfer carries. It reaches 0.8347 balanced accuracy on the teacher-forced replay and 0.7783 [0.766,0.791] on the deployment decoder; the five-seed gain over a no-semantic retrain is +2.2 [1.21,3.14]. Against the teacher's own re-sampling level, round-one R is 97.1% [95.8,98.4] and 92.3% [89.0,95.4]; R is a diagnostic, not an accuracy. A five-seed content-by-schedule study isolates the gain: scheduled self-history carries it (+2.2 [1.21,3.14]); shuffled and constant content controls reproduce it, while content matters teacher-forced. The design rule: invest in the deployment schedule, let content earn its place at the teacher-forced operating point. A 20-seed, 200,000-decision sweep takes 12 minutes with zero LLM calls, versus three hours for the teacher. The paper is a method-and-case-study contribution on one simulator and trace.

cs.AI↗

Macro Control and Capability Evaluation of Black-Box World Models: A Two-Regime Framework for LLM-Agent Markets

Large language model (LLM) agents can be organized into multi-agent markets that act as world models, mapping macro information to emergent macro dynamics while the evaluator observes only macro inputs and outputs. This paper separates a model's capability into two stages that a single aggregate score cannot separate: information simulation, from macro information to micro behavior, and macro emergence, from micro behavior to the macro outcome. We introduce a soft macro-control method that announces a public reference path day by day without touching individual decisions, orders, or matching, and a two-regime evaluation framework that measures both stages from macro observations alone. In the reference-only regime the symmetric true and flipped references share a common level R, which measures realization under the announced reference, while their paired gap B measures structural bias; in the information-injection regime external information dilutes the reference, so only the true-flipped gap Delta remains interpretable. On five single-index LLM-agent markets, announcing a target path shifts the realized path toward the announced target on both announced references under both information conditions. Relative to the no-reference path, the tracking score is higher by 0.306 and 0.415 without news and by 0.183 and 0.329 with news; within the stock/cash scan structural bias is lowest at 70/30 (B=0.042); realization peaks at a 3:7 technical share (R=0.894 versus 0.358 for technical-only); and Delta is 0.0341 with fundamental data and 0.0487 with news. Both stages use the same two observable scores, each read in the regime where it is interpretable.

cs.AI↗

TTRSD: Test-Time Reinforcement Learning with Self-Distillation for Vision-Language Models

Test-time reinforcement learning enables vision-language models (VLMs) to adapt using unlabeled inputs. However, repeated sampling under fixed visual conditions can reinforce shared perceptual errors, while sequence-level rewards fail to isolate visual perception the foundational bottleneck that anchors multimodal reasoning risking the degradation of pre-trained reasoning capabilities. We propose TTRSD, a test-time reinforcement learning framework combining multi-view answer-level self-distillation with visual contrastive token selection. A shared policy aggregates teacher predictions across original, cropped, and downsampled views into an answer distribution. Student trajectories generated from the original image receive rewards based on the support for their final answers in this distribution. To allocate this feedback precisely toward perceptual bottlenecks, we compare the log-probabilities of the same sampled tokens under original and visually ablated inputs while holding their textual prefixes fixed, selecting visually sensitive positions for policy-gradient updates. TTRSD separates update direction, determined by group-relative advantages, from update position, determined by visual sensitivity, without requiring ground-truth labels, external verifiers, or a separate teacher. With only 20 unlabeled adaptation samples, TTRSD improves performance across seven benchmarks and three VLMs, raising InternVL3-2B's MMMU accuracy from 35.79% to 49.32%(+13.53%), demonstrating cross-dataset generalization while preserving inherent reasoning integrity.

cs.CV↗