IG-Search: Step-Level Information Gain Rewards for Search-Augmented Reasoning

The key technical bet here is that measuring confidence shifts in the model itself, rather than checking final answer correctness, gives a cleaner training signal at each retrieval step. This sidesteps the credit assignment problem that plagues trajectory-level rewards, where a single correct answer at the end tells you nothing about which search queries actually helped.

This connects directly to the step-level reasoning theme running through recent coverage. The SpecGuard paper from the same day ('From Tokens to Steps: Verification-Aware Speculative Decoding') also argues that reasoning quality is better evaluated at the step level using internal model signals rather than external judges. Both papers are converging on the same architectural intuition from different directions: that intermediate states carry more useful signal than endpoints. The DiscoTrace work from the same period adds a complementary angle, showing that LLMs already differ from humans in how they construct information-seeking answers, which raises the question of whether optimizing retrieval behavior against model confidence actually reinforces those existing gaps.

The critical test is whether IG-Search's gains hold on multi-hop benchmarks like MuSiQue or 2WikiMultiHopQA, where retrieval chains are longer and confidence calibration errors compound. If performance degrades relative to trajectory-level baselines on those tasks, the information gain signal may be too local to guide complex reasoning chains.