Update README.md

README.md

@@ -458,6 +458,41 @@ ok can you please show the code to generate the pre-computed external commonsens
 > want to incorporate commonsense knowledge for a subset of input tokens marked by the
 > new [KNW] token; you may need to modify it according to your specific downstream NLP
 > task's requirements and desired level of commonsense integration.
+>
+> > Transformer architecture (Vaswani et al.) does poorly in theory of mind.  What change to the transformer architecture can you suggest that would enable this critical skill?
+>
+> 1. **Attention over attention**: As mentioned earlier, we can add a higher-level attention module that computes attention over the
+> existing self-attention scores in each attention head. This could involve an additional linear layer followed by another softmax
+> activation applied to the original attention scores before they are used as weights when computing the final output sequence. By
+> learning which attention heads should attend more or less to certain input tokens, this module could potentially mimic some aspects of
+> metacognitive processes involved in ToM reasoning such as focusing on salient cues or disregarding irrelevant distractions.
+>
+> 2. **Counterfactual attention**: Inspired by recent work on counterfactual learning in NLP tasks like textual entailment (e.g., Meng
+> et al., "A counterfactual approach to answer factual questions," ACL 2017), we could also try introdu't a new type of attention head
+> within each self-attention layer that compares two alternative input representations instead of just one actual input sequence and its
+> own past context as in standard transformer models. Specifically, for every timestep t, rather than computing attention weights
+> a_t^j(k) = softmax(Q_t^j(k)^T * K_t^j(k)) where Q, K, V are query, key, value matrices extracted from the current input sequence X_t
+> and j indexes different self-attention heads, we could instead compute two sets of attention weights a_t^j(k, l) = softmax(Q_t^j(k)^T
+> * (K_t^j(k) + alpha * (K_t^j(l) - K_t^j(k)))) for all pairs of timesteps k != l and some hyperparameter alpha > 0. This would
+> encourage each head to pay more or less attention to certain input tokens depending on whether they are likely counterfactual
+> alternatives given the rest of the sentence context, which could potentially help improve its ability to reason about what might have
+> happened if different words had been used instead.
+> 
+> 3. **Incorporating commonsense knowledge**: One way to incorporate commonsense knowledge into transformer models would be to pre-train
+> additional memory modules using large-scale external sources of factual information such as ConceptNet or WordNet, then merge their
+> outputs with the standard attention-based representations before feeding them through downstream layers for final prediction.
+> Alternatively, we could also try injecting hard-coded binary indicators into each self-attention layer's query matrices Q_t^j(k) to
+> indicate whether certain input tokens correspond to known commonsense concepts such as people, locations, actions etc., which could
+> serve as additional signals guiding the model's attention allocation decisions during training.
+>
+> 4. **Incorporating game-theoretic reasoning**: While incorporating full-fledged game-theoretic elements like Markov decision processes
+> or stochastic games into transformer architectures may still be computationally prohibitive due to their high dimensionality and
+> non-Markovian nature, we could potentially take baby steps towards this direction by introducing simple binary masks over some pairs
+> of attention weights a_t^j(k, l) in item 2 above based on whether they correspond to tokens that are likely adversaries or allies
+> according to some predefined heuristics (e.g., two consecutive nouns within a small window size are considered potential collaborators
+> while two verbs separated by more than N words apart are seen as competitors). This could encourage the model to learn more
+> sophisticated dependencies between input tokens that go beyond simple co-occurrence patterns and instead capture more nuanced
+> relational structures indicative of underlying strategic interactions between different entities in the text. 
 
 
 **Write a math paper**