Source-linked AI summary
SYNLAT: Syntax-Aligned Text-Latent Compression for Chain-of-Thought Reasoning
Yifeng Zhao, Hongjun Yu, Shibo Wang, Yunjiao Zhou, Zixiao Zhu, Zhipeng Ning, Kezhi Mao, Junlang Qian
TL;DR
Long CoT traces raise output-token costs, while explicit compression can harm logical fidelity and existing boundary choices can fragment or bind reasoning relations. SynLat uses syntax-aligned units and answer-conditioned targets to train a compression-conditioned Student. Under achieved-CR selection, it improves aggregate accuracy over evaluated baselines at MEDIUM and HIGH compression, with the largest reported gains at HIGH.
Problem
Long CoT traces increase output-token costs, and aggressive deletion or condensation can reduce logical fidelity or fragment symbolic expressions.
Source: §1 · p. 1Method
SynLat partitions traces into non-overlapping Syntax-Aligned Units, assigns answer-conditioned KEEP/LATENT targets offline, and transfers them to a Student conditioned on the question and compression level.
Source: §1 · p. 1, §3.3 · p. 3Results
Under achieved-CR selection, SynLat exceeds the strongest evaluated baseline overall by 7.0 points for Qwen3-8B at HIGH compression, with aggregate gains increasing with compression across both Student scales.
Source: §5 · p. 4Takeaways & Limitations
The results indicate that CoT compression involves matching reasoning structure with answer-relevant information, not only reducing visible tokens.
Source: §6 · p. 4Baselines still outperform SynLat in some task-level comparisons, including Qwen3-14B on StrategyQA at MEDIUM compression.
Source: §5 · p. 4
Abstract
from arXiv · showhide
Long chain-of-thought (CoT) traces impose substantial output-token costs. Under constrained budgets, compression must preserve answer-critical information, making boundary placement central. Token-level and fixed-length boundaries can fragment coherent spans such as phrases, formulas, and local derivations, whereas step-level boundaries can bind content requiring different compression actions. We introduce SynLat, a text-latent CoT framework that aligns compression boundaries with syntactic structure through non-overlapping Syntax-Aligned Units (SAUs). An answer-conditioned Teacher constructs progressive KEEP/LATENT targets for a single compression-conditioned Student, which generates mixed reasoning from only the question and requested compression level at inference. Across two Qwen3 Student scales, Standard-CoT and Long-CoT groups, and three compression levels, SynLat matches or exceeds the strongest evaluated baseline in all 12 task-group aggregates and strictly leads in 11 under the reported achieved-CR selection protocol. Overall gains reach 3.6/2.6 points at MEDIUM and 7.0/5.5 points at HIGH for Qwen3-8B/14B, with larger advantages under stronger compression, particularly on Long-CoT groups.
1. INTRODUCTION
SynLat addresses the difficulty of compressing long CoT traces without losing answer-critical reasoning by aligning text–latent decisions with syntactic structure. Under the reported achieved-CR selection protocol, it improves the accuracy–compression trade-off, with larger gains under stronger compression, particularly on Long-CoT groups.
- Long CoT traces raise output-token costs; explicit deletion or condensation can reduce logical fidelity and fragment symbolic expressions at aggressive compression ratios.Source: §1 · p. 1
- Token and fixed-window units can fragment local relations, while step-level units can bind distinct relations instead of leaving them separately assignable to text or latent form.Source: Figure 1 · p. 1, §1 · p. 1
- SynLat uses ordered, non-overlapping Syntax-Aligned Units and answer-conditioned KEEP/LATENT targets, then transfers them to a Student that needs only the question and compression level at inference.Source: §1 · p. 1, Figure 2 · p. 2
- SynLat improves the accuracy–compression trade-off under achieved-CR selection, with larger gains at stronger compression and particularly on Long-CoT groups.Source: §1 · p. 1
- SynLat frames budgeted CoT compression as allocating representation fidelity, retaining answer-critical reasoning as text while compressing recoverable content into latent representations.Source: §1 · p. 1
2. RELATION TO PRIOR WORK
Prior work compresses visible CoT, moves reasoning into latent representations, or combines text and latent forms. SynLat distinguishes itself by using trace-specific syntax units for compression-conditioned text–latent allocation.
- Prior work spans visible-text compression, continuous and discrete latent reasoning, and hybrid methods; SynLat instead flattens trace-specific syntax into non-overlapping units for text–latent allocation.Source: §2 · p. 2
3. METHOD
SynLat represents reasoning as retained text mixed with latent tokens, using non-overlapping syntax-aligned units for allocation. An answer-conditioned Teacher builds progressive targets that a single Student learns to generate from the question and requested compression level.
3.1. Preliminaries
- A mixed latent CoT is an ordered trace containing retained text spans and latent tokens alongside the question-to-answer reasoning trace.The standard textual trace is represented as a sequence of tokens before the answer.Source: §3.1 · p. 2
3.2. Syntactic constituency structure
- Constituency parsing yields nested contiguous spans that SynLat flattens into ordered, non-overlapping Syntax-Aligned Units, finer than full reasoning steps.The units use span boundaries rather than syntactic labels.Source: §3.2 · p. 2, Figure 3 · p. 2
3.3. SynLat: syntax-aligned text–latent CoT compression
- Tree-guided construction caps spans, merges adjacent siblings within the cap, and uses residual leaf spans to partition the trace.The resulting ordered units fully reconstruct the original reasoning trace.Source: §3.3 · p. 3
- For each condition, the answer-conditioned Teacher preserves answer-critical units and replaces recoverable units, with higher compression adding LATENT units without restoring earlier replacements.LATENT units receive symbols from a shared finite codebook.Source: §3.3 · p. 3
- Plan-then-generate training transfers the Teacher’s targets to one Student, which generates mixed reasoning from only the question and compression condition at deployment.Stage II removes the reference trace from generation training.Source: §3.3 · p. 4
4. EXPERIMENTS
The experiments evaluate two Qwen3 Student scales across five benchmarks and compare accuracy under a reported achieved-compression-rate selection protocol. They also describe the matched-compression measures used to compare compression units.
Models, data, and benchmarks
- Qwen3-8B and Qwen3-14B Students are evaluated on MATH-500, StrategyQA, OlympiadBench, WorldTree V2, and ProofWriter.Tasks are grouped as Standard-CoT or Long-CoT based on preanswer lengths of correct, uncompressed development traces.Source: §4 · p. 4
- The evaluation uses final-answer accuracy as its primary metric and excludes evaluation instances and reference solutions from target construction, codebook fitting, and model selection.Compression comparisons share Student scales, trace pool, splits, validation-only selection, and decoding.Source: §4 · p. 4
Baselines and evaluation protocol
- LOW, MEDIUM, and HIGH target 25%, 50%, and 75% SAU replacement, while overall achieved compression rates are about 23%, 44%, and 65%.Baseline task points are selected within four percentage points of SynLat’s task-level rate, or the nearest available point otherwise.Source: §4 · p. 4
- Table 2 compares compression units at comparable achieved rates using accuracy retention, average unit length, constituent/formula split rate, and achieved compression rate.Token Deletion is the only compared unit that also changes the compressed representation.Source: Table 2 · p. 4
5. RESULTS
Under the reported achieved-CR selection protocol, SynLat’s aggregate accuracy advantage over the strongest evaluated baseline grows with compression for both Student scales. The results also report unit and planning ablations.
Main results
- At MEDIUM/HIGH, SynLat exceeds the strongest evaluated baseline overall by 3.6/7.0 points for Qwen3-8B and 2.6/5.5 points for Qwen3-14B.The aggregate margins increase with compression under the reported achieved-CR selection protocol.Source: §5 · p. 4
- At HIGH, the Qwen3-8B margin is 4.7 points on Standard-CoT and 8.5 points on Long-CoT.The reported Long-CoT advantage is larger.Source: §5 · p. 4
Compression-unit ablation
- At approximately 44% achieved CR, SAUs retain 94.0% of NONE accuracy; fixed windows, guarded windows, and steps trail by 2.6, 1.8, and 4.4 points.The results attribute the shortfall of steps to coarse units hindering selective compression.Source: §5 · p. 4
Planning and training ablations
- On fixed SAUs, answer-conditioned planning retains 94.0%/89.3% of NONE accuracy at MEDIUM/HIGH, versus 92.0%/86.2% for the strongest answer-free alternatives.Removing action-plan supervision lowers mean accuracy across the three compressed levels from 73.1 to 70.6.Source: §5 · p. 4
6. CONCLUSION
The conclusion frames CoT compression as allocating limited representation fidelity: SynLat uses syntax-aligned units to keep content in text or convert it into latent representations. Experiments report higher final-answer accuracy than representative methods under matched output budgets, particularly on longer-reasoning tasks.
6. CONCLUSION
- Under matched output budgets, SynLat achieves higher final-answer accuracy than representative compression methods, with particularly large gains on tasks requiring longer reasoning traces.The conclusion links the reported results to matching compression structure with answer-relevant information.Source: §6 · p. 4
7. COMPLIANCE WITH ETHICAL STANDARDS
The study uses publicly available benchmark datasets, involves no human or animal subjects, and required no ethical approval.
7. COMPLIANCE WITH ETHICAL STANDARDS
- The study uses public benchmark datasets, involves no human or animal subjects, and required no ethical approval.Source: §7 · p. 5