TL;DR
LLMs often struggle with "hallucinations" or omitting crucial input details during decoding. DIVER (Decoding with span-level Mutual Information VERification) solves this by verifying candidate text spans against the source input using Point-wise Mutual Information (PMI), ensuring that generated outputs are both fluent and strictly faithful to the prompt.
What Problem Does This Paper Address?
Despite their power, Large Language Models (LLMs) are notorious for being "unfaithful" to their inputs. In tasks like machine translation or summarization, they might add irrelevant content (hallucination) or overlook important details (omission).
Standard decoding strategies like greedy search or nucleus sampling select tokens based on their likelihood given the previous tokens, but they don't explicitly check if the new token is strongly correlated with the original input. While some researchers have tried using Point-wise Mutual Information (PMI) to fix this, doing so at the individual token level is often "short-sighted." A single token might seem unrelated to the input initially, but it could be the necessary bridge to a highly relevant phrase. This paper addresses the need for a decoding mechanism that looks ahead at "spans" of text to ensure global faithfulness rather than just local probability.
Methods
The authors introduce DIVER, which enhances decoding through three main steps:
- Divergence Point Identification: DIVER doesn't verify every single token (which would be computationally expensive). Instead, it identifies "divergence points"—steps where the model is uncertain and multiple candidate tokens have high probability.
- Dynamic Span Generation: At these divergence points, the model generates several short candidate "spans" (sequences of tokens) rather than just single tokens. The length of these spans is determined adaptively; DIVER looks ahead until the next potential divergence point to ensure it has enough context for a fair evaluation.
- Span-Level PMI Verification: This is the core innovation. For each candidate span, DIVER calculates a PMI score. It does this by performing "backward teacher-forcing": it asks the model, "Given that I just generated this span, how likely is the original input $x$?"
- Re-ranking: The original model probability is combined with this PMI score to re-rank the candidates. The span that best "explains" or correlates with the input is selected.
By moving from token-level to span-level verification, DIVER avoids the "local optimum" trap where a good long-term choice is rejected because its first token looks weak in isolation.
Datasets
The researchers evaluated DIVER across a massive battery of benchmarks to prove its versatility:
- Code Generation: MBPP.
- Machine Translation: Flores-200 (including challenging pairs like Chinese-English and Arabic-English).
- Summarization: CNN/DailyMail and SAMSum.
- Knowledge QA: Natural Questions and Web Questions.
- Constrained Generation: E2E (table-to-text) and CommonGen.
- Creative/Dialogue: ROCStory and DailyDialogue.
- Instruction Following: AlpacaEval.
Key Insights and Ideas
- The Information Disparity Problem: A single token often carries too little information to be meaningfully compared to a complex input sequence. Spans provide the "semantic mass" necessary for accurate PMI calculation.
- Backward Likelihood as a Truth-Serums: Using the model's own ability to "predict the input from the output" acts as a powerful verification signal. If a generated span makes the input seem more likely, it’s probably faithful.
- Efficiency via Sparsity: By only triggering verification at points of high uncertainty (divergence points), DIVER manages to improve quality without the massive slowdown associated with verifying every single token.
- Small Model Verification: Interestingly, the authors found they could use a much smaller model (like Tiny-LLaMA-1.1B) to verify a larger model (like LLaMA-2-13B) with minimal loss in performance, significantly boosting inference speed.
Key Results
- Superior Faithfulness: In human evaluations for translation, DIVER produced the "most faithful" output 35.94% of the time, easily beating greedy search and beam search.
- Handling "Distant" Languages: DIVER showed massive gains in translating distant language pairs (e.g., +4.28 points in Arabic-English) where standard LLMs often lose track of the source meaning.
- Solving Omissions: In the E2E table-to-text task, DIVER successfully included specific input details (like restaurant names) that every other decoding method consistently ignored.
- Performance vs. Baselines: DIVER consistently outperformed advanced baselines like Contrastive Decoding (CD) and Context-Aware Decoding (CAD) across almost all tasks.
Limitations and Future Work
- Computational Cost: While faster than some alternatives, DIVER is still slower than vanilla greedy decoding (roughly 0.6x to 0.8x the speed) due to the extra backward passes required for PMI calculation.
- Hyper-parameter Sensitivity: The method relies on a hyper-parameter $\gamma$ to define divergence points. While $\gamma \le 0.3$ generally works well, optimal settings can vary slightly by task.
- Future Directions: The authors suggest combining DIVER with speculative decoding. This could potentially allow the model to verify spans in parallel with generation, hiding the latency and making the method viable for real-time applications.
Summary updated